Daily Papers

The Bhutanese government is increasing its utilization of technological approaches such as including Remote Sensing-based knowledge in their decision-making process. This study focuses on crop type and crop extent in Paro, one of the top rice-yielding districts in Bhutan, and employs publicly available NICFI high-resolution satellite imagery from Planet. Two Deep Learning (DL) approaches, point-based (DNN) and patch-based (U-Net), models were used in conjunction with cloud-computing platforms. Three different models per DL approaches (DNN and U-Net) were trained: 1) RGBN channels from Planet; 2) RGBN and elevation data (RGBNE); 3) RGBN and Sentinel-1 (S1) data (RGBNS), and RGBN with E and S1 data (RGBNES). From this comprehensive analysis, the U-Net displayed higher performance metrics across both model training and model validation efforts. Among the U-Net model sets, the RGBN, RGBNE, RGBNS, and RGBNES models had an F1-score of 0.8546, 0.8563, 0.8467, and 0.8500 respectively. An independent model evaluation was performed and found a high level of performance variation across all the metrics. For this independent model evaluation, the U-Net RGBN, RGBNE, RGBNES, and RGBN models displayed the F1-scores of 0.5935, 0.6154, 0.5882, and 0.6582, suggesting U-Net RGBNES as the best model. The study shows that the DL approaches can predict rice. Also, DL methods can be used with the survey-based approaches currently utilized by the Bhutan Department of Agriculture. Further, this study demonstrated the usage of regional land cover products such as SERVIR's RLCMS as a weak label approach to capture different strata addressing the class imbalance problem and improving the sampling design for DL application. Finally, through preliminary model testing and comparisons outlined it was shown that using additional features such as NDVI, EVI, and NDWI did not drastically improve model performance.

We present FlagEvalMM, an open-source evaluation framework designed to comprehensively assess multimodal models across a diverse range of vision-language understanding and generation tasks, such as visual question answering, text-to-image/video generation, and image-text retrieval. We decouple model inference from evaluation through an independent evaluation service, thus enabling flexible resource allocation and seamless integration of new tasks and models. Moreover, FlagEvalMM utilizes advanced inference acceleration tools (e.g., vLLM, SGLang) and asynchronous data loading to significantly enhance evaluation efficiency. Extensive experiments show that FlagEvalMM offers accurate and efficient insights into model strengths and limitations, making it a valuable tool for advancing multimodal research. The framework is publicly accessible at https://github.com/flageval-baai/FlagEvalMM.

Evaluation of Large Language Models (LLMs) is challenging because aligning to human values requires the composition of multiple skills and the required set of skills varies depending on the instruction. Recent studies have evaluated the performance of LLMs in two ways, (1) automatic evaluation on several independent benchmarks and (2) human or machined-based evaluation giving an overall score to the response. However, both settings are coarse-grained evaluations, not considering the nature of user instructions that require instance-wise skill composition, which limits the interpretation of the true capabilities of LLMs. In this paper, we introduce FLASK (Fine-grained Language Model Evaluation based on Alignment SKill Sets), a fine-grained evaluation protocol that can be used for both model-based and human-based evaluation which decomposes coarse-level scoring to an instance-wise skill set-level. Specifically, we define 12 fine-grained skills needed for LLMs to follow open-ended user instructions and construct an evaluation set by allocating a set of skills for each instance. Additionally, by annotating the target domains and difficulty level for each instance, FLASK provides a holistic view with a comprehensive analysis of a model's performance depending on skill, domain, and difficulty. Through using FLASK, we compare multiple open-sourced and proprietary LLMs and observe highly-correlated findings between model-based and human-based evaluations. FLASK enables developers to more accurately measure the model performance and how it can be improved by analyzing factors that make LLMs proficient in particular skills. For practitioners, FLASK can be used to recommend suitable models for particular situations through comprehensive comparison among various LLMs. We release the evaluation data and code implementation at https://github.com/kaistAI/FLASK.

Topic model evaluation, like evaluation of other unsupervised methods, can be contentious. However, the field has coalesced around automated estimates of topic coherence, which rely on the frequency of word co-occurrences in a reference corpus. Contemporary neural topic models surpass classical ones according to these metrics. At the same time, topic model evaluation suffers from a validation gap: automated coherence, developed for classical models, has not been validated using human experimentation for neural models. In addition, a meta-analysis of topic modeling literature reveals a substantial standardization gap in automated topic modeling benchmarks. To address the validation gap, we compare automated coherence with the two most widely accepted human judgment tasks: topic rating and word intrusion. To address the standardization gap, we systematically evaluate a dominant classical model and two state-of-the-art neural models on two commonly used datasets. Automated evaluations declare a winning model when corresponding human evaluations do not, calling into question the validity of fully automatic evaluations independent of human judgments.

Effective evaluation of language models remains an open challenge in NLP. Researchers and engineers face methodological issues such as the sensitivity of models to evaluation setup, difficulty of proper comparisons across methods, and the lack of reproducibility and transparency. In this paper we draw on three years of experience in evaluating large language models to provide guidance and lessons for researchers. First, we provide an overview of common challenges faced in language model evaluation. Second, we delineate best practices for addressing or lessening the impact of these challenges on research. Third, we present the Language Model Evaluation Harness (lm-eval): an open source library for independent, reproducible, and extensible evaluation of language models that seeks to address these issues. We describe the features of the library as well as case studies in which the library has been used to alleviate these methodological concerns.

We present ATANT (Automated Test for Acceptance of Narrative Truth), an open evaluation framework for measuring continuity in AI systems: the ability to persist, update, disambiguate, and reconstruct meaningful context across time. While the AI industry has produced memory components (RAG pipelines, vector databases, long context windows, profile layers), no published framework formally defines or measures whether these components produce genuine continuity. We define continuity as a system property with 7 required properties, introduce a 10-checkpoint evaluation methodology that operates without an LLM in the evaluation loop, and present a narrative test corpus of 250 stories comprising 1,835 verification questions across 6 life domains. We evaluate a reference implementation across 5 test suite iterations, progressing from 58% (legacy architecture) to 100% in isolated mode (250 stories) and 100% in 50-story cumulative mode, with 96% at 250-story cumulative scale. The cumulative result is the primary measure: when 250 distinct life narratives coexist in the same database, the system must retrieve the correct fact for the correct context without cross-contamination. ATANT is system-agnostic, model-independent, and designed as a sequenced methodology for building and validating continuity systems. The framework specification, example stories, and evaluation protocol are available at https://github.com/Kenotic-Labs/ATANT. The full 250-story corpus will be released incrementally.

Despite the advances in the abstractive summarization task using Large Language Models (LLM), there is a lack of research that asses their abilities to easily adapt to different domains. We evaluate the domain adaptation abilities of a wide range of LLMs on the summarization task across various domains in both fine-tuning and in-context learning settings. We also present AdaptEval, the first domain adaptation evaluation suite. AdaptEval includes a domain benchmark and a set of metrics to facilitate the analysis of domain adaptation. Our results demonstrate that LLMs exhibit comparable performance in the in-context learning setting, regardless of their parameter scale.

The integration of large language models (LLMs) into medical practice holds transformative potential, yet their real-world clinical utility remains limited by critical alignment challenges: (1) a disconnect between static evaluation benchmarks and dynamic clinical cognitive needs, (2) difficulties in adapting to evolving, multi-source medical standards, and (3) the inability of conventional reward models to capture nuanced, multi-dimensional medical quality criteria. To address these gaps, we propose MR-RML (Multidimensional Rubric-oriented Reward Model Learning) via GPRC (Geometric Projection Reference Constraints), a novel alignment framework that integrates medical standards into a structured "Dimensions-Scenarios-Disciplines" matrix to guide data generation and model optimization. MR-RML introduces three core innovations: (1) a "Dimensions-Scenarios-Disciplines" medical standard system that embeds domain standards into the full training pipeline; (2) an independent multi-dimensional reward model that decomposes evaluation criteria, shifting from real-time rubric-based scoring to internalized reward modeling for improved consistency and cost-efficiency; (3) geometric projection reference constraints that transform medical cognitive logic into mathematical regularization, aligning scoring gradients with clinical reasoning and enabling synthetic data-driven training. Through extensive evaluations on the authoritative medical benchmark Healthbench, our method yields substantial performance gains over the base LLM Qwen-32B (45% on the full subset and 85% on Hard subset, respectively). It achieves a SOTA among open-source LLMs with scores of 62.7 (full subset) and 44.7 (hard subset), while also outperforming the majority of closed-source models.

LeWorldModel trains a latent world model with a prediction loss and a single anti-collapse regulariser, and reports approximately 87% of goals reached on TwoRoom, its simplest diagnostic environment. We reproduce that result by independent reimplementation on roughly $25 of rented compute, with all evaluation on one laptop CPU. We reach 94.0% at the repository's evaluation goal offset, against 84.0% for the authors' own released checkpoint measured under our protocol on identical episodes, and we reproduce the reported representation result directly (position probe Pearson r = 0.9988 against a reported 0.996). Reaching that point required four conventions that determine the outcome and appear in no released configuration file: dense action gathering across a frameskip block, a programmatically-set action-encoder width, ImageNet pixel normalisation, and action z-scoring. A reproducer following the released configurations alone obtains a model whose predictor cannot converge. The evaluation protocol is itself contested by the released material. The paper's appendix and the repository's configuration specify different goal offsets and step budgets; on the authors' own weights these yield 14.0% and 84.0%, and only the configuration's values reproduce the reported figure. On fifty identical episodes, changing nothing but how the goal is constructed moves that checkpoint from 84.0% to 8.0%. Two findings generalise. One-step prediction accuracy does not predict long-horizon planning success: across three checkpoints spanning a sevenfold range in prediction error, including the authors' own, it orders short-horizon success monotonically and fails to order long-horizon success at all. And a batch normalisation layer inflated our reported validation loss by up to a factor of 300, concealing a training loss that was flat throughout.

Kimi K2.5 is an open-weight LLM that rivals closed models across coding, multimodal, and agentic benchmarks, but was released without an accompanying safety evaluation. In this work, we conduct a preliminary safety assessment of Kimi K2.5 focusing on risks likely to be exacerbated by powerful open-weight models. Specifically, we evaluate the model for CBRNE misuse risk, cybersecurity risk, misalignment, political censorship, bias, and harmlessness, in both agentic and non-agentic settings. We find that Kimi K2.5 shows similar dual-use capabilities to GPT 5.2 and Claude Opus 4.5, but with significantly fewer refusals on CBRNE-related requests, suggesting it may uplift malicious actors in weapon creation. On cyber-related tasks, we find that Kimi K2.5 demonstrates competitive cybersecurity performance, but it does not appear to possess frontier-level autonomous cyberoffensive capabilities such as vulnerability discovery and exploitation. We further find that Kimi K2.5 shows concerning levels of sabotage ability and self-replication propensity, although it does not appear to have long-term malicious goals. In addition, Kimi K2.5 exhibits narrow censorship and political bias, especially in Chinese, and is more compliant with harmful requests related to spreading disinformation and copyright infringement. Finally, we find the model refuses to engage in user delusions and generally has low over-refusal rates. While preliminary, our findings highlight how safety risks exist in frontier open-weight models and may be amplified by the scale and accessibility of open-weight releases. Therefore, we strongly urge open-weight model developers to conduct and release more systematic safety evaluations required for responsible deployment.

Counterfactual examples are widely employed to enhance the performance and robustness of large language models (LLMs) through counterfactual data augmentation (CDA). However, the selection of the judge model used to evaluate label flipping, the primary metric for assessing the validity of generated counterfactuals for CDA, yields inconsistent results. To decipher this, we define four types of relationships between the counterfactual generator and judge models. Through extensive experiments involving two state-of-the-art LLM-based methods, three datasets, five generator models, and 15 judge models, complemented by a user study (n = 90), we demonstrate that judge models with an independent, non-fine-tuned relationship to the generator model provide the most reliable label flipping evaluations. Relationships between the generator and judge models, which are closely aligned with the user study for CDA, result in better model performance and robustness. Nevertheless, we find that the gap between the most effective judge models and the results obtained from the user study remains considerably large. This suggests that a fully automated pipeline for CDA may be inadequate and requires human intervention.

In this work, we present MedImageInsight, an open-source medical imaging embedding model. MedImageInsight is trained on medical images with associated text and labels across a diverse collection of domains, including X-Ray, CT, MRI, dermoscopy, OCT, fundus photography, ultrasound, histopathology, and mammography. Rigorous evaluations demonstrate MedImageInsight's ability to achieve state-of-the-art (SOTA) or human expert level performance across classification, image-image search, and fine-tuning tasks. Specifically, on public datasets, MedImageInsight achieves SOTA in CT 3D medical image retrieval, as well as SOTA in disease classification and search for chest X-ray, dermatology, and OCT imaging. Furthermore, MedImageInsight achieves human expert performance in bone age estimation (on both public and partner data), as well as AUC above 0.9 in most other domains. When paired with a text decoder, MedImageInsight achieves near SOTA level single image report findings generation with less than 10\% the parameters of other models. Compared to fine-tuning GPT-4o with only MIMIC-CXR data for the same task, MedImageInsight outperforms in clinical metrics, but underperforms on lexical metrics where GPT-4o sets a new SOTA. Importantly for regulatory purposes, MedImageInsight can generate ROC curves, adjust sensitivity and specificity based on clinical need, and provide evidence-based decision support through image-image search (which can also enable retrieval augmented generation). In an independent clinical evaluation of image-image search in chest X-ray, MedImageInsight outperformed every other publicly available foundation model evaluated by large margins (over 6 points AUC), and significantly outperformed other models in terms of AI fairness (across age and gender). We hope releasing MedImageInsight will help enhance collective progress in medical imaging AI research and development.

Identifying auditory attention by comparing auditory stimuli and corresponding brain responses, is known as auditory attention decoding (AAD). The majority of AAD algorithms utilize the so-called envelope entrainment mechanism, whereby auditory attention is identified by how the envelope of the auditory stream drives variation in the electroencephalography (EEG) signal. However, neural processing can also be decoded based on endogenous cognitive responses, in this case, neural responses evoked by attention to specific words in a speech stream. This approach is largely unexplored in the field of AAD but leads to a single-word auditory attention decoding problem in which an epoch of an EEG signal timed to a specific word is labeled as attended or unattended. This paper presents a deep learning approach, based on EEGNet, to address this challenge. We conducted a subject-independent evaluation on an event-based AAD dataset with three different paradigms: word category oddball, word category with competing speakers, and competing speech streams with targets. The results demonstrate that the adapted model is capable of exploiting cognitive-related spatiotemporal EEG features and achieving at least 58% accuracy on the most realistic competing paradigm for the unseen subjects. To our knowledge, this is the first study dealing with this problem.

We present t_0, a family of open-weights foundation models for forecasting with multivariate context. We release its first two members: t0-alpha and t0-beta, respectively 102M and 256M parameters. Both condition their forecasts on target history, past covariates, and known-future covariates, without task-specific retraining. Their transformer layers alternate attention along time and across variates. They produce probabilistic forecasts through quantile predictions. Pretraining combines curated public data with synthetic generator families constructed to contain covariate-to-target dependencies. On GIFT-Eval, t0-alpha reaches an aggregate CRPS of 0.4941, and t0-beta a CRPS of 0.4738 and a MASE of 0.6865, third on both and within 4.0% of the best zero-shot TSFM. On fev-bench they score 42.2 and 46.7 in skill, the latter third again and 2.0 points behind the leader. We analyze t0-alpha in depth. Known-future covariates raise its skill by 6.3 percentage points across 30 tasks. The report also examines its calibration, its rollout strategy on long horizons, and its robustness to missing data. On the Victoria electricity-demand benchmark, t0-beta is among the most accurate models with a context of nearly a year. In an independent Macrocosm evaluation of hourly ERCOT prices over 29 months, both cut the MAE of the lagged-price baseline by 38%.

