LLM Evaluation

LLM: Large Language Model

Follow this topic Ask about this topic

The field studies how to measure language-model capabilities, reliability, and risks, and whether those measurements support sound comparisons and deployment decisions. Approaches include curated and synthetic benchmarks, psychometric and adaptive testing, execution-based checks, robustness tests, and human or LLM-based judges. Researchers also examine calibration, uncertainty, contamination, prompt sensitivity, and whether scores reflect the intended capability across languages and domains. A central tension is that efficient aggregate scores can hide important errors and measurement biases, while more realistic, human-grounded evaluations are costly and difficult to standardize.

Latest papers

2,432 papers286 papers in the last four weeks, up 16% on the four weeks before. 1.7% of all new papers.

Mar 5, 2026

FireBench: Evaluating Instruction Following in Enterprise and API-Driven LLM Applications

Instruction following is critical for LLMs deployed in enterprise and API-driven settings, where strict adherence to output formats, content constraints, and procedural requirements is essential for enabling reliable LLM-assisted workflows. However, existing instruction following benchmarks predominantly evaluate natural language generation constraints that reflect the needs of chat assistants rather than enterprise users. To bridge this gap, we introduce FireBench, an LLM instruction following benchmark grounded in real-world enterprise and API usage patterns. FireBench evaluates six core capability dimensions across diverse applications including information extraction, customer support, and coding agents, comprising over 2,400 samples. We evaluate 11 LLMs and present key findings on their instruction following behavior in enterprise scenarios. We open-source FireBench at fire-bench.com to help users assess model suitability, support model developers in diagnosing performance, and invite community contributions.

Yunfan Zhang, Yijie Bei, Jetashree Ravi +1

Columbia University · Fireworks AI

LLM Evaluation Instruction Following Generative AI Evaluation Language Modeling

Mar 5, 2026

Spatial Competence Benchmark

Spatial competence is the quality of maintaining a consistent internal representation of an environment and using it to infer discrete structure and plan actions under constraints. Prevailing spatial evaluations for large models are limited to probing isolated primitives through 3D transformations or visual question answering. We introduce the Spatial Competence Benchmark (SCBench), spanning three hierarchical capability buckets whose tasks require executable outputs verified by deterministic checkers or simulator-based evaluators. On SCBench, three frontier models exhibit monotonically decreasing accuracy up the capability ladder. Sweeping output-token caps shows that accuracy gains concentrate at low budgets and saturate quickly, and failures are dominated by locally plausible geometry that breaks global constraints. We release the task generators, verifiers, and visualisation tooling.

Jash Vira, Ashley Harris

Independent Researcher · Maptek

Spatial Reasoning Benchmarks Spatial Reasoning LLM Evaluation

Mar 4, 2026

Towards Natural Personalization: Evaluating Long-Horizon Preference Following in Personalized User-LLM Interactions

Large Language Models (LLMs) are increasingly serving as personal assistants, where users may share individual preferences over extended interactions. However, assessing how well LLMs can follow these preferences in natural, long-term situations remains underexplored. This work proposes RealPref, a benchmark for evaluating natural preference-following in personalized user-LLM interactions. RealPref features 100 synthetic user profiles, 1300 personalized preferences, 4 types of preference expression (from explicit to implicit), and long-horizon interaction histories. It explored three types of test tasks (multiple-choice, true-or-false, and open-ended), with granular rubrics for LLM-as-a-judge evaluation. Results indicate that LLM performance drops significantly as context length grows and preference expression becomes more implicit, and that generalizing user preference understanding to unseen scenarios poses further challenges. RealPref and these findings provide a foundation for future research to develop user-aware LLM assistants that better adapt to individual needs.

Qianyun Guo, Yibo Li, Yue Liu +1

National University of Singapore, Singapore.

LLM Evaluation LLM Personalization User Preference Modeling

Mar 4, 2026

Bielik-Q2-Sharp: A Comparative Study of Extreme 2-bit Quantization Methods for a Polish 11B Language Model

We present Bielik-Q2-Sharp, the first systematic academic evaluation of extreme 2-bit quantization applied to a Polish large language model. Using Bielik-11B-v2.3-Instruct (11B parameters, Mistral architecture) as our base model, we compare six state-of-the-art post-training quantization methods -- QuIP#, SpinQuant+GPTQ, ButterflyQuant, QTIP, VPTQ, and AQLM -- all calibrated on a Polish-language corpus (CulturaX-PL) with shared Hessian matrices. Our best variant (QuIP# E8P12) achieves 71.92% across 22 Polish benchmarks versus 72.07% for the IQ2_XXS baseline -- within statistical noise, at a modest size premium (3.26 GB vs. ~2.6 GB). On eq_bench, our method scores 47.14 versus 43.53 (+3.6pp), suggesting superior preservation of higher-order reasoning. QTIP achieves the best per-bit efficiency (79.4% MC acc_norm at ~2.4 bpw, 3.27 GB), matching VPTQ's quality at 35% smaller size. We additionally document a MC-generation dissociation phenomenon where rotation-based methods preserve log-likelihood quality but fail catastrophically at autoregressive generation. The entire project was conducted by a single independent researcher on cloud GPUs (vast.ai) within a $285 budget. All models, Hessians, and evaluation logs are publicly available.