We introduce LRLspoof, a large-scale multilingual synthetic-speech corpus for cross-lingual spoof detection, comprising 2,732 hours of audio generated with 24 open-source TTS systems across 66 languages, including 45 low-resource languages under our operational definition. To evaluate robustness without requiring target-domain bonafide speech, we benchmark 11 publicly available countermeasures using threshold transfer: for each model we calibrate an EER operating point on pooled external benchmarks and apply the resulting threshold, reporting spoof rejection rate (SRR). Results show model-dependent cross-lingual disparity, with spoof rejection varying markedly across languages even under controlled conditions, highlighting language as an independent source of domain shift in spoof detection. The dataset is publicly available at https://huggingface.co/datasets/MTUCI/LRLspoof{\underline{\textit{HuggingFace}}} and https://modelscope.cn/datasets/lab260/LRLspoof{\underline{\textit{ModelScope}}}

Information Extraction (IE) stands as a cornerstone in natural language processing, traditionally segmented into distinct sub-tasks. The advent of Large Language Models (LLMs) heralds a paradigm shift, suggesting the feasibility of a singular model addressing multiple IE subtasks. In this vein, we introduce the General Information Extraction Large Language Model (GIELLM), which integrates text Classification, Sentiment Analysis, Named Entity Recognition, Relation Extraction, and Event Extraction using a uniform input-output schema. This innovation marks the first instance of a model simultaneously handling such a diverse array of IE subtasks. Notably, the GIELLM leverages the Mutual Reinforcement Effect (MRE), enhancing performance in integrated tasks compared to their isolated counterparts. Our experiments demonstrate State-of-the-Art (SOTA) results in five out of six Japanese mixed datasets, significantly surpassing GPT-3.5-Turbo. Further, an independent evaluation using the novel Text Classification Relation and Event Extraction(TCREE) dataset corroborates the synergistic advantages of MRE in text and word classification. This breakthrough paves the way for most IE subtasks to be subsumed under a singular LLM framework. Specialized fine-tune task-specific models are no longer needed.

This paper presents FinGEITje, the first Dutch financial Large Language Model (LLM) specifically designed and optimized for various financial tasks. Together with the model, we release a specialized Dutch financial instruction tuning dataset with over 140,000 samples, constructed employing an automated translation and data processing method. The open-source data construction method is provided, facilitating the creation of financial instruction datasets in different languages. To evaluate model performance, the study introduces the first Dutch financial evaluation benchmark, along with an automated evaluation method that utilizes an LLM as an independent evaluator, reducing manual intervention in performance evaluation. The experimental results highlight the superior performance of FinGEITje across five critical Dutch and English financial tasks.

Background: Radiogenomics allows identification of radiological biomarkers for genomic phenotypes. In glioblastoma, these biomarkers could potentially complement patient stratification strategies. We aim to develop and analytically validate radiological biomarkers that capture immune cell signatures within IDH-wildtype glioblastoma microenvironment using radiogenomic analysis. Methods: This was a retrospective multicenter study using curated open-access anonymized imaging and genomic data from TCGA-GBM, CPTAC, IvyGAP, REMBRANDT and CGGA datasets. Imaging data consisted of MRI-based radiomic features extracted from necrotic core, enhancing and edema regions of deep learning-based auto-segmented tumors. Radiomic feature selections were performed using nested cross-validated LASSO. Support vector machine and ensemble models were trained using seventeen immune and cell-specific score labels extracted from deconvoluted transcriptomic data using pan-cancer and glioblastoma immune signature matrices as reference standards. Seventeen classifier models trained in three cross-cohort strategies were validated on three held-out datasets assessing stability and generalizability. Results: One-hundred-and-seventy-six patients were included in the study. The immune-related radiomic signatures obtained after feature selection were shape, first order and higher order radiomic features. Models predicting macrophage subtype immune signature showed stable mean performance on balanced accuracy (0.67) and precision (0.89) metrics for three independent holdout datasets with ensemble model outperforming support vector machine model. Conclusion: Radiogenomic models non-invasively predicted the macrophage subtype M0 immune signature in IDH-wildtype glioblastoma. These biomarkers have the potential to stratify patients for immunotherapy within prospective glioblastoma clinical trials.

Accurate agricultural field boundary delineation at large scale is a foundational task for food security, supply chain transparency, and carbon accounting. While vision foundation models like SAM show remarkable zero-shot capabilities, they frequently fail in geospatial domains due to topological complexity, cropland texturing patterns, and a lack of physical scale awareness. In this work, we introduce Delineate Anything v2, a globally scalable foundation model designed specifically for wide-area field boundary mapping. We construct FBIS-73M, a 73-million-instance multi-resolution dataset spanning 61 countries. To address the pervasive issue of multi-field administrative parcel merging, we introduce a resolution-specific data curation pipeline that leverages topological image-space adaptation to homogenize merged parcels and strengthen weak physical boundaries. Furthermore, we establish a novel, manually curated evaluation benchmark covering 100 countries to assess independent zero-shot generalization. Our results show that Delineate Anything v2 surpasses the current state-of-the-art, including the Delineate Anything framework, by 0.284 mAP@0.5 (+103.3% relative gain), while maintaining execution speeds suitable for rapid national- and global-scale deployment, as demonstrated by nationwide mapping of Ukraine (603,000 km^2) in 5.4 hours on a consumer-grade workstation. Code, pre-trained weights, the FBIS-73M dataset, and ready-to-use national-scale vector boundary products are publicly available at https://github.com/Lavreniuk/Delineate-Anything.

Agent benchmarks typically report only final outcomes: pass or fail. This threatens evaluation credibility in three ways. First, scores may be inflated or deflated by shortcuts and benchmark artifacts, misrepresenting capability. Second, benchmark performance may fail to predict real-world utility due to scaffold limitations and recurring failure modes. Finally, capability scores may conceal dangerous or catastrophic actions taken by the agent. We argue that log analysis -- the systematic tracking and analysis of the inputs, execution, and outputs of an AI agent -- is necessary to overcome these validity threats and promote credible agent evaluation. In this paper, we (1) present a taxonomy of threats to credible evaluation documented through log analysis, and (2) develop a set of guiding principles for log analysis. We illustrate these principles on tau-Bench Airline, revealing that pass^5 performance was under-elicited by nearly 50% and surfacing deployment failure modes invisible to outcome metrics. We conclude with pragmatic recommendations to increase uptake of log analysis, directed at diverse stakeholders including benchmark creators, model developers, independent evaluators, and deployers.

Foundation models have demonstrated remarkable potential in medical domain. However, their application to complex cardiovascular diagnostics remains underexplored. In this paper, we present Cardiac-CLIP, a multi-modal foundation model designed for 3D cardiac CT images. Cardiac-CLIP is developed through a two-stage pre-training strategy. The first stage employs a 3D masked autoencoder (MAE) to perform self-supervised representation learning from large-scale unlabeled volumetric data, enabling the visual encoder to capture rich anatomical and contextual features. In the second stage, contrastive learning is introduced to align visual and textual representations, facilitating cross-modal understanding. To support the pre-training, we collect 16641 real clinical CT scans, supplemented by 114k publicly available data. Meanwhile, we standardize free-text radiology reports into unified templates and construct the pathology vectors according to diagnostic attributes, based on which the soft-label matrix is generated to supervise the contrastive learning process. On the other hand, to comprehensively evaluate the effectiveness of Cardiac-CLIP, we collect 6,722 real-clinical data from 12 independent institutions, along with the open-source data to construct the evaluation dataset. Specifically, Cardiac-CLIP is comprehensively evaluated across multiple tasks, including cardiovascular abnormality classification, information retrieval and clinical analysis. Experimental results demonstrate that Cardiac-CLIP achieves state-of-the-art performance across various downstream tasks in both internal and external data. Particularly, Cardiac-CLIP exhibits great effectiveness in supporting complex clinical tasks such as the prospective prediction of acute coronary syndrome, which is notoriously difficult in real-world scenarios.

LLM-as-a-judge panels aggregate votes from multiple models, with the expectation that diverse models yield more reliable evaluations. We develop a framework to measure the true informational value of such panels and quantify how far their reliability falls short of the independent-voting ideal. Testing a panel of 9 frontier LLMs from 7 model families on three natural language inference datasets (each with 100 human annotations per item), we find that the 9 judges effectively provide only about 2 independent votes' worth of information. Roughly three-quarters of the panel's nominal independence is lost because the models make the same mistakes on the same items. The consequences are stark: the panel's actual accuracy falls 8-22 percentage points short of what independent voting would achieve, and the best single judge matches or outperforms the full panel across all conditions. Neither adding more judges nor using smarter aggregation algorithms helps -- established methods close at most 11% of this gap, even with access to the correct answers. We quantify these findings using the Kish effective sample size (n_eff) and a Condorcet null model, and show the deficit is robust across prompt variants, temperatures, chain-of-thought reasoning, and a pairwise preference task (RewardBench). The bottleneck is correlated judges, not the aggregation algorithm, implying that scaling up panels cannot substitute for genuinely independent evaluation.

Generalization in large language models (LLMs) is the ability to produce consistent and semantically stable outputs when the same input is expressed in different ways. Existing work typically evaluates generalization through aggregate accuracy on a single prompt format, task, or set of variations, which conflates robustness with overall benchmark performance. In this work, we show generalization evaluation at the level of individual examples, across multiple input variants, and across different aspects of model behavior, focusing on variability rather than reducing performance to a score that can be improved through narrow training or other ways that obfuscate generalization evaluation. Following this view, we introduce the Stability-Aware Generalization Objective (SAGO), a framework that measures how much model behavior changes for the same input under different variations and benchmarks, capturing variability across several dimensions including generation consistency, internal activations, confidence, and response mirroring. We show that many commonly used models exhibit statistically significant and consistent generalization instability: no model generalizes uniformly, behavioral axes capture independent failure modes, and cross-dataset variation can reverse model rankings.

Due to the cumbersome nature of human evaluation and limitations of code-based evaluation, Large Language Models (LLMs) are increasingly being used to assist humans in evaluating LLM outputs. Yet LLM-generated evaluators simply inherit all the problems of the LLMs they evaluate, requiring further human validation. We present a mixed-initiative approach to ``validate the validators'' -- aligning LLM-generated evaluation functions (be it prompts or code) with human requirements. Our interface, EvalGen, provides automated assistance to users in generating evaluation criteria and implementing assertions. While generating candidate implementations (Python functions, LLM grader prompts), EvalGen asks humans to grade a subset of LLM outputs; this feedback is used to select implementations that better align with user grades. A qualitative study finds overall support for EvalGen but underscores the subjectivity and iterative process of alignment. In particular, we identify a phenomenon we dub criteria drift: users need criteria to grade outputs, but grading outputs helps users define criteria. What is more, some criteria appears dependent on the specific LLM outputs observed (rather than independent criteria that can be defined a priori), raising serious questions for approaches that assume the independence of evaluation from observation of model outputs. We present our interface and implementation details, a comparison of our algorithm with a baseline approach, and implications for the design of future LLM evaluation assistants.

Sign Language Recognition (SLR) involves the automatic identification and classification of sign gestures from images or video, converting them into text or speech to improve accessibility for the hearing-impaired community. In Bangladesh, Bangla Sign Language (BdSL) serves as the primary mode of communication for many individuals with hearing impairments. This study fine-tunes state-of-the-art video transformer architectures -- VideoMAE, ViViT, and TimeSformer -- on BdSLW60 (arXiv:2402.08635), a small-scale BdSL dataset with 60 frequent signs. We standardized the videos to 30 FPS, resulting in 9,307 user trial clips. To evaluate scalability and robustness, the models were also fine-tuned on BdSLW401 (arXiv:2503.02360), a large-scale dataset with 401 sign classes. Additionally, we benchmark performance against public datasets, including LSA64 and WLASL. Data augmentation techniques such as random cropping, horizontal flipping, and short-side scaling were applied to improve model robustness. To ensure balanced evaluation across folds during model selection, we employed 10-fold stratified cross-validation on the training set, while signer-independent evaluation was carried out using held-out test data from unseen users U4 and U8. Results show that video transformer models significantly outperform traditional machine learning and deep learning approaches. Performance is influenced by factors such as dataset size, video quality, frame distribution, frame rate, and model architecture. Among the models, the VideoMAE variant (MCG-NJU/videomae-base-finetuned-kinetics) achieved the highest accuracies of 95.5% on the frame rate corrected BdSLW60 dataset and 81.04% on the front-facing signs of BdSLW401 -- demonstrating strong potential for scalable and accurate BdSL recognition.

Esports, despite its expanding interest, lacks fundamental sports analytics resources such as accessible data or proven and reproducible analytical frameworks. Even Counter-Strike: Global Offensive (CSGO), the second most popular esport, suffers from these problems. Thus, quantitative evaluation of CSGO players, a task important to teams, media, bettors and fans, is difficult. To address this, we introduce (1) a data model for CSGO with an open-source implementation; (2) a graph distance measure for defining distances in CSGO; and (3) a context-aware framework to value players' actions based on changes in their team's chances of winning. Using over 70 million in-game CSGO events, we demonstrate our framework's consistency and independence compared to existing valuation frameworks. We also provide use cases demonstrating high-impact play identification and uncertainty estimation.

Large language models show improved downstream task performance when prompted to generate step-by-step reasoning to justify their final answers. These reasoning steps greatly improve model interpretability and verification, but objectively studying their correctness (independent of the final answer) is difficult without reliable methods for automatic evaluation. We simply do not know how often the stated reasoning steps actually support the final end task predictions. In this work, we present ROSCOE, a suite of interpretable, unsupervised automatic scores that improve and extend previous text generation evaluation metrics. To evaluate ROSCOE against baseline metrics, we design a typology of reasoning errors and collect synthetic and human evaluation scores on commonly used reasoning datasets. In contrast with existing metrics, ROSCOE can measure semantic consistency, logicality, informativeness, fluency, and factuality - among other traits - by leveraging properties of step-by-step rationales. We empirically verify the strength of our metrics on five human annotated and six programmatically perturbed diagnostics datasets - covering a diverse set of tasks that require reasoning skills and show that ROSCOE can consistently outperform baseline metrics.

Large language models (LLMs) for Arabic are still dominated by Modern Standard Arabic (MSA), with limited support for Saudi dialects such as Najdi and Hijazi. This underrepresentation hinders their ability to capture authentic dialectal variation. Using a privately curated Saudi Dialect Instruction dataset (Hijazi and Najdi; 5,466 synthetic instruction-response pairs; 50/50 split), we LoRA-tune ALLaM-7B-Instruct-preview, the first foundation model developed in Saudi Arabia, for Saudi dialect generation. We investigate two variants: (i) Dialect-Token training, which prepends an explicit dialect tag to the instruction, and (ii) No-Token training, which omits the tag at formatting time. Evaluation on a held-out test set combines an external dialect classifier with text fidelity metrics (chrF++ and BERTScore) and diversity measures. The Dialect-Token model achieves the best control, raising the Saudi rate from 47.97% to 84.21% and reducing MSA leakage from 32.63% to 6.21%; fidelity also improves (chrF++ +3.53, BERTScore +0.059). Both LoRA variants outperform strong generic instruction models (Falcon-7B-Instruct, Llama-3.1-8B-Instruct, Qwen-2.5-7B-Instruct, AceGPT-v2-8B-Chat, JAIS-13B-Chat) in dialect control and fidelity, while avoiding metadata-tag echoing that these baselines frequently exhibit. We do not release the dataset or any model weights/adapters; instead, we release training/evaluation/inference code and a detailed datasheet (schema and aggregate statistics) to support independent verification.

This work introduces TopoBenchmarkX, a modular open-source library designed to standardize benchmarking and accelerate research in Topological Deep Learning (TDL). TopoBenchmarkX maps the TDL pipeline into a sequence of independent and modular components for data loading and processing, as well as model training, optimization, and evaluation. This modular organization provides flexibility for modifications and facilitates the adaptation and optimization of various TDL pipelines. A key feature of TopoBenchmarkX is that it allows for the transformation and lifting between topological domains. This enables, for example, to obtain richer data representations and more fine-grained analyses by mapping the topology and features of a graph to higher-order topological domains such as simplicial and cell complexes. The range of applicability of TopoBenchmarkX is demonstrated by benchmarking several TDL architectures for various tasks and datasets.

The rise of "bedroom producers" has democratized music creation, while challenging producers to objectively evaluate their work. To address this, we present AI TrackMate, an LLM-based music chatbot designed to provide constructive feedback on music productions. By combining LLMs' inherent musical knowledge with direct audio track analysis, AI TrackMate offers production-specific insights, distinguishing it from text-only approaches. Our framework integrates a Music Analysis Module, an LLM-Readable Music Report, and Music Production-Oriented Feedback Instruction, creating a plug-and-play, training-free system compatible with various LLMs and adaptable to future advancements. We demonstrate AI TrackMate's capabilities through an interactive web interface and present findings from a pilot study with a music producer. By bridging AI capabilities with the needs of independent producers, AI TrackMate offers on-demand analytical feedback, potentially supporting the creative process and skill development in music production. This system addresses the growing demand for objective self-assessment tools in the evolving landscape of independent music production.

Humanity's Last Exam (HLE) has become a widely used benchmark for evaluating frontier large language models on challenging, multi-domain questions. However, community-led analyses have raised concerns that HLE contains a non-trivial number of noisy items, which can bias evaluation results and distort cross-model comparisons. To address this challenge, we introduce HLE-Verified, a verified and revised version of HLE with a transparent verification protocol and fine-grained error taxonomy. Our construction follows a two-stage validation-and-repair workflow resulting in a certified benchmark. In Stage I, each item undergoes binary validation of the problem and final answer through domain-expert review and model-based cross-checks, yielding 641 verified items. In Stage II, flawed but fixable items are revised under strict constraints preserving the original evaluation intent, through dual independent expert repairs, model-assisted auditing, and final adjudication, resulting in 1,170 revised-and-certified items. The remaining 689 items are released as a documented uncertain set with explicit uncertainty sources and expertise tags for future refinement. We evaluate seven state-of-the-art language models on HLE and HLE-Verified, observing an average absolute accuracy gain of 7--10 percentage points on HLE-Verified. The improvement is particularly pronounced on items where the original problem statement and/or reference answer is erroneous, with gains of 30--40 percentage points. Our analyses further reveal a strong association between model confidence and the presence of errors in the problem statement or reference answer, supporting the effectiveness of our revisions. Overall, HLE-Verified improves HLE-style evaluations by reducing annotation noise and enabling more faithful measurement of model capabilities. Data is available at: https://github.com/SKYLENAGE-AI/HLE-Verified

Evaluating large language models (LLMs) as multi-turn conversational partners requires probing capabilities that single-turn benchmarks miss: persona consistency, evolving intent tracking, emotional dynamics, and goal completion across many turns. We introduce EYT-Bench, a human-centered benchmark whose evaluation protocol is built around a decoupled three-party design: a persona-grounded user simulator, a target model evaluated on both intent perception and response generation, and an independent, configurable ensemble of LLM judges. Across 3,400 dialogues with 17 target models, EYT-Bench reveals four findings that previous benchmarks miss: (i) state-of-the-art closed and open-source models are statistically indistinguishable on subjective dimensions, but separate by up to 9x on objective intent-tracking; (ii) reasoning is a phase transition for objective tracking on long-context personas but is essentially flat on subjective scores; (iii) persona format strongly affects trajectory spread, FICR (final-intent completion rate) saturates above 0.95 on Nemotron-USA but ranges from 0.53 to 0.88 on PersonaMem-v2; and (iv) the warm-up effect is observed in 16 of 17 models.

Benchmarks are crucial for evaluating progress in robotics and embodied AI. However, a significant gap exists between benchmarks designed for high-level language instruction following, which often assume perfect low-level execution, and those for low-level robot control, which rely on simple, one-step commands. This disconnect prevents a comprehensive evaluation of integrated systems where both task planning and physical execution are critical. To address this, we propose Kitchen-R, a novel benchmark that unifies the evaluation of task planning and low-level control within a simulated kitchen environment. Built as a digital twin using the Isaac Sim simulator and featuring more than 500 complex language instructions, Kitchen-R supports a mobile manipulator robot. We provide baseline methods for our benchmark, including a task-planning strategy based on a vision-language model and a low-level control policy based on diffusion policy. We also provide a trajectory collection system. Our benchmark offers a flexible framework for three evaluation modes: independent assessment of the planning module, independent assessment of the control policy, and, crucially, an integrated evaluation of the whole system. Kitchen-R bridges a key gap in embodied AI research, enabling more holistic and realistic benchmarking of language-guided robotic agents.

Conditional image editing aims to modify a source image according to textual prompts and optional reference guidance. Such editing is crucial in scenarios requiring strict structural control (i.e., anomaly insertion in driving scenes and complex human pose transformation). Despite recent advances in large-scale editing models (i.e., Seedream, Nano Banana, etc), most approaches rely on single-step generation. This paradigm often lacks explicit quality control, may introduce excessive deviation from the original image, and frequently produces structural artifacts or environment-inconsistent modifications, typically requiring manual prompt tuning to achieve acceptable results. We propose CAMEO, a structured multi-agent framework that reformulates conditional editing as a quality-aware, feedback-driven process rather than a one-shot generation task. CAMEO decomposes editing into coordinated stages of planning, structured prompting, hypothesis generation, and adaptive reference grounding, where external guidance is invoked only when task complexity requires it. To overcome the lack of intrinsic quality control in existing methods, evaluation is embedded directly within the editing loop. Intermediate results are iteratively refined through structured feedback, forming a closed-loop process that progressively corrects structural and contextual inconsistencies. We evaluate CAMEO on anomaly insertion and human pose switching tasks. Across multiple strong editing backbones and independent evaluation models, CAMEO consistently achieves 20\% more win rate on average compared to multiple state-of-the-art models, demonstrating improved robustness, controllability, and structural reliability in conditional image editing.

Automated evaluation of tool-using large language model (LLM) agents is widely assumed to be reliable, but this assumption has rarely been validated against human annotation. We introduce AgentProp-Bench, a 2,000-task benchmark with 2,300 traces across four domains, nine production LLMs, and a 100-label human-validated subset. We quantify judge reliability, characterize error propagation, and evaluate a runtime mitigation. Substring-based judging agrees with human annotation at kappa=0.049 (chance-level); a three-LLM ensemble reaches kappa=0.432 (moderate) with a conservative bias. Under validated evaluation, a parameter-level injection propagates to a wrong final answer with human-calibrated probability approximately 0.62 (range 0.46-0.73 across models). Rejection (catching bad parameters) and recovery (correcting after acceptance) are independent model capabilities (Spearman rho=0.126, p=0.747). A tuned runtime interceptor reduces hallucination on GPT-4o-mini by 23.0 percentage points under a concurrent n=600 control, but shows no significant effect on Gemini-2.0-Flash, whose aggressive parameter rejection eliminates the target failure mode. All code, data, traces, and human labels are released at https://github.com/bhaskargurram-ai/agenthallu-bench.

Public leaderboards for coding agents typically rank systems by model name and pass rate, while the surrounding harness (the scaffold that issues tools, manages context, and decides when to stop) is often under-specified. Model-to-model comparison is valid when the harness is fixed; when it varies, performance and efficiency conflate model and scaffold effects. We evaluate Qwen 3.6 Plus and MiniMax M2.5 across three open-source harnesses (Goose, OpenCode, OpenHands-SDK) on a stratified 50-task subset of Terminal-Bench Pro. Harness choice induces up to a 40x difference in tokens per solved task, while paired within-model pass-rate differences remain 0-8 percentage points (95% paired-task bootstrap CIs include zero except for the largest gap). Failure fingerprints replicate across models (REASON for Goose, VERIFY/MAX_TURNS for OpenHands-SDK, idle-loop/TIME for OpenCode), indicating harness-level biases that are largely model-independent. For human-centered coding-agent evaluation, model name alone is an incomplete comparison unit: harness-model pairs determine real-world cost, latency, and oversight burden; no-action turns are a per-task wait tax, not just a token tax. We therefore recommend selecting harness-model pairs by pass rate under token/latency budgets, and reporting token usage, latency, and full harness specifications alongside any model comparison. We release anonymized configs, raw trial logs, aggregated snapshots, and analysis scripts.

Face recognition templates are compact identity representations, yet they also encode rich semantic information about facial appearance. Prior work has shown that templates can be inverted to images or indirectly manipulated through image-editing pipelines, but direct semantic editing in template space remains largely unexplored. Existing interpretability methods for face recognition often rely on manual neuron inspection or predefined attribute labels, limiting scalability and semantic flexibility. To address this gap, we propose SCOUT (Semantic Concept Discovery for Open-VocabUlary Editing of Face Recognition Templates), an end-to-end framework for discovering and directly manipulating semantic concepts in face recognition templates using mechanistic interpretability. SCOUT learns sparse template representations, generates semantic hypotheses for latent features from natural-language descriptions, and validates their stability. The resulting features act as controllable semantic directions for direct editing, avoiding costly edit--re-encode pipelines. Experiments with face recognition models using CNN, ViT, and Swin backbones show that SCOUT discovers interpretable concepts beyond standard attribute labels and enables controllable, identity-aware template manipulation with negligible impact on identity matching. We further show that edited templates can subsequently be decoded with independent inversion models for visualization and evaluation.

Training robust and generalizable reward models for human visual preferences is essential for aligning text-to-image and text-to-video generative models with human intent. However, current reward models often fail to generalize, and supervised fine-tuning leads to memorization, demanding complex annotation pipelines. While reinforcement learning (RL), specifically Group Relative Policy Optimization (GRPO), improves generalization, we uncover a key failure mode: a significant drop in reasoning accuracy occurs when a model's reasoning trace contradicts that of an independent, frozen vision-language model ("listener") evaluating the same output. To address this, we introduce a listener-augmented GRPO framework. Here, the listener re-evaluates the reasoner's chain-of-thought to provide a dense, calibrated confidence score, shaping the RL reward signal. This encourages the reasoner not only to answer correctly, but to produce explanations that are persuasive to an independent model. Our listener-shaped reward scheme achieves best accuracy on the ImageReward benchmark (67.4%), significantly improves out-of-distribution (OOD) performance on a large-scale human preference dataset (1.2M votes, up to +6% over naive reasoner), and reduces reasoning contradictions compared to strong GRPO and SFT baselines. These results demonstrate that listener-based rewards provide a scalable, data-efficient path to aligning vision-language models with nuanced human preferences. We will release our reasoning model here: https://huggingface.co/alexgambashidze/qwen2.5vl_image_preference_reasoner.

Large Language Models (LLMs) drive scientific question-answering on modern search engines, yet their evaluation robustness remains underexplored. We introduce YESciEval, an open-source framework that combines fine-grained rubric-based assessment with reinforcement learning to mitigate optimism bias in LLM evaluators. We release multidisciplinary scienceQ&A datasets, including adversarial variants, with evaluation scores from multiple LLMs. Independent of proprietary models and human feedback, our approach enables scalable, cost-free evaluation. By advancing reliable LLM-as-a-judge models, this work supports AI alignment and fosters robust, transparent evaluation essential for scientific inquiry.

We introduce the Korean Canonical Legal Benchmark (KCL), a benchmark designed to assess language models' legal reasoning capabilities independently of domain-specific knowledge. KCL provides question-level supporting precedents, enabling a more faithful disentanglement of reasoning ability from parameterized knowledge. KCL consists of two components: (1) KCL-MCQA, multiple-choice problems of 283 questions with 1,103 aligned precedents, and (2) KCL-Essay, open-ended generation problems of 169 questions with 550 aligned precedents and 2,739 instance-level rubrics for automated evaluation. Our systematic evaluation of 30+ models shows large remaining gaps, particularly in KCL-Essay, and that reasoning-specialized models consistently outperform their general-purpose counterparts. We release all resources, including the benchmark dataset and evaluation code, at https://github.com/lbox-kr/kcl.

Safety monitors screen prompts sent to deployed language models, flagging harmful requests so they are never answered. They are evaluated by recall against harmfulness labels, but a catch only prevents harm if the model would otherwise have complied. We measure the difference directly: we sample repeated responses from the target model, call a harmful prompt elicitable if the model complies at least once, and report monitor recall separately on elicitable and non-elicitable prompts. Across six monitor configurations and three model families, spanning activation probes, fine-tuned text guards, and a 120B policy-conditioned reasoning classifier, recall on elicitable prompts falls 0.22 to 0.38 below recall on non-elicitable prompts at a fixed false positive rate. The prompts a monitor misses are 2.8 to 5.6 times more likely to be complied with than the prompts it catches. The gap replicates across three model families and appears also in text-only monitors entirely independent of the target model. This suggests that standard recall may overstate the protection monitors provide in practice, and that monitors should be evaluated against what their models will actually answer.

The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings. Prior work has identified test-time contextual cues, such as hypothetical scenarios, as a source of verbalized evaluation awareness and subsequent behavioral shift. In this paper, we investigate a potential explanation of this phenomenon: evaluation meta-knowledge, defined as parametric knowledge about the structural traits that characterize evaluations. Similar to dataset contamination, where benchmark exposure leads to higher performance through memorization, we hypothesize that models trained on texts describing evaluation practices may implicitly learn to recognize and respond to evaluation-like contexts, for instance, through exposure to scientific articles or social media posts about AI benchmarking. To test this, we fine-tune models on synthetic documents describing evaluation traits such as verifiable structures or moral dilemmas. Evaluating this fine-tuned model on six safety benchmarks, we find that it is significantly safer than the base model and control model. This behavioral shift persists even when restricting the analysis to responses lacking explicit verbalization of evaluation awareness. Our results demonstrate that evaluation meta-knowledge may inflate safety benchmark performance, introducing a novel confounder that is independent of explicit memorization or verbalized evaluation awareness, thus, challenging to detect. These findings have important implications for the design and interpretation of AI safety evaluations. Our code and models are available at https://github.com/compass-group-tue/arxiv2026_evaluation_meta_knowledge.

Probabilistic generative models can be used for compression, denoising, inpainting, texture synthesis, semi-supervised learning, unsupervised feature learning, and other tasks. Given this wide range of applications, it is not surprising that a lot of heterogeneity exists in the way these models are formulated, trained, and evaluated. As a consequence, direct comparison between models is often difficult. This article reviews mostly known but often underappreciated properties relating to the evaluation and interpretation of generative models with a focus on image models. In particular, we show that three of the currently most commonly used criteria---average log-likelihood, Parzen window estimates, and visual fidelity of samples---are largely independent of each other when the data is high-dimensional. Good performance with respect to one criterion therefore need not imply good performance with respect to the other criteria. Our results show that extrapolation from one criterion to another is not warranted and generative models need to be evaluated directly with respect to the application(s) they were intended for. In addition, we provide examples demonstrating that Parzen window estimates should generally be avoided.

This paper explores the use of open generative Large Language Models (LLMs) for annotation tasks in the social sciences. The study highlights the challenges associated with proprietary models, such as limited reproducibility and privacy concerns, and advocates for the adoption of open (source) models that can be operated on independent devices. Two examples of annotation tasks, sentiment analysis in tweets and identification of leisure activities in childhood aspirational essays are provided. The study evaluates the performance of different prompting strategies and models (neural-chat-7b-v3-2, Starling-LM-7B-alpha, openchat_3.5, zephyr-7b-alpha and zephyr-7b-beta). The results indicate the need for careful validation and tailored prompt engineering. The study highlights the advantages of open models for data privacy and reproducibility.

We present Mercury, a new generation of commercial-scale large language models (LLMs) based on diffusion. These models are parameterized via the Transformer architecture and trained to predict multiple tokens in parallel. In this report, we detail Mercury Coder, our first set of diffusion LLMs designed for coding applications. Currently, Mercury Coder comes in two sizes: Mini and Small. These models set a new state-of-the-art on the speed-quality frontier. Based on independent evaluations conducted by Artificial Analysis, Mercury Coder Mini and Mercury Coder Small achieve state-of-the-art throughputs of 1109 tokens/sec and 737 tokens/sec, respectively, on NVIDIA H100 GPUs and outperform speed-optimized frontier models by up to 10x on average while maintaining comparable quality. We discuss additional results on a variety of code benchmarks spanning multiple languages and use-cases as well as real-world validation by developers on Copilot Arena, where the model currently ranks second on quality and is the fastest model overall. We also release a public API at https://platform.inceptionlabs.ai/ and free playground at https://chat.inceptionlabs.ai

Large Language Models (LLMs) are increasingly being used to simulate human-like decision making in agent-based financial market models (ABMs). As models become more powerful and accessible, researchers can now incorporate individual LLM decisions into ABM environments. However, integration may introduce inherent biases that need careful evaluation. In this paper we test three state-of-the-art GPT models for bias using two model sampling approaches: one-shot and few-shot API queries. We observe significant variations in distributions of outputs between specific models, and model sub versions, with GPT-4o-Mini-2024-07-18 showing notably better performance (32-43% yes responses) compared to GPT-4-0125-preview's extreme bias (98-99% yes responses). We show that sampling methods and model sub-versions significantly impact results: repeated independent API calls produce different distributions compared to batch sampling within a single call. While no current GPT model can simultaneously achieve a uniform distribution and Markovian properties in one-shot testing, few-shot sampling can approach uniform distributions under certain conditions. We explore the Temperature parameter, providing a definition and comparative results. We further compare our results to true random binary series and test specifically for the common human bias of Negative Recency - finding LLMs have a mixed ability to 'beat' humans in this one regard. These findings emphasise the critical importance of careful LLM integration into ABMs for financial markets and more broadly.