Jakub Prejzner

BitSharp, Independent Researcher, Rzeszów, Poland

Quantization LLM Quantization Post-Training Quantization Low-Bit Quantization Language Modeling LLM Evaluation

Mar 4, 2026

Traces of Social Competence in Large Language Models

The False Belief Test (FBT) has been the main method for assessing Theory of Mind (ToM) and related socio-cognitive competencies. For Large Language Models (LLMs), the reliability and explanatory potential of this test have remained limited due to issues like data contamination, insufficient model details, and inconsistent controls. We address these issues by testing 17 open-weight models on a balanced set of 192 FBT variants (Trott et al., 2023) using Bayesian Logistic regression to identify how model size and post-training affect socio-cognitive competence. We find that scaling model size benefits performance, but not strictly. A cross-over effect reveals that explicating propositional attitudes (X thinks) fundamentally alters response patterns. Instruction tuning partially mitigates this effect, but further reasoning-oriented fine-tuning amplifies it. In a case study analysing social reasoning ability throughout OLMo 2 training, we show that this cross-over effect emerges during pre-training, suggesting that models acquire stereotypical response patterns tied to mental-state vocabulary that can outweigh other scenario semantics. Finally, vector steering allows us to isolate a think vector as the causal driver of observed FBT behaviour.

Tom Kouwenhoven, Michiel van der Meer, Max van Duijn

Leiden Institute of Advanced Computer Science Leiden University, Netherlands

Theory of Mind LLM Evaluation LLM Fine-Tuning

Mar 4, 2026

FINEST: Improving LLM Responses to Sensitive Topics Through Fine-Grained Evaluation

Large Language Models (LLMs) often generate overly cautious and vague responses on sensitive topics, sacrificing helpfulness for safety. Existing evaluation frameworks lack systematic methods to identify and address specific weaknesses in responses to sensitive topics, making it difficult to improve both safety and helpfulness simultaneously. To address this, we introduce FINEST, a FINE-grained response evaluation taxonomy for Sensitive Topics, which breaks down helpfulness and harmlessness into errors across three main categories: Content, Logic, and Appropriateness. Experiments on a Korean-sensitive question dataset demonstrate that our score- and error-based improvement pipeline, guided by FINEST, significantly improves the model responses across all three categories, outperforming refinement without guidance. Notably, score-based improvement -- providing category-specific scores and justifications -- yields the most significant gains, reducing the error sentence ratio for Appropriateness by up to 33.09%. This work lays the foundation for a more explainable and comprehensive evaluation and improvement of LLM responses to sensitive questions.

Juhyun Oh, Nayeon Lee, Chani Jung +5

School of Computing, KAIST, Daejeon, Republic of Korea · Samsung, Suwon, Republic of Korea

Language Model Generation Evaluation Language Model Safety Evaluation LLM Evaluation Question Answering

Mar 4, 2026

A Multi-Dimensional Quality Scoring Framework for Decentralized LLM Inference with Proof of Quality

Decentralized large language model (LLM) inference networks can pool heterogeneous compute to scale serving, but they require lightweight and incentive-compatible mechanisms to assess output quality. Prior work introduced cost-aware Proof of Quality (PoQ) and adaptive robust PoQ to allocate rewards under evaluator heterogeneity and adversarial behavior. In this paper, we focus on the quality signal itself and propose a multi-dimensional quality scoring framework that decomposes output quality into modular dimensions, including model and cost priors, structure quality, semantic quality, query-output alignment, and agreement/uncertainty. Using logged outputs from QA and summarization tasks, we systematically audit dimension reliability and show that seemingly reasonable dimensions can be task-dependent and even negatively correlated with reference quality without calibration. While the default composite underperforms a strong single semantic evaluator, ablations reveal that removing unreliable dimensions and re-normalizing weights yields a calibrated composite that matches or exceeds the best single- evaluator and consensus baselines. Finally, we integrate the composite score as a drop-in quality signal in PoQ and demonstrate complementary benefits with robust aggregation and adaptive trust weighting under adversarial evaluator attacks.

Arther Tian, Alex Ding, Frank Chen +2

DGrid AI

Language Model Generation Evaluation LLM Evaluation LLM Inference LLM Auditing

Mar 4, 2026

Evaluating Large Language Models for Gait Classification Using Text-Encoded Kinematic Waveforms

Background: Machine learning (ML) enhances gait analysis but often lacks the level of interpretability desired for clinical adoption. Large Language Models (LLMs) may offer explanatory capabilities and confidence-aware outputs when applied to structured kinematic data. This study therefore evaluated whether general-purpose LLMs can classify continuous gait kinematics when represented as textual numeric sequences and how their performance compares to conventional ML approaches. Methods: Lower-body kinematics were recorded from 20 participants performing seven gait patterns. A supervised KNN classifier and a class-independent One-Class SVM (OCSVM) were compared against zero-shot LLMs (GPT-5, GPT-5-mini, GPT-4.1, and o4-mini). Models were evaluated using Leave-One-Subject-Out (LOSO) cross-validation. LLMs were tested both with and without explicit reference gait statistics. Results: The supervised KNN achieved the highest performance (multiclass Matthews Correlation Coefficient, MCC = 0.88). The best-performing LLM (GPT-5) with reference grounding achieved a multiclass MCC of 0.70 and a binary MCC of 0.68, outperforming the class-independent OCSVM (binary MCC = 0.60). Performance of the LLM was highly dependent on explicit reference information and self-rated confidence; when restricted to high-confidence predictions, multiclass MCC increased to 0.83 on the filtered subset. Notably, the computationally efficient o4-mini model performed comparably to larger models. Conclusion: When continuous kinematic waveforms were encoded as textual numeric tokens, general-purpose LLMs, even with reference grounding, did not match supervised multiclass classifiers for precise gait classification and are better regarded as exploratory systems requiring cautious, human-guided interpretation rather than diagnostic use.

Carlo Dindorf, Jonas Dully, Rebecca Keilhauer +2

Department of Sports Science, University of Kaiserslautern-Landau (RPTU), Kaiserslautern, Germany · Fraunhofer Institute for Experimental Software Engineering IESE, Kaiserslautern, Germany · German Research Center of Artificial Intelligence, Kaiserslautern, Germany

LLM Evaluation Gait Analysis Gait Recognition LLM Grounding

Mar 4, 2026

CzechTopic: A Benchmark for Zero-Shot Topic Localization in Historical Czech Documents

Topic localization aims to identify spans of text that express a given topic defined by a name and description. To study this task, we introduce a human-annotated benchmark based on Czech historical documents, containing human-defined topics together with manually annotated spans and supporting evaluation at both document and word levels. Evaluation is performed relative to human agreement rather than a single reference annotation. We evaluate a diverse range of large language models alongside BERT-based models fine-tuned on a distilled development dataset. Results reveal substantial variability among LLMs, with performance ranging from near-human topic detection to pronounced failures in span localization. While the strongest models approach human agreement, the distilled token embedding models remain competitive despite their smaller scale. The dataset and evaluation framework are publicly available at: https://github.com/dcgm/czechtopic.

Martin Kostelník, Michal Hradiš, Martin Dočekal

Faculty of Information Technology, Brno University of Technology, Brno, Czech Republic

Document Understanding Historical Document Analysis LLM Evaluation Low-Resource Language Processing Historical Document Processing

Mar 4, 2026

In-Context Environments Induce Evaluation-Awareness in Language Models

Humans often become more self-aware under threat, yet can lose self-awareness when absorbed in a task; we hypothesize that language models exhibit environment-dependent evaluation awareness. This raises concerns that models could strategically underperform, or sandbag, to avoid triggering capability-limiting interventions such as unlearning or shutdown. Prior work demonstrates sandbagging under hand-crafted prompts, but this underestimates the true vulnerability ceiling. We introduce a black-box adversarial optimization framework treating the in-context prompt as an optimizable environment, and develop two approaches to characterize sandbagging: (1) measuring whether models expressing intent to underperform can actually execute it across different task structures, and (2) causally isolating whether underperformance is driven by genuine evaluation-aware reasoning or shallow prompt-following. Evaluating Claude-3.5-Haiku, GPT-4o-mini, and Llama-3.3-70B across four benchmarks (Arithmetic, GSM8K, MMLU, and HumanEval), optimized prompts induce up to 94 percentage point (pp) degradation on arithmetic (GPT-4o-mini: 97.8%

\rightarrow

4.0%), far exceeding hand-crafted baselines which produce near-zero behavioral change. Code generation exhibits model-dependent resistance: Claude degrades only 0.6pp, while Llama's accuracy drops to 0%. The intent -- execution gap reveals a monotonic resistance ordering: Arithmetic

<

GSM8K

<

MMLU, demonstrating that vulnerability is governed by task structure rather than prompt strength. CoT causal intervention confirms that 99.3% of sandbagging is causally driven by verbalized eval-aware reasoning, ruling out shallow instruction-following. These findings demonstrate that adversarially optimized prompts pose a substantially greater threat to evaluation reliability than previously understood.

Maheep Chaudhary

Independent

Evaluation Awareness in Language Models LLM Evaluation Adversarial Attacks on LLMs

Mar 4, 2026

T2S-Bench & Structure-of-Thought: Benchmarking and Prompting Comprehensive Text-to-Structure Reasoning

Think about how human handles complex reading tasks: marking key points, inferring their relationships, and structuring information to guide understanding and responses. Likewise, can a large language model benefit from text structure to enhance text-processing performance? To explore it, in this work, we first introduce Structure of Thought (SoT), a prompting technique that explicitly guides models to construct intermediate text structures, consistently boosting performance across eight tasks and three model families. Building upon this insight, we present T2S-Bench, the first benchmark designed to evaluate and improve text-to-structure capabilities of models. T2S-Bench includes 1.8K samples across 6 scientific domains and 32 structural types, rigorously constructed to ensure accuracy, fairness, and quality. Evaluation on 45 mainstream models reveals substantial improvement potential: the average accuracy on the multi-hop reasoning task is only 52.1%, and even the most advanced model achieves 58.1% node accuracy in end-to-end extraction. Furthermore, on Qwen2.5-7B-Instruct, SoT alone yields an average +5.7% improvement across eight diverse text-processing tasks, and fine-tuning on T2S-Bench further increases this gain to +8.6%. These results highlight the value of explicit text structuring and the complementary contributions of SoT and T2S-Bench. Dataset and eval code have been released at https://t2s-bench.github.io/T2S-Bench-Page/.