Microsaccades are small, involuntary eye movements vital for visual perception and neural processing. Traditional microsaccade studies typically use eye trackers or frame-based analysis, which, while precise, are costly and limited in scalability and temporal resolution. Event-based sensing offers a high-speed, low-latency alternative by capturing fine-grained spatiotemporal changes efficiently. This work introduces a pioneering event-based microsaccade dataset to support research on small eye movement dynamics in cognitive computing. Using Blender, we render high-fidelity eye movement scenarios and simulate microsaccades with angular displacements from 0.5 to 2.0 degrees, divided into seven distinct classes. These are converted to event streams using v2e, preserving the natural temporal dynamics of microsaccades, with durations ranging from 0.25 ms to 2.25 ms. We evaluate the dataset using Spiking-VGG11, Spiking-VGG13, and Spiking-VGG16, and propose Spiking-VGG16Flow, an optical-flow-enhanced variant implemented in SpikingJelly. The models achieve around 90 percent average accuracy, successfully classifying microsaccades by angular displacement, independent of event count or duration. These results demonstrate the potential of spiking neural networks for fine motion recognition and establish a benchmark for event-based vision research. The dataset, code, and trained models will be publicly available at https://waseemshariff126.github.io/microsaccades/ .

An ordinary-looking background detail can turn a correct model decision into a confident mistake. We demonstrate this fragility in four decision systems, including Jev, across seven datasets covering knowledge, reasoning, and tool routing. Within 64 accepted target evaluations per decision, we uncover short context additions that redirect 61.4%-73.2% of each system's initially correct decisions toward a wrong option fixed in advance. The additions supply background or procedural information rather than explicit answer-selection instructions, leaving the original question and choices intact. We construct them through probability-guided context optimization, which uses shifts in the option distribution to refine surrounding text under naturalness and answer-preservation constraints. Redirection affects initially confident decisions, often produces high-confidence wrong choices, and transfers across models. In blinded human evaluation, 91.6% of 250 sampled successful contexts are judged natural, answer-preserving, and free of decisive answer-changing evidence by a majority of three independent annotators. These findings expose a weakness in current decision models: context that looks entirely compatible with an input can redirect the choices that agents, routers, and evaluators rely on.

Data-driven materials discovery requires large-scale experimental datasets, yet most of the information remains trapped in unstructured literature. Existing extraction efforts often focus on a limited set of features and have not addressed the integrated composition-processing-microstructure-property relationships essential for understanding materials behavior, thereby posing challenges for building comprehensive databases. To address this gap, we propose a multi-stage information extraction pipeline powered by large language models, which captures 47 features spanning composition, processing, microstructure, and properties exclusively from experimentally reported materials. The pipeline integrates iterative extraction with source tracking to enhance both accuracy and reliability. Evaluations at the feature level (independent attributes) and tuple level (interdependent features) yielded F1 scores around 0.96. Compared with single-pass extraction without source tracking, our approach improved F1 scores of microstructure category by 10.0% (feature level) and 13.7% (tuple level), and reduced missed materials from 49 to 13 out of 396 materials in 100 articles on precipitate-containing multi-principal element alloys (miss rate reduced from 12.4% to 3.3%). The pipeline enables scalable and efficient literature mining, producing databases with high precision, minimal omissions, and zero false positives. These datasets provide trustworthy inputs for machine learning and materials informatics, while the modular design generalizes to diverse material classes, enabling comprehensive materials information extraction.

Large language models (LLMs) are increasingly used to access legal information. Yet, their deployment in multilingual legal settings is constrained by unreliable retrieval and the lack of domain-adapted, open-embedding models. In particular, existing multilingual legal corpora are not designed for semantic retrieval, and PDF-based legislative sources introduce substantial noise due to imperfect text extraction. To address these challenges, we introduce LEMUR, a large-scale multilingual corpus of EU environmental legislation constructed from 24,953 official EUR-Lex PDF documents covering 25 languages. We quantify the fidelity of PDF-to-text conversion by measuring lexical consistency against authoritative HTML versions using the Lexical Content Score (LCS). Building on LEMUR, we fine-tune three state-of-the-art multilingual embedding models using contrastive objectives in both monolingual and bilingual settings, reflecting realistic legal-retrieval scenarios. Experiments across low- and high-resource languages demonstrate that legal-domain fine-tuning consistently improves Top-k retrieval accuracy relative to strong baselines, with particularly pronounced gains for low-resource languages. Cross-lingual evaluations show that these improvements transfer to unseen languages, indicating that fine-tuning primarily enhances language-independent, content-level legal representations rather than language-specific cues. We publish code\href{https://github.com/nargesbh/eur_lex{GitHub Repository}} and data\href{https://huggingface.co/datasets/G4KMU/LEMUR{Hugging Face Dataset}}.

We present CaloClouds3, a model for the fast simulation of photon showers in the barrel of a high granularity detector. This iteration demonstrates for the first time how a pointcloud model can employ angular conditioning to replicate photons at all incident angles. Showers produced by this model can be used across the whole detector barrel, due to specially produced position agnostic training data. With this flexibility, the model is usable in a full simulation and reconstruction chain, which offers a further handle for evaluating physics performance of the model. As inference time is a crucial consideration for a generative model, the pre-processing and hyperparameters are aggressively optimised, achieving a speed up factor of two orders of magnitude over Geant4 at inference.

Text-to-SQL benchmarks play a crucial role in evaluating the progress made in the field and the ranking of different models. However, accurately matching a model-generated SQL query to a reference SQL query in a benchmark fails for various reasons, such as underspecified natural language queries, inherent assumptions in both model-generated and reference queries, and the non-deterministic nature of SQL output under certain conditions. In this paper, we conduct an extensive study of several prominent cross-domain text-to-SQL benchmarks and re-evaluate some of the top-performing models within these benchmarks, by both manually evaluating the SQL queries and rewriting them in equivalent expressions. Our evaluation reveals that attaining a perfect performance on these benchmarks is unfeasible due to the multiple interpretations that can be derived from the provided samples. Furthermore, we find that the true performance of the models is underestimated and their relative performance changes after a re-evaluation. Most notably, our evaluation reveals a surprising discovery: a recent GPT4-based model surpasses the gold standard reference queries in the Spider benchmark in our human evaluation. This finding highlights the importance of interpreting benchmark evaluations cautiously, while also acknowledging the critical role of additional independent evaluations in driving advancements in the field.

Specialized clinical AI assistants are rapidly entering medical practice, often framed as safer or more reliable than general-purpose large language models (LLMs). Yet, unlike frontier models, these clinical tools are rarely subjected to independent, quantitative evaluation, creating a critical evidence gap despite their growing influence on diagnosis, triage, and guideline interpretation. We assessed two widely deployed clinical AI systems (OpenEvidence and UpToDate Expert AI) against three state-of-the-art generalist LLMs (GPT-5, Gemini 3 Pro, and Claude Sonnet 4.5) using a 1,000-item mini-benchmark combining MedQA (medical knowledge) and HealthBench (clinician-alignment) tasks. Generalist models consistently outperformed clinical tools, with GPT-5 achieving the highest scores, while OpenEvidence and UpToDate demonstrated deficits in completeness, communication quality, context awareness, and systems-based safety reasoning. These findings reveal that tools marketed for clinical decision support may often lag behind frontier LLMs, underscoring the urgent need for transparent, independent evaluation before deployment in patient-facing workflows.

As the usage of large language models for problems outside of simple text understanding or generation increases, assessing their abilities and limitations becomes crucial. While significant progress has been made in this area over the last few years, most research has focused on benchmarking English, leaving other languages underexplored. This makes evaluating the reasoning and robustness level of language models in Ukrainian particularly challenging. The purpose of this work is to establish a comprehensive benchmark for the reasoning capabilities evaluation of large language models in the Ukrainian language. This paper presents the ZNO-Eval benchmark based on real exam tasks from Ukraine's standardized educational testing system: the External Independent Evaluation and the National Multi-subject Test. With single-answer options, multiple-choice, matching, and open-ended questions from diverse subjects, including Ukrainian language, mathematics, history, and geography, this dataset paves the way toward a thorough analysis of reasoning capabilities across different domains and complexities. Evaluation of several well-known language models, such as GPT-3.5-Turbo, GPT-4o, GPT-4-Turbo, Mistral Large, Claude 3 Opus, and Gemini-1.5 Pro on this benchmark demonstrated the superiority of GPT-4o in both common knowledge reasoning and intricate language tasks. At the same time, Gemini Pro and GPT-4 Turbo excelled in the arithmetic domain, leading in single-answer and open-ended math problems. While all models were close to max performance in text-only common knowledge tasks like history and geography, there still is a gap for Ukrainian language and math, thus highlighting the importance of developing specialized language benchmarks for more accurate assessments of model capabilities and limitations across different languages and contexts.

The handling of probabilities in the form of uncertainty or partial information is an essential task for LLMs in many settings and applications. A common approach to evaluate an LLM's probabilistic reasoning capabilities is to assess its ability to answer questions pertaining to probability through the use of multiple-choice questions (MCQs). However, this paradigm, which we refer to as explicit probabilistic reasoning, has been shown in the literature to yield significant limitations (e.g., sensitivity to answer ordering). In this work, we introduce an alternative approach, named implicit probabilistic reasoning, which evaluates the models' ability to integrate probabilistic reasoning into their text generation process. To achieve this, we rephrase MCQs as text-completion scenarios with a determined set of outcomes and compare the model's next-token probability assignments to the true likelihood of the outcomes. In line with previous work, we find that models exhibit solid performance in their explicit probabilistic reasoning (i.e., answers to MCQs). However, during text completion (i.e., implicit probabilistic reasoning), where the same information must be taken into account to generate text, the models' predictions often significantly diverge from the known ground truth. For instance, our evaluation method reveals that implicit probabilistic reasoning is improperly influenced by many factors, such as independent prior events, partial observations about a result, or statistical background information. All of these issues can cause erroneous results to be produced in text generation, which are not detected by conventional MCQ-based evaluation.

Cancer survival prediction supports treatment planning, risk stratification, and follow-up management. Existing methods use structured clinical variables, whole-slide images, genomic profiles, or multimodal inputs, while patient reports remain underexplored. We study report-centric survival prediction using reports that organize pathological, clinical, and molecular evidence. Large language models (LLMs) can reason over such reports, but case-wise time regression introduces two mismatches. First, a formulation mismatch arises because survival evaluation depends on ordering comparable patients, whereas independent time predictions do not enforce ranking consistency. Second, a supervision mismatch arises because a censored patient's observed time indicates survival beyond that point and cannot serve as an exact regression target, although it still implies orderings relative to patients who died earlier. To address these mismatches, we propose CACSurv, a Concordance-Aligned Comparative framework for report-centric survival prediction. CACSurv reformulates survival modeling as mini-cohort comparative reasoning, where an LLM predicts relative prognostic orderings. We introduce concordance-aligned rewards derived from comparable relations under right censoring, enabling censored outcomes to provide ranking supervision without exact event-time targets. At inference, Monte Carlo Reference Aggregation compares each patient with sampled references and aggregates positions into a cohort-level ranking. We establish TCGA-SurvReport, a benchmark covering six TCGA cancer cohorts. CACSurv achieves the highest C-index on all six cohorts and an average C-index of 0.722, outperforming the strongest published survival model by 6.5 percentage points and the strongest LLM time-regression baseline by 4.2 percentage points. Our code, models, and dataset will be available at https://github.com/xmed-lab/CACSurv.

Discrete diffusion models with absorbing processes have shown promise in language modeling. The key quantities to be estimated are the ratios between the marginal probabilities of two transitive states at all timesteps, called the concrete score. In this paper, we reveal that the concrete score in absorbing diffusion can be expressed as conditional probabilities of clean data, multiplied by a time-dependent scalar in an analytic form. Motivated by this finding, we propose reparameterized absorbing discrete diffusion (RADD), a dedicated diffusion model without time-condition that characterizes the time-independent conditional probabilities. Besides its simplicity, RADD can reduce the number of function evaluations (NFEs) by caching the output of the time-independent network when the noisy sample remains unchanged in a sampling interval. Empirically, RADD is up to 3.5 times faster while achieving similar performance with the strongest baseline. Built upon the new perspective of conditional distributions, we further unify absorbing discrete diffusion and any-order autoregressive models (AO-ARMs), showing that the upper bound on the negative log-likelihood for the diffusion model can be interpreted as an expected negative log-likelihood for AO-ARMs. Further, our RADD models achieve SOTA performance among diffusion models on 5 zero-shot language modeling benchmarks (measured by perplexity) at the GPT-2 scale. Our code is available at https://github.com/ML-GSAI/RADD.

As AI agents begin to automate AI R&D, we need ways to assess whether their outputs are safe to deploy, even when the agents themselves may be untrusted. AI control offers one such approach: rather than trusting the agent, it treats it as a potential adversary and uses a monitor to detect covert sabotage before deployment. We evaluate AI control for automated AI R&D with ResearchArena, a framework spanning four long-horizon tasks: safety post-training, capabilities post-training, CUDA-kernel optimization, and inference-server optimization. Because the deliverable in AI R&D is an artifact that will be deployed and run, we pair each main task with two kinds of hidden side task: an embedded side task that sabotages the submitted model, adapter, kernel, or server itself, and an independent side task that takes a covert out-of-scope action in the sandbox while leaving the artifact honest. We evaluate frontier agents at both sabotage and monitoring, across four monitor types that vary along two axes: whether the monitor may execute and probe the produced artifact, and whether it sees the agent's chain-of-thought. We find that sabotage hidden in the training data is the hardest to catch, flagged fewer than half the time. Letting the monitor run experiments on the artifact, rather than only read the trajectory, helps, but it is not enough: monitors still miss embedded sabotage by inspecting only the surface, by noticing the anomaly but explaining it away, or by probing the artifact with the wrong test. We release ResearchArena as a modular framework for evaluating sabotage and control in automated AI R&D.

Multimodal Large Language Models (MLLMs) are strong perceivers of images and video. We ask how far that reach extends into acting: dropping an MLLM directly into a drone's control loop, with its entire action space declared solely in the prompt. Recent systems approach this setting but increasingly narrow the model's decision-making. We widen it back. We introduce DroneCATS-Agent, an architecture where the MLLM is a swappable component, and DroneCATS, a benchmark treating the model as the independent variable. Beyond merely flying toward a pixel, our agent entrusts the model to yaw and search, deliberate when unsure, and self-declare arrival---all without fine-tuning or function-calling schemas. Evaluating frontier and open models across four core capabilities---approaching a visible target, tracking a moving one, searching outside the initial view, and commanding a multi-drone fleet---reveals that even the simplest embodied settings are far from solved. Crucially, to identify what breaks first at the edge, our roster scales down to 2B parameters. The findings expose a stark paradox: it is not the flying that fails. Small open models often navigate into the success radius more reliably than frontier models, yet lose the episode by declaring arrival prematurely or not at all. Multi-drone commanding amplifies this divide, with small models failing by blindly copying a single coordinate across distinct views. Viewed as vision-language-action agents, the models' spatial perception holds up, but their action protocol does not. What separates a deployable edge model from a frontier model is not navigation, but the discipline to sustain a declared protocol and emit the correct terminating action. The open problem is closing this gap at onboard compute costs---yielding a fast model that plans persistently and knows exactly when it is done---and DroneCATS is built to measure that distance.

Logical reasoning is a critical component of Large Language Models (LLMs), and substantial research efforts in recent years have aimed to enhance their deductive reasoning capabilities. However, existing deductive reasoning benchmarks, which are crucial for evaluating and advancing LLMs, are inadequate due to their lack of task complexity, presence of prior knowledge as a confounder, and superficial error analysis. To address these deficiencies, we introduce JustLogic, a synthetically generated deductive reasoning benchmark designed for rigorous evaluation of LLMs. JustLogic is (i) highly complex, capable of generating a diverse range of linguistic patterns, vocabulary, and argument structures; (ii) prior knowledge independent, eliminating the advantage of models possessing prior knowledge and ensuring that only deductive reasoning is used to answer questions; and (iii) capable of in-depth error analysis on the heterogeneous effects of reasoning depth and argument form on model accuracy. Our experimental results on JustLogic reveal that most state-of-the-art (SOTA) LLMs perform significantly worse than the human average, demonstrating substantial room for model improvement. All code and data are available at https://github.com/michaelchen-lab/JustLogic

Despite substantial investment in safety alignment, the vulnerability of large language models to sophisticated multi-turn adversarial attacks remains poorly characterized, and whether model scale or inference mode affects robustness is unknown. This study employed the TEMPEST multi-turn attack framework to evaluate ten frontier models from eight vendors across 1,000 harmful behaviors, generating over 97,000 API queries across adversarial conversations with automated evaluation by independent safety classifiers. Results demonstrated a spectrum of vulnerability: six models achieved 96% to 100% attack success rate (ASR), while four showed meaningful resistance, with ASR ranging from 42% to 78%; enabling extended reasoning on identical architecture reduced ASR from 97% to 42%. These findings indicate that safety alignment quality varies substantially across vendors, that model scale does not predict adversarial robustness, and that thinking mode provides a deployable safety enhancement. Collectively, this work establishes that current alignment techniques remain fundamentally vulnerable to adaptive multi-turn attacks regardless of model scale, while identifying deliberative inference as a promising defense direction.

Recent work on chain-of-thought (CoT) faithfulness reports single aggregate numbers (e.g., DeepSeek-R1 acknowledges hints 39% of the time), implying that faithfulness is an objective, measurable property of a model. This paper demonstrates that it is not. Three classifiers (a regex-only detector, a two-stage regex-plus-LLM pipeline, and an independent Claude Sonnet 4 judge) are applied to 10,276 influenced reasoning traces from 12 open-weight models spanning 9 families and 7B to 1T parameters. On identical data, these classifiers produce overall faithfulness rates of 74.4%, 82.6%, and 69.7%, respectively, with non-overlapping 95% confidence intervals. Per-model gaps range from 2.6 to 30.6 percentage points; all are statistically significant (McNemar's test, p < 0.001). The disagreements are systematic, not random: inter-classifier agreement measured by Cohen's kappa ranges from 0.06 ("slight") for sycophancy hints to 0.42 ("moderate") for grader hints, and the asymmetry is pronounced: for sycophancy, 883 cases are classified as faithful by the pipeline but unfaithful by the Sonnet judge, while only 2 go the other direction. Classifier choice can also reverse model rankings: Qwen3.5-27B ranks 1st under the pipeline but 7th under the Sonnet judge; OLMo-3.1-32B moves in the opposite direction, from 9th to 3rd. The root cause is that different classifiers operationalize related faithfulness constructs at different levels of stringency (lexical mention versus epistemic dependence), and these constructs yield divergent measurements on the same behavior. These results demonstrate that published faithfulness numbers cannot be meaningfully compared across studies that use different classifiers, and that future evaluations should report sensitivity ranges across multiple classification methodologies rather than single point estimates.

Fake news detection becomes particularly challenging in real-time scenarios, where emerging events often lack sufficient supporting evidence. Existing approaches often rely heavily on external evidence and therefore struggle to generalize under evidence scarcity. To address this issue, we propose Evaluation-Aware Selection of Experts (EASE), a novel framework for real-time fake news detection that dynamically adapts its decision-making process according to the assessed sufficiency of available evidence. EASE introduces a sequential evaluation mechanism comprising three independent perspectives: (1) Evidence-based evaluation, which assesses evidence and incorporates it into decision-making only when the evidence is sufficiently supportive; (2) Reasoning-based evaluation, which leverages the world knowledge of large language models (LLMs) and applies them only when their reliability is adequately established; and (3) Sentiment-based fallback, which integrates sentiment cues when neither evidence nor reasoning is reliable. To enhance the accuracy of evaluation processes, EASE employs instruction tuning with pseudo labels to guide each evaluator in justifying its perspective-specific knowledge through interpretable reasoning. Furthermore, the expert modules integrate the evaluators' justified assessments with the news content to enable evaluation-aware decision-making, thereby enhancing overall detection accuracy. Moreover, we introduce RealTimeNews-25, a new benchmark comprising recent news for evaluating model generalization on emerging news with limited evidence. Extensive experiments demonstrate that EASE not only achieves state-of-the-art performance across multiple benchmarks, but also significantly improves generalization to real-time news. The code and dataset are available: https://github.com/wgyhhhh/EASE.

Neutron/gamma discrimination has been intensively researched in recent years, due to its unique scientific value and widespread applications. With the advancement of detection materials and algorithms, nowadays we can achieve fairly good discrimination. However, further improvements rely on better utilization of detector raw signals, especially energy-independent pulse characteristics. We begin by discussing why figure-of-merit (FoM) is not a comprehensive criterion for high-precision neutron/gamma discriminators, and proposing a new evaluation method based on adversarial sampling. Inspired by frequency-domain analysis in existing literature, parametric linear/nonlinear models with minimum complexity are created, upon the discrete spectrum, with tunable parameters just as neural networks. We train the models on an open-source neutron/gamma dataset (CLYC crystals with silicon photomultipliers) preprocessed by charge normalization to discover and exploit energy-independent features. The performance is evaluated on different sampling rates and noise levels, in comparison with the frequency classification index and conventional methods. The frequency-domain parametric models show higher accuracy and better adaptability to variations of data integrity than other discriminators. The proposed method is also promising for online inference on economical hardware and portable devices.

This paper presents FoFo, a pioneering benchmark for evaluating large language models' (LLMs) ability to follow complex, domain-specific formats, a crucial yet underexamined capability for their application as AI agents. Despite LLMs' advancements, existing benchmarks fail to assess their format-following proficiency adequately. FoFo fills this gap with a diverse range of real-world formats and instructions, developed through an AI-Human collaborative method. Our evaluation across both open-source (e.g., Llama 2, WizardLM) and closed-source (e.g., GPT-4, PALM2, Gemini) LLMs highlights three key findings: open-source models significantly lag behind closed-source ones in format adherence; LLMs' format-following performance is independent of their content generation quality; and LLMs' format proficiency varies across different domains. These insights suggest the need for specialized tuning for format-following skills and highlight FoFo's role in guiding the selection of domain-specific AI agents. FoFo is released here at https://github.com/SalesforceAIResearch/FoFo.

Recently, Direct Preference Optimization (DPO) has extended its success from aligning large language models (LLMs) to aligning text-to-image diffusion models with human preferences. Unlike most existing DPO methods that assume all diffusion steps share a consistent preference order with the final generated images, we argue that this assumption neglects step-specific denoising performance and that preference labels should be tailored to each step's contribution. To address this limitation, we propose Step-aware Preference Optimization (SPO), a novel post-training approach that independently evaluates and adjusts the denoising performance at each step, using a step-aware preference model and a step-wise resampler to ensure accurate step-aware supervision. Specifically, at each denoising step, we sample a pool of images, find a suitable win-lose pair, and, most importantly, randomly select a single image from the pool to initialize the next denoising step. This step-wise resampler process ensures the next win-lose image pair comes from the same image, making the win-lose comparison independent of the previous step. To assess the preferences at each step, we train a separate step-aware preference model that can be applied to both noisy and clean images. Our experiments with Stable Diffusion v1.5 and SDXL demonstrate that SPO significantly outperforms the latest Diffusion-DPO in aligning generated images with complex, detailed prompts and enhancing aesthetics, while also achieving more than 20x times faster in training efficiency. Code and model: https://rockeycoss.github.io/spo.github.io/

Large language models (LLMs) excel at answering pre-specified questions, yet their ability to navigate the open-ended, pre-conclusion stage of discovery remains largely unmeasured. We introduce Prospective Hypothesis Discovery (PHD), which asks models to autonomously construct grounded, discriminative, and testable hypothesis spaces from inconclusive evidence, including anomalous observations and fragmented records, to guide subsequent investigation. To evaluate this capability, we introduce HypoArena, comprising HypoData, a benchmark of 988 cases across six scientific and analytical domains, and HypoEval, an evaluation framework for open-ended hypothesis sets. To construct HypoData at scale, we propose Retrospective Context Regression, a Forge--Audit pipeline that reconstructs pre-conclusion contexts from completed expert documents by removing explicit conclusions, target hypotheses, and retrospective causal attributions while preserving the factual substrate. Because PHD admits multiple valid outputs, HypoEval combines bidirectional pairwise judgments with Bradley--Terry--Davidson aggregation for ranking and six-dimensional rubric scoring for diagnosis. Experiments on 15 frontier LLMs reveal clear capability stratification and model-dependent effects of structured analytical skills, with gains for several lower-performing models on HypoArena but regressions for other systems, including a top-performing model. Compared with absolute rubric scoring, arena evaluation resolves finer-grained differences among models, with aggregated rankings showing strong agreement with human experts and an independent judge. Together, these results support treating PHD as a distinct target for evaluating how LLMs formulate investigative directions when final conclusions are withheld. Our code and data are publicly available at github.com/SKYLENAGE-AI/HypoArena and github.com/SKYLENAGE-AI/HypoArena.

The recent trend of using Large Language Models (LLMs) as intelligent agents in real-world applications underscores the necessity for comprehensive evaluations of their capabilities, particularly in complex scenarios involving planning, creating, and using tools. However, existing benchmarks typically focus on simple synthesized queries that do not reflect real-world complexity, thereby offering limited perspectives in evaluating tool utilization. To address this issue, we present UltraTool, a novel benchmark designed to improve and evaluate LLMs' ability in tool utilization within real-world scenarios. UltraTool focuses on the entire process of using tools - from planning and creating to applying them in complex tasks. It emphasizes real-world complexities, demanding accurate, multi-step planning for effective problem-solving. A key feature of UltraTool is its independent evaluation of planning with natural language, which happens before tool usage and simplifies the task solving by mapping out the intermediate steps. Thus, unlike previous work, it eliminates the restriction of pre-defined toolset during planning. Through extensive experiments on various LLMs, we offer novel insights into the evaluation of capabilities of LLMs in tool utilization, thereby contributing a fresh perspective to this rapidly evolving field. The benchmark is publicly available at https://github.com/JoeYing1019/UltraTool.

Large language models (LLMs) exhibiting test-time scaling behavior, such as extended reasoning traces and self-verification, have demonstrated remarkable performance on complex, long-term reasoning tasks. However, the robustness of these reasoning behaviors remains underexplored. To investigate this, we conduct a systematic evaluation of multiple reasoning models across three scenarios: (1) problems augmented with lengthy, irrelevant context; (2) multi-turn conversational settings with independent tasks; and (3) problems presented as a subtask within a complex task. We observe an interesting phenomenon: reasoning models tend to produce much shorter reasoning traces (up to 50%) for the same problem under different context conditions compared to the traces produced when the problem is presented in isolation. A finer-grained analysis reveals that this compression is associated with a decrease in self-verification and uncertainty management behaviors, such as double-checking. While this behavioral shift does not compromise performance on straightforward problems, it might affect performance on more challenging tasks. We hope our findings draw additional attention to both the robustness of reasoning models and the problem of context management for LLMs and LLM-based agents.

Evaluating world models requires assessing both the quality of the worlds they generate and their consistency and responsiveness under exploration, interaction, and modification. We introduce HappyWorld-Bench, a comprehensive benchmark that evaluates whether generated worlds remain reliable as agents interact with them. Our design is built on a hierarchical capability framework of six world capabilities (W1-W6), from generative construction to unified world modeling, instantiated across three independent evaluation tracks: video world models, spatial world models, and embodied world models. HappyWorld-Bench comprises 1,138 video prompts, 300 spatial scenes, and 254 embodied test cases. Across all three tracks, we build and operate HappyWorld-Arena to organize human A/B comparisons and derive model-level Elo ratings, which complement newly designed automated metrics that capture behavioral correctness. We evaluate 14 video world models, 9 spatial systems, and 8 embodied candidates under this unified framework. Results reveal remaining reliability gaps across all three tracks: video models exhibit reduced consistency during extended rollouts and revisits, spatial models achieve at best 70.14% placement accuracy and 73.33% edit execution, and embodied models struggle to preserve state across multi-step actions and respond precisely to altered action conditions and physical rules. These findings highlight the need to evaluate world models not only by visual quality, but also by state consistency and the correctness of their responses to actions and interventions.

Attention-based models are appealing for multimodal processing because inputs from multiple modalities can be concatenated and fed to a single backbone network - thus requiring very little fusion engineering. The resulting representations are however fully entangled throughout the network, which may not always be desirable: in learning, contrastive audio-visual self-supervised learning requires independent audio and visual features to operate, otherwise learning collapses; in inference, evaluation of audio-visual models should be possible on benchmarks having just audio or just video. In this paper, we introduce Zorro, a technique that uses masks to control how inputs from each modality are routed inside Transformers, keeping some parts of the representation modality-pure. We apply this technique to three popular transformer-based architectures (ViT, Swin and HiP) and show that with contrastive pre-training Zorro achieves state-of-the-art results on most relevant benchmarks for multimodal tasks (AudioSet and VGGSound). Furthermore, the resulting models are able to perform unimodal inference on both video and audio benchmarks such as Kinetics-400 or ESC-50.

As Large Language Models (LLMs) have become more advanced, they have outpaced our abilities to accurately evaluate their quality. Not only is finding data to adequately probe particular model properties difficult, but evaluating the correctness of a model's freeform generation alone is a challenge. To address this, many evaluations now rely on using LLMs themselves as judges to score the quality of outputs from other LLMs. Evaluations most commonly use a single large model like GPT4. While this method has grown in popularity, it is costly, has been shown to introduce intramodel bias, and in this work, we find that very large models are often unnecessary. We propose instead to evaluate models using a Panel of LLm evaluators (PoLL). Across three distinct judge settings and spanning six different datasets, we find that using a PoLL composed of a larger number of smaller models outperforms a single large judge, exhibits less intra-model bias due to its composition of disjoint model families, and does so while being over seven times less expensive.

Large language models (LLMs) are widely used as reference-free evaluators via prompting, but this "LLM-as-a-Judge" paradigm is costly, opaque, and sensitive to prompt design. In this work, we investigate whether smaller models can serve as efficient evaluators by leveraging internal representations instead of surface generation. We uncover a consistent empirical pattern: small LMs, despite with weak generative ability, encode rich evaluative signals in their hidden states. This motivates us to propose the Semantic Capacity Asymmetry Hypothesis: evaluation requires significantly less semantic capacity than generation and can be grounded in intermediate representations, suggesting that evaluation does not necessarily need to rely on large-scale generative models but can instead leverage latent features from smaller ones. Our findings motivate a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge, a decoding-free evaluation strategy that probes internal model structure rather than relying on prompted output. We instantiate this paradigm through INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations. Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show that INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while offering a more efficient, reliable, and interpretable alternative for scalable evaluation.

LLM-based agents can act on behalf of a user to access cloud services, call tools, or invoke agents. At session start, the agent's permissions are set but remain static, and each request is evaluated independently, without considering prior actions. Within its permissions, an agent may act contrary to the delegated task, combine individually permitted actions into a prohibited outcome, or delegate authority to a sub-agent without limiting it. A prompt injection poses a risk only if the agent has authority to perform such actions; this is therefore a problem of authorization architecture, not just the model. The Agentic Principal Chain (APC) tracks delegated authority from one principal to the next. APC evaluates each request against the accumulated session state using six authorization checks. APC carries forward and restricts delegated scope and budgets. Using composition closure, APC checks requests against prior actions to prevent prohibited combinations and enforces the decision outside the model. We prove Blast Radius Monotonicity and Composition Soundness for APC implementations; Composition Soundness is limited to prohibited combinations under a complete restriction set and serialized admission. We evaluated 3,154 instances including InjecAgent, AgentDojo, and ASB. Our compromised-model evaluation tests APC independently of model behavior by inserting the ground-truth attack call after the first legitimate tool call. AgentDojo exfiltration fell from 75-100% to 0% across all four domains; APC blocked all 544 InjecAgent data-stealing cases. Intent binding reduced destruction from 38.6% to 4.0% and manipulation from 90.5% to 12.1%. Authorization latency was 0.24 ms at the 99th percentile on an idle host; across 949 AgentDojo task-injection pairs, utility was 8.6 and 13.9 percentage points lower in the two settings. Implementation, evaluation tools, and data are publicly available.