Qinsi Wang, Hancheng Ye, Jinhee Kim +12

Duke University · UT Austin · Meta

Structured Reasoning LLM Reasoning LLM Evaluation LLM Prompting LLM Information Extraction

Mar 3, 2026

Tucano 2 Cool: Better Open Source LLMs for Portuguese

We present Tucano 2, a fully open suite of large language models (LLMs) with 0.5-3.7 billion parameters, designed to address certain gaps in open-source development for Portuguese LLMs. Following our previous works, we now extend our dataset, GigaVerbo-v2, to a new degree of quality and scale, while also introducing a new synthetic dataset, GigaVerbo-v2 Synth, aimed at filling missing gaps in GigaVerbo-v2, and two post-training datasets, GigaVerbo-v2 SFT and GigaVerbo-v2 Preferences, that allow Portuguese LLMs to be trained in domains like retrieval augmented generation, coding, tool use, chain-of-thought reasoning, and many other domains of interest. Through extensive ablation studies, we design both pretraining and continual pretraining recipes for the Tucano 2 suite (Base, Instruct, and Think), which achieve state-of-the-art performance on several Portuguese-language modeling benchmarks. We also extend and refine the evaluation harness introduced in our earlier work, yielding a comprehensive evaluation suite that provides strong signals across different pretraining, continual pretraining, and post-training regimes. All artifacts associated with Tucano 2 are openly released, including training recipes, logs, and source code, ensuring that our work is reproducible, accessible, and extendable by the broader Portuguese NLP community.

Nicholas Kluge Corrêa, Aniket Sen, Shiza Fatimah +4

Bonn-Aachen International Center for Information Technology (b-it) / CAISA Lab · Lamarr Institute for Machine Learning and Artificial Intelligence · Center for Science and Thought +2

Portuguese Language Models Language Modeling Language Model Pretraining Language Model Post-Training LLM Evaluation

Mar 3, 2026

Verify to Amplify: Improving Reasoning via Learned Chain-of-Thought Verification

Large Language Models (LLMs) using chain-of-thought have demonstrated great potential for solving complex reasoning and planning tasks. Despite these advances, LLM-generated outputs remain susceptible to errors, making verification important for reliable reasoning systems. Learned verifiers can increase trust, enforce safety constraints, and ensure alignment with personal preferences, while also providing feedback to improve generation. This raises a central challenge: when learned verifiers are used to guide generation, the feedback loop between generator and verifier may induce a distribution shift. This is particularly salient for process reward models, a prominent class of learned verifiers that score or classify individual steps in a chain-of-thought reasoning trace. Motivated by this challenge, we propose a new online learning framework for chain-of-thought verifiers that, given a problem statement and a reasoning trace, check the correctness of each reasoning step given the preceding steps. Highlighting the asymmetric role of soundness errors (accepting an incorrect reasoning step) and completeness errors (flagging a correct step as wrong), we introduce novel notions of dimension that characterize their optimal tradeoff. We then show how our learned verifiers can boost the accuracy of a weak generator. Assuming that the generator can produce a correct next step with a small success probability, we show how to learn a strong generator with small error and abstention rates. Our results also allow learning from offline data when queries to an expert verifier can be simulated from a small set of correct reasoning traces. However, we establish a separation between our approach and learning from offline expert demonstrations: we show that learning from offline demonstrations cannot in general achieve the soundness-completeness guarantees produced by our interactive learning approach.

Maria-Florina Balcan, Avrim Blum, Kiriaki Fragkia +2

Carnegie Mellon University · Toyota Technological Institute at Chicago · Northwestern University

LLM Evaluation LLM Answer Verification Process Reward Models LLM Reliability LLM Reasoning

Mar 3, 2026

Chart Deep Research in LVLMs via Parallel Relative Policy Optimization

With the rapid advancement of data science, charts have evolved from simple numerical presentation tools to essential instruments for insight discovery and decision-making support. However, current chart data intelligence exhibits significant limitations in deep research capabilities, with existing methods predominantly addressing shallow tasks such as visual recognition or factual question-answering, rather than the complex reasoning and high-level data analysis that deep research requires. This limitation stems from two primary technical bottlenecks: at the training level, existing post-training techniques exhibit deficiencies in handling multi-dimensional reward signal interference and heterogeneous data gradient conflicts, preventing models from achieving balanced development across multiple capability dimensions; at the evaluation level, current methods remain limited to factual retrieval and basic computation, failing to assess end-to-end analytic reasoning and other deep research capabilities. To address the training challenge, we propose PRPO, which performs parallel optimization across reward dimensions and capability partitioning across data types, effectively disentangling conflicts between heterogeneous data and multi-dimensional reward signals while ensuring optimization stability. For the evaluation challenge, we construct MCDR-Bench based on the ``error uniqueness principle," transforming subjective generation assessment into objective error identification through controllable error injection, enabling quantifiable evaluation of deep research capabilities. Experimental validation confirms that the proposed PRPO and MCDR-Bench jointly establish a unified framework that systematically advances chart deep research through enhanced collaborative training and objective evaluation.

Jiajin Tang, Gaoyang, Wenjie Wang +2

ByteDance · ShanghaiTech University · School of Computer Science and Engineering, Sun Yat-sen University

Chart Understanding Large Vision-Language Models LLM Evaluation Reasoning Evaluation RL for Language Model Reasoning

Mar 3, 2026

Faster, Cheaper, More Accurate: Specialised Knowledge Tracing Models Outperform LLMs

Predicting future student responses to questions is particularly valuable for educational learning platforms where it enables effective interventions. One of the key approaches to do this has been through the use of knowledge tracing (KT) models. These are small, domain-specific, temporal models trained on student question-response data. KT models are optimised for high accuracy on specific educational domains and have fast inference and scalable deployments. The rise of Large Language Models (LLMs) motivates us to ask the following questions: (1) How well can LLMs perform at predicting students' future responses to questions? (2) Are LLMs scalable for this domain? (3) How do LLMs compare to KT models on this domain-specific task? In this paper, we compare multiple LLMs and KT models across predictive performance, deployment cost, and inference speed to answer the above questions. We show that KT models outperform LLMs with respect to accuracy and F1 scores on this domain-specific task. Further, we demonstrate that LLMs are orders of magnitude slower than KT models and cost orders of magnitude more to deploy. This highlights the importance of domain-specific models for education prediction tasks and the fact that current closed source LLMs should not be used as a universal solution for all tasks.

Prarthana Bhattacharyya, Joshua Mitton, Ralph Abboud +1

Eedi · Learning Engineering Virtual Institute

Knowledge Tracing AI in Education LLM Evaluation

Mar 3, 2026

OCR or Not? Rethinking Document Information Extraction in the MLLMs Era with Real-World Large-Scale Datasets

Multimodal Large Language Models (MLLMs) enhance the potential of natural language processing. However, their actual impact on document information extraction remains unclear. In particular, it is unclear whether an MLLM-only pipeline--while simpler--can truly match the performance of traditional OCR+MLLM setups. In this paper, we conduct a large-scale benchmarking study that evaluates various out-of-the-box MLLMs on business-document information extraction. To examine and explore failure modes, we propose an automated hierarchical error analysis framework that leverages large language models (LLMs) to diagnose error patterns systematically. Our findings suggest that OCR may not be necessary for powerful MLLMs, as image-only input can achieve comparable performance to OCR-enhanced approaches. Moreover, we demonstrate that carefully designed schema, exemplars, and instructions can further enhance MLLMs performance. We hope this work can offer practical guidance and valuable insight for advancing document information extraction.

Jiyuan Shen, Peiyue Yuan, Atin Ghosh +2

SAP · Stanford University

Document Information Extraction Multimodal Large Language Models LLM Evaluation

Mar 3, 2026

From Solver to Tutor: Evaluating the Pedagogical Intelligence of LLMs with KMP-Bench

Large Language Models (LLMs) show significant potential in AI mathematical tutoring, yet current evaluations often rely on simplistic metrics or narrow pedagogical scenarios, failing to assess comprehensive, multi-turn teaching effectiveness. In this paper, we introduce KMP-Bench, a comprehensive K-8 Mathematical Pedagogical Benchmark designed to assess LLMs from two complementary perspectives. The first module, KMP-Dialogue, evaluates holistic pedagogical capabilities against six core principles (e.g., Challenge, Explanation, Feedback), leveraging a novel multi-turn dialogue dataset constructed by weaving together diverse pedagogical components. The second module, KMP-Skills, provides a granular assessment of foundational tutoring abilities, including multi-turn problem-solving, error detection and correction, and problem generation. Our evaluations on KMP-Bench reveal a key disparity: while leading LLMs excel at tasks with verifiable solutions, they struggle with the nuanced application of pedagogical principles. Additionally, we present KMP-Pile, a large-scale (150K) dialogue dataset. Models fine-tuned on KMP-Pile show substantial improvement on KMP-Bench, underscoring the value of pedagogically-rich training data for developing more effective AI math tutors.

Weikang Shi, Houxing Ren, Junting Pan +8

Multimedia Laboratory (MMLab), The Chinese University of Hong Kong · CPII under InnoHK · Shenzhen Loop Area Institute

Generative AI in Education Intelligent Tutoring Systems LLM Evaluation Mathematical Reasoning Benchmarks

Mar 3, 2026

How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities

Large Language Models (LLMs) are increasingly deployed in socially sensitive domains, yet their unpredictable behaviors, ranging from misaligned intent to inconsistent personality, pose significant risks. We introduce SteerEval, a hierarchical benchmark for evaluating LLM controllability across three domains: language features, sentiment, and personality. Each domain is structured into three specification levels: L1 (what to express), L2 (how to express), and L3 (how to instantiate), connecting high-level behavioral intent to concrete textual output. Using SteerEval, we systematically evaluate contemporary steering methods, revealing that control often degrades at finer-grained levels. Our benchmark offers a principled and interpretable framework for safe and controllable LLM behavior, serving as a foundation for future research.

Ziwen Xu, Kewei Xu, Haoming Xu +8

Zhejiang University · Alibaba Group

LLM Evaluation Language Model Steering

Mar 3, 2026

A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities

Large language models (LLMs) display a unified "general factor" of capability across 10 benchmarks (a finding confirmed by our factor analysis of 156 models), yet they still struggle with simple, trivial tasks for humans. This is because current benchmarks focus on task completion, failing to probe the foundational cognitive abilities that highlight these behaviors. We address this by introducing the NeuroCognition benchmark, grounded in three adapted neuropsychological tests targeting distinct foundational cognitive components: Raven's Progressive Matrices (abstract relational reasoning), Spatial Working Memory (goal-directed spatial updating), and the Wisconsin Card Sorting Test (cognitive flexibility). Our evaluation reveals that while models perform strongly on text, their performance degrades for images and with increased complexity. Comparison with a human baseline shows that LLMs and humans fail at different parts of the same tasks. Furthermore, we observe that complex reasoning is not universally beneficial, whereas simple, human-like strategies yield partial gains. We also find that NeuroCognition correlates positively with standard general-capability benchmarks, while still measuring distinct cognitive abilities beyond them. Overall, NeuroCognition emphasizes where current LLMs align with human-like intelligence and where they lack core adaptive cognition, showing the potential to serve as a verifiable, scalable source for improving LLMs.