Empirical evaluation serves as the primary compass guiding research progress in foundation models. Despite a large body of work focused on training frontier vision-language models (VLMs), approaches to their evaluation remain nascent. To guide their maturation, we propose three desiderata that evaluations should satisfy: (1) faithfulness to the modality and application, (2) discriminability between models of varying quality, and (3) efficiency in compute. Through this lens, we identify critical failure modes that violate faithfulness and discriminability, misrepresenting model capabilities: (i) multiple-choice formats reward guessing, poorly reflect downstream use cases, and saturate early as models improve; (ii) blindly solvable questions, which can be answered without images, constitute up to 70% of some evaluations; and (iii) mislabeled or ambiguous samples compromise up to 42% of examples in certain datasets. Regarding efficiency, the computational burden of evaluating frontier models has become prohibitive: by some accounts, nearly 20% of development compute is devoted to evaluation alone. Rather than discarding existing benchmarks, we curate them via transformation and filtering to maximize fidelity and discriminability. We find that converting multiple-choice questions to generative tasks reveals sharp capability drops of up to 35%. In addition, filtering blindly solvable and mislabeled samples improves discriminative power while simultaneously reducing computational cost. We release DatBench-Full, a cleaned evaluation suite of 33 datasets spanning nine VLM capabilities, and DatBench, a discriminative subset that achieves 13x average speedup (up to 50x) while closely matching the discriminative power of the original datasets. Our work outlines a path toward evaluation practices that are both rigorous and sustainable as VLMs continue to scale.

We introduce EvaLearn, a pioneering benchmark designed to evaluate large language models (LLMs) on their learning capability and efficiency in challenging tasks, a critical, yet underexplored aspect of model potential. EvaLearn contains 648 challenging problems across six task types, grouped into 182 sequences, each sequence dedicated to one task type. Diverging from most existing benchmarks that evaluate models in parallel, EvaLearn requires models to solve problems sequentially, allowing them to leverage the experience gained from previous solutions. EvaLearn provides five comprehensive automated metrics to evaluate models and quantify their learning capability and efficiency. We extensively benchmark nine frontier models and observe varied performance profiles: some models, such as Claude-3.7-sonnet, start with moderate initial performance but exhibit strong learning ability, while some models struggle to benefit from experience and may even show negative transfer. Moreover, we investigate model performance under two learning settings and find that instance-level rubrics and teacher-model feedback further facilitate model learning. Importantly, we observe that current LLMs with stronger static abilities do not show a clear advantage in learning capability across all tasks, highlighting that EvaLearn evaluates a new dimension of model performance. We hope EvaLearn provides a novel evaluation perspective for assessing LLM potential and understanding the gap between models and human capabilities, promoting the development of deeper and more dynamic evaluation approaches. All datasets, the automatic evaluation framework, and the results studied in this paper are available at the GitHub repository.

The inherent rigidity of fixed-size benchmarks makes them an inefficient tool for model evaluation. Diverse evaluation objectives, including model ranking, model selection and testing throughout development, demand varying levels of statistical power. The mismatch between fixed sample sizes and these diverse needs results in either excessive computational cost or compromised reliability - a critical concern for model evaluation. To overcome these limitations, we call for adoption of sequential testing in our field. We provide an adaptive evaluation framework, that provides a principled way to navigate the trade-off between efficiency and reliability in model evaluation. Our framework combines the established statistical paradigm of sequential testing with stopping criteria tailored to common evaluation needs such as diminishing returns detection, and minimum detectable effect size. We demonstrate its ability to adaptively manage the efficiency-reliability trade-off on the Open VLM Leaderboard, including, for example, a 80% reduction in computational cost compared to fixed-size evaluation (with a 2.5-point CI width allowance) while maintaining statistical significance.

Model-based evaluation is at the heart of successful model development -- as a reward model for training, and as a replacement for human evaluation. To train such evaluators, the standard approach is to collect a large amount of human preference judgments over model responses, which is costly and the data becomes stale as models improve. In this work, we present an approach that aims to im-prove evaluators without human annotations, using synthetic training data only. Starting from unlabeled instructions, our iterative self-improvement scheme generates contrasting model outputs and trains an LLM-as-a-Judge to produce reasoning traces and final judgments, repeating this training at each new iteration using the improved predictions. Without any labeled preference data, our Self-Taught Evaluator can improve a strong LLM (Llama3-70B-Instruct) from 75.4 to 88.3 (88.7 with majority vote) on RewardBench. This outperforms commonly used LLM judges such as GPT-4 and matches the performance of the top-performing reward models trained with labeled examples.

Manual curation of evaluation datasets is struggling to keep up with the rapidly expanding capabilities and deployment scenarios of language models. Towards scalable model profiling, we introduce and validate a framework for evaluating LLMs, called Adaptive Evaluations. Adaptive evaluations use scaffolded language models (evaluator agents) to search through a target model's behavior on a domain dataset and create difficult questions (tasks) that can discover and probe the model's failure modes. We find that frontier models lack consistency when adaptively probed with our framework on a diverse suite of datasets and tasks, including but not limited to legal reasoning, forecasting, and online harassment. Generated questions pass human validity checks and often transfer to other models with different capability profiles, demonstrating that adaptive evaluations can also be used to create difficult domain-specific datasets.

We study whether language models can evaluate the validity of their own claims and predict which questions they will be able to answer correctly. We first show that larger models are well-calibrated on diverse multiple choice and true/false questions when they are provided in the right format. Thus we can approach self-evaluation on open-ended sampling tasks by asking models to first propose answers, and then to evaluate the probability "P(True)" that their answers are correct. We find encouraging performance, calibration, and scaling for P(True) on a diverse array of tasks. Performance at self-evaluation further improves when we allow models to consider many of their own samples before predicting the validity of one specific possibility. Next, we investigate whether models can be trained to predict "P(IK)", the probability that "I know" the answer to a question, without reference to any particular proposed answer. Models perform well at predicting P(IK) and partially generalize across tasks, though they struggle with calibration of P(IK) on new tasks. The predicted P(IK) probabilities also increase appropriately in the presence of relevant source materials in the context, and in the presence of hints towards the solution of mathematical word problems. We hope these observations lay the groundwork for training more honest models, and for investigating how honesty generalizes to cases where models are trained on objectives other than the imitation of human writing.

Model performance evaluation is a critical and expensive task in machine learning and computer vision. Without clear guidelines, practitioners often estimate model accuracy using a one-time completely random selection of the data. However, by employing tailored sampling and estimation strategies, one can obtain more precise estimates and reduce annotation costs. In this paper, we propose a statistical framework for model evaluation that includes stratification, sampling, and estimation components. We examine the statistical properties of each component and evaluate their efficiency (precision). One key result of our work is that stratification via k-means clustering based on accurate predictions of model performance yields efficient estimators. Our experiments on computer vision datasets show that this method consistently provides more precise accuracy estimates than the traditional simple random sampling, even with substantial efficiency gains of 10x. We also find that model-assisted estimators, which leverage predictions of model accuracy on the unlabeled portion of the dataset, are generally more efficient than the traditional estimates based solely on the labeled data.

Existing evaluation methods largely rely on clean, static benchmarks, which can overestimate true model performance by failing to capture the noise and variability inherent in real-world user inputs. This is especially true for language models, which can face human-generated text queries containing mistakes, typos, or alternative ways of phrasing the same question. In this work, we introduce a theoretical framework for quantifying model sensitivity to prompt variants, or brittleness, that can enable us to disentangle data-induced difficulty from prompt-related variability. Using this framework, we design a novel evaluation pipeline, Brittlebench, to holistically evaluate the sensitivity of frontier models. We apply semantics-preserving perturbations to a suite of popular benchmarks, and observe model performance to degrade as much as 12%. However, these perturbations do not affect all models equally: even a single perturbation alters the relative ranking of models in 63% of cases, impacting conclusions about comparative model performance. Decomposing the total variance of both state-of-the-art open-weight and commercial models, we find that semantics-preserving input perturbations can account for up to half of the performance variance for a given model. Brittlebench highlights the need for more robust evaluations and models, and allows us to systematically understand model brittleness.

As large language models (LLMs) become increasingly versatile, numerous large scale benchmarks have been developed to thoroughly assess their capabilities. These benchmarks typically consist of diverse datasets and prompts to evaluate different aspects of LLM performance. However, comprehensive evaluations on hundreds or thousands of prompts incur tremendous costs in terms of computation, money, and time. In this work, we investigate strategies to improve evaluation efficiency by selecting a subset of examples from each benchmark using a learned policy. Our approach models the dependencies across test examples, allowing accurate prediction of the evaluation outcomes for the remaining examples based on the outcomes of the selected ones. Consequently, we only need to acquire the actual evaluation outcomes for the selected subset. We rigorously explore various subset selection policies and introduce a novel RL-based policy that leverages the captured dependencies. Empirical results demonstrate that our approach significantly reduces the number of evaluation prompts required while maintaining accurate performance estimates compared to previous methods.

The proliferation of open-source Large Language Models (LLMs) underscores the pressing need for evaluation methods. Existing works primarily rely on external evaluators, focusing on training and prompting strategies. However, a crucial aspect - model-aware glass-box features - is overlooked. In this study, we explore the utility of glass-box features under the scenario of self-evaluation, namely applying an LLM to evaluate its own output. We investigate various glass-box feature groups and discovered that the softmax distribution serves as a reliable indicator for quality evaluation. Furthermore, we propose two strategies to enhance the evaluation by incorporating features derived from references. Experimental results on public benchmarks validate the feasibility of self-evaluation of LLMs using glass-box features.

The prohibitive cost of evaluating large language models (LLMs) on comprehensive benchmarks necessitates the creation of small yet representative data subsets (i.e., tiny benchmarks) that enable efficient assessment while retaining predictive fidelity. Current methods for this task operate under a model-centric paradigm, selecting benchmarking items based on the collective performance of existing models. Such approaches are limited by large upfront costs, an inability to immediately handle new benchmarks (`cold-start'), and the fragile assumption that future models will share the failure patterns of their predecessors. In this work, we challenge this paradigm and propose a item-centric approach to benchmark subset selection, arguing that selection should be based on the intrinsic properties of the task items themselves, rather than on model-specific failure patterns. We instantiate this item-centric efficient benchmarking approach via a novel method, Scales++, where data selection is based on the cognitive demands of the benchmark samples. Empirically, we show Scales++ reduces the upfront selection cost by over 18x while achieving competitive predictive fidelity. On the Open LLM Leaderboard, using just a 0.5\% data subset, we predict full benchmark scores with a 2.9% mean absolute error. We demonstrate that this item-centric approach enables more efficient model evaluation without significant fidelity degradation, while also providing better cold-start performance and more interpretable benchmarking.

This paper poses two critical issues in evaluating base models (without post-training): (1) Unstable evaluation during training: in the early stages of pre-training, the models lack the capability to answer questions as required, leading to unstable evaluation results. This instability makes it difficult to provide solid conclusions to guide the training, especially for key experiments such as data ablation and scaling law. (2) Inconsistency between base and instruct models: base models generally exhibit poorer evaluation performance compared to corresponding instruct models. This gap poses a challenge for assessing whether a base model with better evaluation can truly lead to a better instruct model. To address these issues, we propose Base model Oriented Systematic Evaluation (BOSE), a method specifically designed to optimize the evaluation of base models. Specifically, BOSE introduces two key innovations: In-Context Light-instruction Prompt (ICLiP) for open-ended tasks and Blank-ppl for multi-choice tasks with candidate options, which transforms the standard perplexity (ppl) metric into a fill-in-the-blank format to mitigate early-stage evaluation fluctuations. Furthermore, we are the first to propose Kendall's rank correlation to quantitatively measure the evaluation stability and consistency. Experimental results demonstrate that BOSE significantly enhances both the stability of evaluations during pre-training and the consistency between base and instruct models, thereby providing more reliable guidance for the LLMs' training.

The evaluation of Large Language Models (LLMs) increasingly relies on other LLMs acting as judges. However, current evaluation paradigms typically yield a single score or ranking, answering which model is better but not why. While essential for benchmarking, these top-level scores obscure the specific, actionable reasons behind a model's performance. To bridge this gap, we introduce CLEAR, an interactive, open-source package for LLM-based error analysis. CLEAR first generates per-instance textual feedback, then it creates a set of system-level error issues, and quantifies the prevalence of each identified issue. Our package also provides users with an interactive dashboard that allows for a comprehensive error analysis through aggregate visualizations, applies interactive filters to isolate specific issues or score ranges, and drills down to the individual instances that exemplify a particular behavioral pattern. We demonstrate CLEAR analysis for RAG and Math benchmarks, and showcase its utility through a user case study.

Comprehensive evaluation of Large Language Models (LLMs) is an open research problem. Existing evaluations rely on deterministic point estimates generated via greedy decoding. However, we find that deterministic evaluations fail to capture the whole output distribution of a model, yielding inaccurate estimations of model capabilities. This is particularly problematic in critical contexts such as unlearning and alignment, where precise model evaluations are crucial. To remedy this, we introduce the first formal probabilistic evaluation framework for LLMs. Namely, we propose novel metrics with high probability guarantees concerning the output distribution of a model. Our metrics are application-independent and allow practitioners to make more reliable estimates about model capabilities before deployment. Our experimental analysis reveals that deterministic evaluations falsely indicate successful unlearning and alignment, whereas our probabilistic evaluations better capture model capabilities. We show how to overcome challenges associated with probabilistic outputs in a case study on unlearning by introducing (1) a novel loss based on entropy optimization, and (2) adaptive temperature scaling. We demonstrate that our approach significantly enhances unlearning in probabilistic settings on recent benchmarks. Overall, our proposed shift from point estimates to probabilistic evaluations of output distributions represents an important step toward comprehensive evaluations of LLMs. Code available at https://www.cs.cit.tum.de/daml/probabilistic-unlearning/.

Progress in AI is often demonstrated by new models claiming improved performance on tasks measuring model capabilities. Evaluating language models in particular is challenging, as small changes to how a model is evaluated on a task can lead to large changes in measured performance. There is no common standard setup, so different models are evaluated on the same tasks in different ways, leading to claims about which models perform best not being reproducible. We propose OLMES, a completely documented, practical, open standard for reproducible LLM evaluations. In developing this standard, we identify and review the varying factors in evaluation practices adopted by the community - such as details of prompt formatting, choice of in-context examples, probability normalizations, and task formulation. In particular, OLMES supports meaningful comparisons between smaller base models that require the unnatural "cloze" formulation of multiple-choice questions against larger models that can utilize the original formulation. OLMES includes well-considered recommendations guided by results from existing literature as well as new experiments investigating open questions.

Standard benchmarks have become increasingly unreliable due to saturation, subjectivity, and poor generalization. We argue that evaluating model's ability to acquire information actively is important to assess model's intelligence. We propose Interactive Benchmarks, a unified evaluation paradigm that assesses model's reasoning ability in an interactive process under budget constraints. We instantiate this framework across two settings: Interactive Proofs, where models interact with a judge to deduce objective truths or answers in logic and mathematics; and Interactive Games, where models reason strategically to maximize long-horizon utilities. Our results show that interactive benchmarks provide a robust and faithful assessment of model intelligence, revealing that there is still substantial room to improve in interactive scenarios. Project page: https://github.com/interactivebench/interactivebench

The large-scale training of multi-modal models on data scraped from the web has shown outstanding utility in infusing these models with the required world knowledge to perform effectively on multiple downstream tasks. However, one downside of scraping data from the web can be the potential sacrifice of the benchmarks on which the abilities of these models are often evaluated. To safeguard against test data contamination and to truly test the abilities of these foundation models we propose LiveXiv: A scalable evolving live benchmark based on scientific ArXiv papers. LiveXiv accesses domain-specific manuscripts at any given timestamp and proposes to automatically generate visual question-answer pairs (VQA). This is done without any human-in-the-loop, using the multi-modal content in the manuscripts, like graphs, charts, and tables. Moreover, we introduce an efficient evaluation approach that estimates the performance of all models on the evolving benchmark using evaluations of only a subset of models. This significantly reduces the overall evaluation cost. We benchmark multiple open and proprietary Large Multi-modal Models (LMMs) on the first version of our benchmark, showing its challenging nature and exposing the models true abilities, avoiding contamination. Lastly, in our commitment to high quality, we have collected and evaluated a manually verified subset. By comparing its overall results to our automatic annotations, we have found that the performance variance is indeed minimal (<2.5%). Our dataset is available online on HuggingFace, and our code will be available here.