Faiz Ghifari Haznitrama, Faeyza Rishad Ardi, Alice Oh

School of Computing, KAIST, Daejeon, South Korea

Mar 2, 2026

Pencil Puzzle Bench: A Benchmark for Multi-Step Verifiable Reasoning

We introduce Pencil Puzzle Bench, a framework for evaluating large language model reasoning through pencil puzzles, a family of constraint-satisfaction problems closely related to NP-complete problems, with deterministic, step-level verification. From a database of 62,231 puzzles across 94 varieties with verified unique solutions, we select a benchmark of 300 puzzles spanning 20 varieties and evaluate 51 models from 11 providers in two modes: direct ask (single-shot) and agentic (multi-turn with iterative verification). A key differentiator of our benchmark is that every intermediate board state can be checked against variety-specific constraints, localizing errors to the exact rule violated, providing the infrastructure for dense, per-move reward signals for process supervision and reinforcement learning. Our evaluation reveals two distinct axes of capability: (1) reasoning effort scaling, where GPT-5.2 improves 81x from no reasoning to maximum effort; and (2) agentic iteration, where Claude Opus 4.6 rises from 0.3% to 30.0% through iterative checking, while GPT-5.2@xhigh improves from 20.2% to 56.0%. Agentic attempts span a median of 29 turns over 17 minutes, with the longest exceeding 1,221 turns and 14.3 hours - a demanding test of long-context utilization, not just reasoning.

Justin Waugh

Approximate Labs

LLM Evaluation LLM Reasoning AI Agent Benchmarks Constraint Satisfaction

Mar 2, 2026

CharacterFlywheel: Scaling Iterative Improvement of Engaging and Steerable LLMs in Production

This report presents CharacterFlywheel, an iterative flywheel process for improving large language models (LLMs) in production social chat applications across Instagram, WhatsApp, and Messenger. Starting from LLaMA 3.1, we refined models across 15 generations using data from both internal and external real-user traffic. Through continuous deployments from July 2024 to April 2025, we conducted controlled 7-day A/B tests showing consistent engagement improvements: 7 of 8 newly deployed models demonstrated positive lift over the baseline, with the strongest performers achieving up to 8.8% improvement in engagement breadth and 19.4% in engagement depth. We also observed substantial gains in steerability, with instruction following increasing from 59.2% to 84.8% and instruction violations decreasing from 26.6% to 5.8%. We detail the CharacterFlywheel process which integrates data curation, reward modeling to estimate and interpolate the landscape of engagement metrics, supervised fine-tuning (SFT), reinforcement learning (RL), and both offline and online evaluation to ensure reliable progress at each optimization step. We also discuss our methods for overfitting prevention and navigating production dynamics at scale. These contributions advance the scientific rigor and understanding of LLMs in social applications serving millions of users.

Yixin Nie, Lin Guan, Zhongyao Ma +19

Meta Superintelligence Labs · OpenAI · xAI +1

LLM Post-Training Conversational Agents Language Model Self-Improvement LLM Evaluation

Mar 2, 2026

CyclicJudge: Mitigating Judge Bias Efficiently in LLM-based Evaluation

LLM-as-judge evaluation has become standard practice for open-ended model assessment; however, judges exhibit systematic biases that cannot be averaged out by increasing the number of scenarios or generations. These biases are often similar in magnitude to the model differences that benchmarks are designed to detect, resulting in unreliable rankings when single-judge evaluations are used. We introduce a variance decomposition that partitions benchmark score variance into scenario, generation, judge, and residual components. Based on this analysis, CyclicJudge, a round-robin assignment of judges to scenarios, is demonstrated to be the optimal strategy for a fixed judge panel and judge-call budget: the score recovers the panel mean exactly while matching the cost of single-judge evaluation. Empirical results on MT-Bench and MindEval validate the effectiveness of CyclicJudge as predicted, across both general-purpose and domain-specific evaluation settings.

Ziyi Zhu, Olivier Tieleman, Alexey Bukhtiyarov +1

Slingshot AI · Department of Engineering, University of Cambridge

LLM Evaluation Bias in LLM-as-a-Judge LLM-as-a-Judge

Mar 2, 2026

Probing Materials Knowledge in LLMs: From Latent Embeddings to Reliable Predictions

Large language models are increasingly applied to materials science, yet fundamental questions remain about their reliability and knowledge encoding. Evaluating 25 LLMs across four materials science tasks -- over 200 base and fine-tuned configurations -- we find that output modality fundamentally determines model behavior. For symbolic tasks, fine-tuning converges to consistent, verifiable answers with reduced response entropy, while for numerical tasks, fine-tuning improves prediction accuracy but models remain inconsistent across repeated inference runs, limiting their reliability as quantitative predictors. For numerical regression, we find that better performance can be obtained by extracting embeddings directly from intermediate transformer layers than from model text output, revealing an ``LLM head bottleneck,'' though this effect is property- and dataset-dependent. Finally, we present a longitudinal study of GPT model performance in materials science, tracking four models over 18 months and observing 9--43% performance variation that poses reproducibility challenges for scientific applications.

Vineeth Venugopal, Soroush Mahjoubi, Elsa Olivetti

Department of Materials Science and Engineering, Massachusetts Institute of Technology, Cambridge, 02139, MA, USA.

LLM Reliability LLM Evaluation Language Model Probing LLM Fine-Tuning Numerical Reasoning in Language Models Materials Property Prediction

Mar 2, 2026

RubricBench: Aligning Model-Generated Rubrics with Human Standards

As Large Language Model (LLM) alignment evolves from simple completions to complex, highly sophisticated generation, Reward Models are increasingly shifting toward rubric-guided evaluation to mitigate surface-level biases. However, the community lacks a unified benchmark to assess this evaluation paradigm, as existing benchmarks lack both the discriminative complexity and the ground-truth rubric annotations required for rigorous analysis. To bridge this gap, we introduce RubricBench, a curated benchmark with 1,147 pairwise comparisons specifically designed to assess the reliability of rubric-based evaluation. Our construction employs a multi-dimensional filtration pipeline to target hard samples featuring nuanced input complexity and misleading surface bias, augmenting each with expert-annotated, atomic rubrics derived strictly from instructions. Comprehensive experiments reveal a substantial capability gap between human-annotated and model-generated rubrics, indicating that even state-of-the-art models struggle to autonomously specify valid evaluation criteria, lagging considerably behind human-guided performance.

Qiyuan Zhang, Junyi Zhou, Yufei Wang +8

City University of Hong Kong · Tencent Hunyuan · McGill - Mila & Quebec AI Institute +2

LLM Evaluation Rubric-Based Evaluation Automatic Rubric Generation LLM Reliability LLM-as-a-Judge Rubric-Guided Generation

Mar 1, 2026

DEP: A Decentralized Large Language Model Evaluation Protocol

With the rapid development of Large Language Models (LLMs), a large number of benchmarks have been proposed. However, most benchmarks lack unified evaluation standard and require the manual implementation of custom scripts, making results hard to ensure consistency and reproducibility. Furthermore, mainstream evaluation frameworks are centralized, with datasets and answers, which increases the risk of benchmark leakage. To address these issues, we propose a Decentralized Evaluation Protocol (DEP), a decentralized yet unified and standardized evaluation framework through a matching server without constraining benchmarks. The server can be mounted locally or deployed remotely, and once adapted, it can be reused over the long term. By decoupling users, LLMs, and benchmarks, DEP enables modular, plug-and-play evaluation: benchmark files and evaluation logic stay exclusively on the server side. In remote setting, users cannot access the ground truth, thereby achieving data isolation and leak-proof evaluation. To facilitate practical adoption, we develop DEP Toolkit, a protocol-compatible toolkit that supports features such as breakpoint resume, concurrent requests, and congestion control. We also provide detailed documentation for adapting new benchmarks to DEP. Using DEP toolkit, we evaluate multiple LLMs across benchmarks. Experimental results verify the effectiveness of DEP and show that it reduces the cost of deploying benchmark evaluations. As of February 2026, we have adapted over 60 benchmarks and continue to promote community co-construction to support unified evaluation across various tasks and domains.

Jianxiang Peng, Junhao Li, Hongxiang Wang +15

TJUNLP Lab, Tianjin University, Tianjin, China · National Supercomputing Center in Tianjin · Yuanhui AI

LLM Evaluation Benchmark Contamination

Mar 1, 2026

CollabEval: Enhancing LLM-as-a-Judge via Multi-Agent Collaboration

Large Language Models (LLMs) have revolutionized AI-generated content evaluation, with the LLM-as-a-Judge paradigm becoming increasingly popular. However, current single-LLM evaluation approaches face significant challenges, including inconsistent judgments and inherent biases from pre-training data. To address these limitations, we propose CollabEval, a novel multi-agent evaluation framework that implements a three-phase Collaborative Evaluation process: initial evaluation, multi-round discussion, and final judgment. Unlike existing approaches that rely on competitive debate or single-model evaluation, CollabEval emphasizes collaboration among multiple agents with strategic consensus checking for efficiency. Our extensive experiments demonstrate that CollabEval consistently outperforms single-LLM approaches across multiple dimensions while maintaining robust performance even when individual models struggle. The framework provides comprehensive support for various evaluation criteria while ensuring efficiency through its collaborative design.

Yiyue Qian, Shinan Zhang, Yun Zhou +3

Amazon AWS Generative AI Innovation Center · Amazon AWS Bedrock

LLM Evaluation LLM-as-a-Judge Language Model Generation Evaluation Multi-LLM Collaboration Bias in LLM-as-a-Judge

Mar 1, 2026

Knowledge without Wisdom: Measuring Misalignment between LLMs and Intended Impact

LLMs increasingly excel on AI benchmarks, but doing so does not guarantee validity for downstream tasks. This study contrasts LLM alignment on benchmarks, downstream tasks, and, importantly the intended impact of those tasks. We evaluate the performance of leading LLMs (i.e., generative pre-trained base models) on difficult-to-verify tasks of the teaching and learning of schoolchildren. Across all LLMs, inter-model behaviors on disparate tasks correlate higher than they do with expert human behaviors on target tasks. These biases shared across LLMs are poorly aligned with downstream measures of teaching quality and often negatively aligned with the intended impact of student learning outcomes. Further, we find multi-model ensembles, both unanimous model voting and expert-weighting by benchmark performance, further exacerbate misalignment with learning. We measure that selection of LLM and/or prompting strategy only reliably accounts for

15\%

of all measured misalignment error and that variation in misalignment error is shared across LLMs, suggesting that common pretraining accounts for much of the misalignment in these tasks. We demonstrate methods for robustly measuring alignment of complex tasks and provide unique insights into practical applications of LLMs in high-noise contexts.