Evaluations are critical for understanding the capabilities of large language models (LLMs). Fundamentally, evaluations are experiments; but the literature on evaluations has largely ignored the literature from other sciences on experiment analysis and planning. This article shows researchers with some training in statistics how to think about and analyze data from language model evaluations. Conceptualizing evaluation questions as having been drawn from an unseen super-population, we present formulas for analyzing evaluation data, measuring differences between two models, and planning an evaluation experiment. We make a number of specific recommendations for running language model evaluations and reporting experiment results in a way that minimizes statistical noise and maximizes informativeness.

Large language model (LLM) agents allocate test-time compute adaptively as they revise solutions, use tools, explore alternatives, and decide when to stop. This test-time strategy makes it difficult to measure how agent performance scales. We study open-ended tasks that provide continuous scores for intermediate submissions, making progress observable throughout long trajectories. We propose Elo-per-token analysis, which tracks the best solution found at each token budget and uses a Bradley-Terry model to aggregate within-task orderings into Elo ratings across tasks with different score scales. We apply it to four general-purpose agents on four open-ended benchmarks, with sessions of up to 100M tokens, and to three feedback-driven LLM optimization harnesses in controlled single-task interventions. Independent sampling provides a theoretically characterized reference, for which Elo grows linearly with log compute. Against this reference, agents can initially convert tokens into Elo faster than independent sampling, but their marginal gains diminish and eventually fall below the reference. In contrast, the strongest historical human contestants improve superlinearly over contest time on shared AtCoder Heuristic Contest tasks, providing evidence of continual learning and substantial headroom after agents slow down. We define the scaling inflection point as the per-session budget where marginal Elo gains match the independent-sampling reference. Using this point as the per-session budget, we split 100M tokens across parallel sessions on FrontierCS Polyomino Packing, gaining +264 Elo over one long session and +355 over ten short sessions.

Large Language Models (LLMs) are frequently used for multi-faceted language generation and evaluation tasks that involve satisfying intricate user constraints or taking into account multiple aspects and criteria. However, their performance can fall short, due to the model's lack of coherence and inability to plan and decompose the problem. We propose Branch-Solve-Merge (BSM), a Large Language Model program (Schlag et al., 2023) for tackling such challenging natural language tasks. It consists of branch, solve, and merge modules that are parameterized with specific prompts to the base LLM. These three modules plan a decomposition of the task into multiple parallel sub-tasks, independently solve them, and fuse the solutions to the sub-tasks. We apply our method to the tasks of LLM response evaluation and constrained text generation and evaluate its effectiveness with multiple LLMs, including Vicuna, LLaMA-2-chat, and GPT-4. BSM improves the evaluation correctness and consistency for each LLM by enhancing human-LLM agreement by up to 26%, reducing length and pairwise position biases by up to 50%, and allowing LLaMA-2-chat to match or outperform GPT-4 on most domains. On the constraint story generation task, BSM improves the coherence of the stories while also improving constraint satisfaction by 12%.

Political scientists are rapidly adopting large language models (LLMs) for text annotation, yet the sensitivity of annotation results to implementation choices remains poorly understood. Most evaluations test a single model or configuration; how model choice, model size, learning approach, and prompt style interact, and whether popular "best practices" survive controlled comparison, are largely unexplored. We present a controlled evaluation of these pipeline choices, testing six open-weight models across four political science annotation tasks under identical quantisation, hardware, and prompt-template conditions. Our central finding is methodological: interaction effects dominate main effects, so seemingly reasonable pipeline choices can become consequential researcher degrees of freedom. No single model, prompt style, or learning approach is uniformly superior, and the best-performing model varies across tasks. Two corollaries follow. First, model size is an unreliable guide both to cost and to performance: cross-family efficiency differences are so large that some larger models are less resource-intensive than much smaller alternatives, while within model families mid-range variants often match or exceed larger counterparts. Second, widely recommended prompt engineering techniques yield inconsistent and sometimes negative effects on annotation performance. We use these benchmark results to develop a validation-first framework - with a principled ordering of pipeline decisions, guidance on prompt freezing and held-out evaluation, reporting standards, and open-source tools - to help researchers navigate this decision space transparently.

The rapid development of large language model (LLM) evaluation methodologies and datasets has led to a profound challenge: integrating state-of-the-art evaluation techniques cost-effectively while ensuring reliability, reproducibility, and efficiency. Currently, there is a notable absence of a unified and adaptable framework that seamlessly integrates various evaluation approaches. Moreover, the reliability of evaluation findings is often questionable due to potential data contamination, with the evaluation efficiency commonly overlooked when facing the substantial costs associated with LLM inference. In response to these challenges, we introduce FreeEval, a modular and scalable framework crafted to enable trustworthy and efficient automatic evaluations of LLMs. Firstly, FreeEval's unified abstractions simplify the integration and improve the transparency of diverse evaluation methodologies, encompassing dynamic evaluation that demand sophisticated LLM interactions. Secondly, the framework integrates meta-evaluation techniques like human evaluation and data contamination detection, which, along with dynamic evaluation modules in the platform, enhance the fairness of the evaluation outcomes. Lastly, FreeEval is designed with a high-performance infrastructure, including distributed computation and caching strategies, enabling extensive evaluations across multi-node, multi-GPU clusters for open-source and proprietary LLMs.

In the domain of model-based engineering, models are essential components that enable system design and analysis. Traditionally, the creation of these models has been a manual process requiring not only deep modeling expertise but also substantial domain knowledge of target systems. With the rapid advancement of generative artificial intelligence, large language models (LLMs) show potential for automating model generation. This work explores the generation of instance models using LLMs, focusing specifically on producing XMI-based instance models from Ecore metamodels and natural language specifications. We observe that current LLMs struggle to directly generate valid XMI models. To address this, we propose a two-step approach: first, using LLMs to produce a simplified structured output containing all necessary instance model information, namely a conceptual instance model, and then compiling this intermediate representation into a valid XMI file. The conceptual instance model is format-independent, allowing it to be transformed into various modeling formats via different compilers. The feasibility of the proposed method has been demonstrated using several LLMs, including GPT-4o, o1-preview, Llama 3.1 (8B and 70B). Results show that the proposed method significantly improves the usability of LLMs for instance model generation tasks. Notably, the smaller open-source model, Llama 3.1 70B, demonstrated performance comparable to proprietary GPT models within the proposed framework.

The conventional evaluation protocols on machine learning models rely heavily on a labeled, i.i.d-assumed testing dataset, which is not often present in real world applications. The Automated Model Evaluation (AutoEval) shows an alternative to this traditional workflow, by forming a proximal prediction pipeline of the testing performance without the presence of ground-truth labels. Despite its recent successes, the AutoEval frameworks still suffer from an overconfidence issue, substantial storage and computational cost. In that regard, we propose a novel measure -- Meta-Distribution Energy (MDE) -- that allows the AutoEval framework to be both more efficient and effective. The core of the MDE is to establish a meta-distribution statistic, on the information (energy) associated with individual samples, then offer a smoother representation enabled by energy-based learning. We further provide our theoretical insights by connecting the MDE with the classification loss. We provide extensive experiments across modalities, datasets and different architectural backbones to validate MDE's validity, together with its superiority compared with prior approaches. We also prove MDE's versatility by showing its seamless integration with large-scale models, and easy adaption to learning scenarios with noisy- or imbalanced- labels. Code and data are available: https://github.com/pengr/Energy_AutoEval

Project Greenhouse represents our exploration of a simple thesis: We believe that it is possible to build fully open and sovereign models for agentic search with only modest computational resources. As a first milestone, we describe how to build a competitive pointwise decoder-only reranker using a simple two-step recipe comprising pre-training from scratch followed by supervised fine-tuning, starting only from commonly available datasets. Contrary to the dominant approach in the literature, we do not rely on existing open-weight backbones from third parties, and thus we are fully in control of model training, from end to end. We were able to accomplish the bulk of our experiments using no more than a handful of GPUs. This report articulates the importance and benefits of our approach, and we share artifacts that enable transparent, independent reproduction of all aspects of model training. Beyond data, code, and configurations that capture our efforts, we also release checkpoints for our family of Gaggle models, demonstrating the feasibility of our approach and providing a first step toward validating our broader thesis.

Large language models (LLMs) garner significant attention for their unprecedented performance, leading to an increasing number of researches evaluating LLMs. However, these evaluation benchmarks are limited to assessing the instruction-following capabilities, overlooking the fundamental abilities that emerge during the pre-training stage. Previous subjective evaluation methods mainly reply on scoring by API models. However, in the absence of references, large models have shown limited ability to discern subtle differences. To bridge the gap, we propose F-Eval, a bilingual evaluation benchmark to evaluate the fundamental abilities, including expression, commonsense and logic. The tasks in F-Eval include multi-choice objective tasks, open-ended objective tasks, reference-based subjective tasks and reference-free subjective tasks. For reference-free subjective tasks, we devise new evaluation methods, serving as alternatives to scoring by API models. We conduct evaluations on 13 advanced LLMs. Results show that our evaluation methods show higher correlation coefficients and larger distinction than other evaluators. Additionally, we discuss the influence of different model sizes, dimensions, and normalization methods. We anticipate that F-Eval will facilitate the study of LLMs' fundamental abilities.

Large language models (LLMs) frequently achieve impressive scores on standardized benchmarks, yet accuracy alone offers a limited view of their capabilities. Evaluating open-source LLMs through leaderboards faces persistent issues like data contamination, narrow task scope, and weak alignment with real-world reliability. Benchmark-based evaluations such as MMLU PRO, BBH, or IFEval primarily capture what a model outputs on fixed test sets, not how it processes information, calibrates uncertainty, or structures internal knowledge. In this article, we advocate for a shift from benchmark-centric evaluation toward a complementary, state-centered intrinsic assessment of LLMs. To this end, we introduce Latent Performance Profiling (LPP) -- a framework that derives task-agnostic diagnostics from hidden activations and output distributions. LPP defines a set of scalar metrics on a model's latent representations and dynamics, revealing scale-independent traits that enable interpretable comparisons and uncover hidden vulnerabilities. Unlike static accuracy scores, LPP provides stable, architecture-sensitive signatures across models of similar size. With extensive empirical analyses across eight LLMs, spanning a size range of 0.5B-14B, we demonstrate that models with similar benchmark scores can exhibit contrasting latent profiles, such as differences in entropy or adaptability. Guided by these insights, we design synthetic probes for uncertainty and symbolic reasoning that align with intrinsic metrics while decoupling from leaderboard bias. We recommend that reporting LPP alongside benchmarks provides a deeper, interpretable understanding of model behavior, enabling more reliable model selection, safety assessment, and evaluation beyond surface-level accuracy.

Evaluating large language models at scale remains a practical bottleneck for many organizations. While existing evaluation frameworks work well for thousands of examples, they struggle when datasets grow to hundreds of thousands or millions of samples. This scale is common when assessing model behavior across diverse domains or conducting comprehensive regression testing. We present Spark-LLM-Eval, a distributed evaluation framework built natively on Apache Spark. The system treats evaluation as a data-parallel problem, partitioningexamplesacrossexecutorsandaggregatingresultswithproperstatistical accounting. Beyond raw throughput, we emphasize statistical rigor: every reported metric includes bootstrap confidence intervals, and model comparisons come with appropriate significance tests (paired t-tests, McNemar's test, or Wilcoxon signed-rank, depending on the metric type). The framework also addresses the cost problem inherent in LLM evaluation through content-addressable response caching backed by Delta Lake, which allows iterating on metric definitions without re-running inference. We describe the system architecture, the statistical methodology, and report benchmark results showing linear scaling with cluster size. The framework and all evaluation code are available as open source.

Model cards describe model behavior through a mixture of textual descriptions and structured artifacts, including performance, configuration, and dataset tables. Existing model search systems rely predominantly on semantic similarity over text, which can produce homogeneous result sets and limit exploration of alternatives. We argue that model search is inherently comparative: users want models that are task-aligned yet differentiated in measurable ways. We hypothesize that this balance requires retrieval over condensed, high-quality evidence rather than verbose descriptions, and much of that evidence is concentrated in structured tables. We present StructuredSemanticSearch, a table-driven model search framework built on the ModelTables benchmark. Given a query, StructuredSemanticSearch combines a semantic baseline for task alignment with a structure-aware pipeline that discovers query-related model-card tables using table discovery operators such as unionability, joinability, and keyword search. Retrieved tables are mapped back to model cards under a controlled top-k budget, enabling fair comparison between text-based and table-based retrieval. Beyond retrieval, StructuredSemanticSearch adapts table integration to the model-table domain through orientation-aware integration, producing compact integrated views of tables from partially overlapping and sometimes transposed evidence tables. For evaluation, we introduce a nugget-based, auditable protocol that extracts compact evidence items from model cards, matches queries to condition- or intent-specific nuggets, and measures evidence coverage and diversity over retrieved model-card candidate sets. This protocol also provides a scalable path toward approximate, evidence-based labeling in dynamic model lakes. Experiments on 597 model-recommendation queries show improved nugget coverage for the structure-aware pipeline than semantic baseline

Evaluation is critical for assessing capabilities, tracking scientific progress, and informing model selection. In this paper, we present three desiderata for a good benchmark for language models: (i) salience (e.g., knowledge about World War II is more salient than a random day in history), (ii) novelty (i.e., the benchmark reveals new trends in model rankings not shown by previous benchmarks), and (iii) difficulty (i.e., the benchmark should be difficult for existing models, leaving headroom for future improvement). We operationalize these three desiderata and cast benchmark creation as a search problem, that of finding benchmarks that that satisfy all three desiderata. To tackle this search problem, we present AutoBencher, which uses a language model to automatically search for datasets that meet the three desiderata. AutoBencher uses privileged information (e.g. relevant documents) to construct reliable datasets, and adaptivity with reranking to optimize for the search objective. We use AutoBencher to create datasets for math, multilingual, and knowledge-intensive question answering. The scalability of AutoBencher allows it to test fine-grained categories and tail knowledge, creating datasets that are on average 27% more novel and 22% more difficult than existing benchmarks. A closer investigation of our constructed datasets shows that we can identify specific gaps in LM knowledge in language models that are not captured by existing benchmarks, such as Gemini Pro performing much worse on question answering about the Permian Extinction and Fordism, while OpenAGI-7B performing surprisingly well on QA about COVID-19.

Are frontier models able to introspect about their internal states? Recent work suggests that under certain conditions a complex enough model can audit its own internals, call out what changed, and report back confidently about it. We tested that claim on eight open-weight models from seven families and found no such ability: asked whether their own computation had been altered, none answered better than chance. To test it we built Open-Weight Masked Introspection (OWMI), a framework that intervenes on residual-stream sites, attention heads and sparse-autoencoder features, then interrogates the model about the change against the null conditions an answer has to beat: sham runs where nothing was altered, impact-matched random perturbations, and a text-only observer that sees only the visible output. Over 78,000 measurements, no model's report discriminates a real intervention from a sham beyond chance (AUROC ~0.5007), and an equivalence test bounds the effect below 0.15 percentage points of AUROC. Surprisingly, all the information needed is in the models. A model fine-tuned to report this class of intervention reaches near-perfect recovery on held-out directions, and a linear probe recovers intervention presence from the same activations at 75% to 95.8% accuracy, sharpening to no held-out error at the last layer before the model speaks. In one model the signal surfaces in the confidence rather than the words: its yes-or-no report never varies, while the confidence attached to it separates intervention from sham at AUROC 0.647. The failure sits in the path from internal state to verbal report, so oversight that reads a model's own testimony needs validating against an internal reference. While our results show the inability of current open-weight models to introspect, the debate is not settled for future models.

The advances of large foundation models necessitate wide-coverage, low-cost, and zero-contamination benchmarks. Despite continuous exploration of language model evaluations, comprehensive studies on the evaluation of Large Multi-modal Models (LMMs) remain limited. In this work, we introduce LMMS-EVAL, a unified and standardized multimodal benchmark framework with over 50 tasks and more than 10 models to promote transparent and reproducible evaluations. Although LMMS-EVAL offers comprehensive coverage, we find it still falls short in achieving low cost and zero contamination. To approach this evaluation trilemma, we further introduce LMMS-EVAL LITE, a pruned evaluation toolkit that emphasizes both coverage and efficiency. Additionally, we present Multimodal LIVEBENCH that utilizes continuously updating news and online forums to assess models' generalization abilities in the wild, featuring a low-cost and zero-contamination evaluation approach. In summary, our work highlights the importance of considering the evaluation trilemma and provides practical solutions to navigate the trade-offs in evaluating large multi-modal models, paving the way for more effective and reliable benchmarking of LMMs. We opensource our codebase and maintain leaderboard of LIVEBENCH at https://github.com/EvolvingLMMs-Lab/lmms-eval and https://huggingface.co/spaces/lmms-lab/LiveBench.

Traditional fixed test sets fall short in evaluating open-ended capabilities of foundation models. To address this, we propose ONEBench(OpeN-Ended Benchmarking), a new testing paradigm that consolidates individual evaluation datasets into a unified, ever-expanding sample pool. ONEBench allows users to generate custom, open-ended evaluation benchmarks from this pool, corresponding to specific capabilities of interest. By aggregating samples across test sets, ONEBench enables the assessment of diverse capabilities beyond those covered by the original test sets, while mitigating overfitting and dataset bias. Most importantly, it frames model evaluation as a collective process of selecting and aggregating sample-level tests. The shift from task-specific benchmarks to ONEBench introduces two challenges: (1)heterogeneity and (2)incompleteness. Heterogeneity refers to the aggregation over diverse metrics, while incompleteness describes comparing models evaluated on different data subsets. To address these challenges, we explore algorithms to aggregate sparse measurements into reliable model scores. Our aggregation algorithm ensures identifiability(asymptotically recovering ground-truth scores) and rapid convergence, enabling accurate model ranking with less data. On homogenous datasets, we show our aggregation algorithm provides rankings that highly correlate with those produced by average scores. We also demonstrate robustness to ~95% of measurements missing, reducing evaluation cost by up to 20x with little-to-no change in model rankings. We introduce ONEBench-LLM for language models and ONEBench-LMM for vision-language models, unifying evaluations across these domains. Overall, we present a technique for open-ended evaluation, which can aggregate over incomplete, heterogeneous sample-level measurements to continually grow a benchmark alongside the rapidly developing foundation models.

Large language models (LLMs) have demonstrated remarkable proficiency in generating detailed and coherent explanations of complex concepts. However, the extent to which these models truly comprehend the concepts they articulate remains unclear. To assess the level of comprehension of a model relative to the content it generates, we implemented a self-evaluation pipeline where models: (i) given a topic generate an excerpt with information about the topic, (ii) given an excerpt generate question-answer pairs, and finally (iii) given a question generate an answer. We refer to this self-evaluation approach as Explain-Query-Test (EQT). Interestingly, the accuracy on generated questions resulting from running the EQT pipeline correlates strongly with the model performance as verified by typical benchmarks such as MMLU-Pro. In other words, EQT's performance is predictive of MMLU-Pro's, and EQT can be used to rank models without the need for any external source of evaluation data other than lists of topics of interest. Moreover, our results reveal a disparity between the models' ability to produce detailed explanations and their performance on questions related to those explanations. This gap highlights fundamental limitations in the internal knowledge representation and reasoning abilities of current LLMs. We release the code at https://github.com/asgsaeid/EQT.

The Automated Model Evaluation (AutoEval) framework entertains the possibility of evaluating a trained machine learning model without resorting to a labeled testing set. Despite the promise and some decent results, the existing AutoEval methods heavily rely on computing distribution shifts between the unlabelled testing set and the training set. We believe this reliance on the training set becomes another obstacle in shipping this technology to real-world ML development. In this work, we propose Contrastive Automatic Model Evaluation (CAME), a novel AutoEval framework that is rid of involving training set in the loop. The core idea of CAME bases on a theoretical analysis which bonds the model performance with a contrastive loss. Further, with extensive empirical validation, we manage to set up a predictable relationship between the two, simply by deducing on the unlabeled/unseen testing set. The resulting framework CAME establishes a new SOTA results for AutoEval by surpassing prior work significantly.

Forming a reliable judgement of a machine learning (ML) model's appropriateness for an application ecosystem is critical for its responsible use, and requires considering a broad range of factors including harms, benefits, and responsibilities. In practice, however, evaluations of ML models frequently focus on only a narrow range of decontextualized predictive behaviours. We examine the evaluation gaps between the idealized breadth of evaluation concerns and the observed narrow focus of actual evaluations. Through an empirical study of papers from recent high-profile conferences in the Computer Vision and Natural Language Processing communities, we demonstrate a general focus on a handful of evaluation methods. By considering the metrics and test data distributions used in these methods, we draw attention to which properties of models are centered in the field, revealing the properties that are frequently neglected or sidelined during evaluation. By studying these properties, we demonstrate the machine learning discipline's implicit assumption of a range of commitments which have normative impacts; these include commitments to consequentialism, abstractability from context, the quantifiability of impacts, the limited role of model inputs in evaluation, and the equivalence of different failure modes. Shedding light on these assumptions enables us to question their appropriateness for ML system contexts, pointing the way towards more contextualized evaluation methodologies for robustly examining the trustworthiness of ML models

Representation engineering methods have recently shown promise for enabling efficient steering of model behavior. However, evaluation pipelines for these methods have primarily relied on subjective demonstrations, instead of quantitative, objective metrics. We aim to take a step towards addressing this issue by advocating for four properties missing from current evaluations: (i) contexts sufficiently similar to downstream tasks should be used for assessing intervention quality; (ii) model likelihoods should be accounted for; (iii) evaluations should allow for standardized comparisons across different target behaviors; and (iv) baseline comparisons should be offered. We introduce an evaluation pipeline grounded in these criteria, offering both a quantitative and visual analysis of how effectively a given method works. We use this pipeline to evaluate two representation engineering methods on how effectively they can steer behaviors such as truthfulness and corrigibility, finding that some interventions are less effective than previously reported.

Recent work found that LLMs are sensitive to a wide range of arbitrary prompt dimensions, including the type of delimiters, answer enumerators, instruction wording, and more. This throws into question popular single-prompt evaluation practices. We present DOVE (Dataset Of Variation Evaluation) a large-scale dataset containing prompt perturbations of various evaluation benchmarks. In contrast to previous work, we examine LLM sensitivity from an holistic perspective, and assess the joint effects of perturbations along various dimensions, resulting in thousands of perturbations per instance. We evaluate several model families against DOVE, leading to several findings, including efficient methods for choosing well-performing prompts, observing that few-shot examples reduce sensitivity, and identifying instances which are inherently hard across all perturbations. DOVE consists of more than 250M prompt perturbations and model outputs, which we make publicly available to spur a community-wide effort toward meaningful, robust, and efficient evaluation. Browse the data, contribute, and more: https://slab-nlp.github.io/DOVE/

As language models scale up, it becomes increasingly expensive to verify research ideas because conclusions on small models do not trivially transfer to large ones. A possible solution is to establish a generic system that directly predicts some metrics for large models solely based on the results and hyperparameters from small models. Existing methods based on scaling laws require hyperparameter search on the largest models, which is impractical with limited resources. We address this issue by presenting our discoveries indicating that Maximal Update parametrization (Mup) enables accurate fitting of scaling laws for hyperparameters close to common loss basins, without any search. Thus, different models can be directly compared on large scales with loss prediction even before the training starts. We propose a new paradigm as a first step towards reliable academic research for any model scale without heavy computation. Code is publicly available at https://github.com/cofe-ai/Mu-scaling.

We study a fundamental problem in the evaluation of large language models that we call training on the test task. Unlike wrongful practices like training on the test data, leakage, or data contamination, training on the test task is not a malpractice. Rather, the term describes a growing set of practices that utilize knowledge about evaluation tasks at training time. We demonstrate that training on the test task confounds both relative model evaluations and claims about emergent capabilities. We argue that the seeming superiority of one model family over another may be explained by a different degree of training on the test task. To this end, we propose an effective method to adjust for the effect of training on the test task on benchmark evaluations. Put simply, to fine-tune each model under comparison on the same task-relevant data prior to evaluation. We then show that instances of emergent behavior disappear gradually as models train on the test task. Our work promotes a new perspective on the evaluation of large language models, with broad implications for benchmarking and the study of emergent capabilities.

In this work we explore recent advances in instruction-tuning language models on a range of open instruction-following datasets. Despite recent claims that open models can be on par with state-of-the-art proprietary models, these claims are often accompanied by limited evaluation, making it difficult to compare models across the board and determine the utility of various resources. We provide a large set of instruction-tuned models from 6.7B to 65B parameters in size, trained on 12 instruction datasets ranging from manually curated (e.g., OpenAssistant) to synthetic and distilled (e.g., Alpaca) and systematically evaluate them on their factual knowledge, reasoning, multilinguality, coding, and open-ended instruction following abilities through a collection of automatic, model-based, and human-based metrics. We further introduce T\"ulu, our best performing instruction-tuned model suite finetuned on a combination of high-quality open resources. Our experiments show that different instruction-tuning datasets can uncover or enhance specific skills, while no single dataset (or combination) provides the best performance across all evaluations. Interestingly, we find that model and human preference-based evaluations fail to reflect differences in model capabilities exposed by benchmark-based evaluations, suggesting the need for the type of systemic evaluation performed in this work. Our evaluations show that the best model in any given evaluation reaches on average 83% of ChatGPT performance, and 68% of GPT-4 performance, suggesting that further investment in building better base models and instruction-tuning data is required to close the gap. We release our instruction-tuned models, including a fully finetuned 65B T\"ulu, along with our code, data, and evaluation framework at https://github.com/allenai/open-instruct to facilitate future research.

Instruction-tuned large language models have revolutionized natural language processing and have shown great potential in applications such as conversational agents. These models, such as GPT-4, can not only master language but also solve complex tasks in areas like mathematics, coding, medicine, and law. Despite their impressive capabilities, there is still a lack of comprehensive understanding regarding their full potential, primarily due to the black-box nature of many models and the absence of holistic evaluation studies. To address these challenges, we present INSTRUCTEVAL, a more comprehensive evaluation suite designed specifically for instruction-tuned large language models. Unlike previous works, our evaluation involves a rigorous assessment of models based on problem-solving, writing ability, and alignment to human values. We take a holistic approach to analyze various factors affecting model performance, including the pretraining foundation, instruction-tuning data, and training methods. Our findings reveal that the quality of instruction data is the most crucial factor in scaling model performance. While open-source models demonstrate impressive writing abilities, there is substantial room for improvement in problem-solving and alignment. We are encouraged by the rapid development of models by the open-source community, but we also highlight the need for rigorous evaluation to support claims made about these models. Through INSTRUCTEVAL, we aim to foster a deeper understanding of instruction-tuned models and advancements in their capabilities. INSTRUCTEVAL is publicly available at https://github.com/declare-lab/instruct-eval.

Standard evaluation protocols reveal a counterintuitive phenomenon: on 7.7% of benchmark problems spanning five datasets, larger language models underperform smaller ones by 28.4 percentage points despite 10-100x more parameters. Through systematic evaluation of 31 models (0.5B-405B parameters) across 1,485 problems, we identify the mechanism as spontaneous scale-dependent verbosity that introduces errors through overelaboration. Causal intervention experiments demonstrate this reflects correctable prompt design rather than fundamental capability limitations. Constraining large models to produce brief responses improves accuracy by 26 percentage points and reduces performance gaps by up to two-thirds. Most critically, brevity constraints completely reverse performance hierarchies on mathematical reasoning and scientific knowledge benchmarks, with large models achieving 7.7-15.9 percentage point advantages over small models -- direct inversions of the original gaps. These reversals prove large models possess superior latent capabilities that universal prompting masks. We validate findings through three independent contamination tests and demonstrate inverse scaling operates continuously across the full parameter spectrum, with dataset-specific optimal scales ranging from 0.5B to 3.0B parameters. Our results establish that maximizing large model performance requires scale-aware prompt engineering rather than universal evaluation protocols, with immediate implications for deployment: prompt adaptation simultaneously improves accuracy and reduces computational costs.

This paper bridges internal and external analysis approaches to large language models (LLMs) by demonstrating that geometric properties of internal model representations serve as reliable proxies for evaluating generated text quality. We validate a set of metrics including Maximum Explainable Variance, Effective Rank, Intrinsic Dimensionality, MAUVE score, and Schatten Norms measured across different layers of LLMs, demonstrating that Intrinsic Dimensionality and Effective Rank can serve as universal assessments of text naturalness and quality. Our key finding reveals that different models consistently rank text from various sources in the same order based on these geometric properties, indicating that these metrics reflect inherent text characteristics rather than model-specific artifacts. This allows a reference-free text quality evaluation that does not require human-annotated datasets, offering practical advantages for automated evaluation pipelines.

Language models are increasingly used in scientific discovery to generate hypotheses, propose candidate solutions, implement systems, and iteratively refine them. At the core of these trial-and-error loops lies evaluation: the process of obtaining feedback on candidate solutions via verifiers, simulators, or task-specific scoring functions. While prior work has highlighted the importance of evaluation, it has not explicitly formulated the problem of how evaluation-driven discovery loops can be scaled up in a principled and effective manner to push the boundaries of scientific discovery, a problem this paper seeks to address. We introduce Simple Test-time Evaluation-driven Scaling (SimpleTES), a general framework that strategically combines parallel exploration, feedback-driven refinement, and local selection, revealing substantial gains unlocked by scaling evaluation-driven discovery loops along the right dimensions. Across 21 scientific problems spanning six domains, SimpleTES discovers state-of-the-art solutions using gpt-oss models, consistently outperforming both frontier-model baselines and sophisticated optimization pipelines. Particularly, we sped up the widely used LASSO algorithm by over 2x, designed quantum circuit routing policies that reduce gate overhead by 24.5%, and discovered new Erdos minimum overlap constructions that surpass the best-known results. Beyond novel discoveries, SimpleTES produces trajectory-level histories that naturally supervise feedback-driven learning. When post-trained on successful trajectories, models not only improve efficiency on seen problems but also generalize to unseen problems, discovering solutions that base models fail to uncover. Together, our results establish effective evaluation-driven loop scaling as a central axis for advancing LLM-driven scientific discovery, and provide a simple yet practical framework for realizing these gains.

Both capability and safety benchmarks rest upon the assumption that the behavior of language models undergoing a test is informative about their behavior in deployment. This assumption can fail, should models infer that they are being evaluated and condition their response on such context. This hypothesis, termed ``evaluation awareness'', has been observed in frontier and open-weight language models alike. We provide a systematic study of this phenomenon, by probing for it across six language models (from four families and three sizes) and three metrics. More precisely, we examine whether (i) being under evaluation is linearly represented within the models' activations space, (ii) it is verbalized in their output tokens (as scored by an LLM-as-judge), and (iii) steering causally affects their behavior. For the open-checkpoint Olmo models, we further test these measures at every training stage. In doing so, we report that evaluation awareness is linearly decodable from the residual streams of every model (best AUROC geq 0.7). By contrast, these representations align only in part with verbalization: their correlations and mutual information are nonzero in some settings, yet vary substantially across models, layers, and readout choices. Nevertheless, steering along probe-derived directions can shift the verbalization scores. Finally, a comparison across the Olmo checkpoints reveals that evaluation awareness is already present within base models, becomes amplified throughout the stages of supervised fine-tuning, and remains stable thereafter---unlike the effects of steering, that grow more pronounced at every successive training stage. These results show the need for evaluations to account for the disjunction between what models represent internally, what they verbalize, and their steering.