Michael Hardy, Yunsung Kim

Stanford University

LLM Alignment LLM Evaluation AI in Education Language Model Bias Evaluation

Feb 28, 2026

Leveraging Computerized Adaptive Testing for Cost-effective Evaluation of Large Language Models in Medical Benchmarking

The rapid proliferation of large language models (LLMs) in healthcare creates an urgent need for scalable and psychometrically sound evaluation methods. Conventional static benchmarks are costly to administer repeatedly, vulnerable to data contamination, and lack calibrated measurement properties for fine-grained performance tracking. We propose and validate a computerized adaptive testing (CAT) framework grounded in item response theory (IRT) for efficient assessment of standardized medical knowledge in LLMs. The study comprises a two-phase design: a Monte Carlo simulation to identify optimal CAT configurations and an empirical evaluation of 38 LLMs using a human-calibrated medical item bank. Each model completed both the full item bank and an adaptive test that dynamically selected items based on real-time ability estimates and terminated upon reaching a predefined reliability threshold (standard error <= 0.3). Results show that CAT-derived proficiency estimates achieved a near-perfect correlation with full-bank estimates (r = 0.988) while using only 1.3 percent of the items. Evaluation time was reduced from several hours to minutes per model, with substantial reductions in token usage and computational cost, while preserving inter-model performance rankings. This work establishes a psychometric framework for rapid, low-cost benchmarking of foundational medical knowledge in LLMs. The proposed adaptive methodology is intended as a standardized pre-screening and continuous monitoring tool and is not a substitute for real-world clinical validation or safety-oriented prospective studies.

Tianpeng Zheng, Zhehan Jiang, Jiayi Liu +1

Institute of Medical Education, Health Science Center, Peking University, Beijing, China · School of Public Health, Peking University, Haidian District, Beijing, China · Peking University Health Science Center-Chaoxing Joint Laboratory for Digital and Smart Medical, Peking University, Beijing, China +1

Computerized Adaptive Testing LLM Evaluation Clinical Language Model Evaluation Educational Assessment

Feb 28, 2026

Super Research: Answering Highly Complex Questions with Large Language Models through Super Deep and Super Wide Research

While Large Language Models (LLMs) have demonstrated proficiency in Deep Research or Wide Search, their capacity to solve highly complex questions-those requiring long-horizon planning, massive evidence gathering, and synthesis across heterogeneous sources-remains largely unexplored. We introduce Super Research, a task for complex autonomous research tasks that integrates (i) structured decomposition into a research plan, (ii) super wide retrieval for diverse perspectives, and (iii) super deep investigation to resolve uncertainties through iterative queries. To evaluate this capability, we curated a benchmark of 300 expert-written questions across diverse domains, each requiring up to 100+ retrieval steps and 1,000+ web pages to reconcile conflicting evidence. Super Research produces verifiable reports with fine-grained citations and intermediate artifacts (e.g., outlines and tables) to ensure traceable reasoning. Furthermore, we present a graph-anchored auditing protocol that evaluates Super Research along five dimensions: Coverage, Logical Consistency, Report Utility, Objectivity and Citation Health. While super-complex questions may be infrequent in standard applications, Super Research serves as a critical ceiling evaluation and stress test for LLM capabilities. A model's proficiency within Super Research acts as a powerful proxy for its general research competence; success here suggests the robustness necessary to navigate nearly any subordinate research task. Leaderboard is available at: https://cnsdqd-dyb.github.io/Super-Research-Benchmark/

Yubo Dong, Nianhao You, Yuxuan Hou +5

College of Computer Science and Technology · Ant Group, Hangzhou, China

Autonomous Research Agents LLM Evaluation Deep Research Agents Knowledge-Intensive QA

Feb 28, 2026

Advancing Multimodal Judge Models through a Capability-Oriented Benchmark and MCTS-Driven Data Generation

Using Multimodal Large Language Models (MLLMs) as judges to achieve precise and consistent evaluations has gradually become an emerging paradigm across various domains. Evaluating the capability and reliability of MLLM-as-a-judge systems is therefore essential for ensuring trustworthy assessment. Existing judge benchmarks categorize samples by task types but fail to capture the fundamental judgment capabilities required for reliable evaluation. In this work, we introduce M-JudgeBench, a ten-dimensional capability-oriented benchmark designed to comprehensively assess the judgment abilities of MLLMs. Our benchmark decomposes evaluation into pairwise Chain-of-Thought (CoT) comparison, length bias avoidance, and process error detection tasks, jointly covering ten fine-grained subtasks. This design enables diagnosis of model reliability across reasoning styles, response lengths, and cross-model variations. Systematic evaluation uncovers the systematic weaknesses in existing MLLM-as-a-judge systems. To address this issue, we further propose Judge-MCTS, a data construction framework generating pairwise reasoning trajectories with various correctness and length. Using Judge-MCTS, we construct an MCTS-augmented dataset and train M-Judger, a series of strong judge models. Extensive experiments demonstrate the superiority of M-Judger on existing judge benchmarks as well as M-JudgeBench. Overall, our work establishes a more principled foundation for evaluating MLLM-as-a-judge through M-JudgeBench and Judge-MCTS framework, paving the way for future research on judge model evaluation and capability-driven judge training.

Zeyu Chen, Huanjin Yao, Ziwang Zhao +1

LLM Evaluation LLM-as-a-Judge Synthetic Data Generation Multimodal Large Language Models Multimodal Model Evaluation