LLM Evaluation

LLM: Large Language Model

Follow this topic Ask about this topic

Momentum

194 papers in the last four weeks, up 87% on the four weeks before. 1.9% of all new papers.

Jul 13Week of Sep 28

Latest papers 1,631

Oct 5, 2026cs.CL

JEV versus LLMs: Accuracy, Cost and Calibration on Seven Political Science Replications

Large language models (LLMs) annotate and scale political text or constructs by generating text tokens. A new class of models, which TypeSafe markets as "System One" models, instead returns decisions and probability distributions across a user-supplied fixed answer set. A commercial model, JEV, is advertised as having a dramatic cost and speed advantage over traditional LLMs along with better calibrated decisions. As such, it might be useful for social scientists looking to quickly and cost-effectively annotate or scale large corpora of text and have a reliable indicator of a classifier's uncertainty. Yet, the accuracy of these claims and the broader model accuracy in social science text-based tasks are not yet established. In this paper, we do just that and hope to establish the suitability of JEV for social science tasks. We compare JEV with LLMs and human coders from published research, and with a current mid-tier commercial LLM (GPT-6 Luna) and an open-weight alternative (Qwen3.8-27B). We find that JEV matches, or comes close to, the capabilities of both LLMs in a variety of tasks. However, we find no cost advantage over GPT-6 Luna at OpenAI's batch prices. Further, we find that, when each question is asked once, JEV's probabilities are better calibrated than GPT-6 Luna's token probabilities, but not consistently better than Qwen3.8-27B's. We conclude that unless researchers have a need for speed, JEV's only obvious advantage is ease of parsing the underlying choice probabilities.

Steven Denney, Matthew DiGiuseppe

Leiden Institute for Area Studies, Leiden University. · Institute of Political Science, Leiden University.

LLM Evaluation Computational Social Science Language Model Calibration

Oct 5, 2026cs.AI

FREA: A Multi-Source Expert Benchmark for Reaction Feasibility Verification

As generative models and AI agents propose chemical reactions at a scale beyond expert review, feasibility verifiers decide which proposals enter synthesis planning. But do their decisions agree with chemists across different kinds of candidates? We introduce FREA, a benchmark of 751 reactions labeled by expert chemists under an explicit feasibility criterion, drawn from retrosynthesis model proposals, zero-yield experimental records, edits by large language models (LLMs), and five negative candidate generation methods. Our evaluation finds that no verifier leads across all sources: LLMs given only the criterion are competitive with dedicated verifiers, while forward models perform best on retrosynthesis proposals but reject most feasible edits of recorded reactions at the evaluated operating points. Looking beyond aggregate scores, both forward models perform below chance when separating infeasible alternative disconnections from feasible generated candidates. To study whether negative supervision addresses these weaknesses, we also release a corpus of over 14 million recorded reactions and generated negative candidates. In matched training comparisons, adding a mixture of generated negatives to forward training raises mean AUROC across sources, but these gains do not extend to retrosynthesis proposals. Varying the generation method further shows that the largest gain on generated candidates coincides with worse proposal screening. These findings motivate evaluating verifiers against experts across sources and designing negatives for transfer to model proposals.

Botao Yu, Bo Zhou, Daniel Adu-Ampratwum +7

Department of Computer Science and Engineering, The Ohio State University · Department of Pharmaceutical Sciences, University of Illinois Chicago · College of Pharmacy, The Ohio State University +3

LLM Evaluation AI-Assisted Scientific Research

Oct 5, 2026cs.LG

Will the Judge Flip? Predicting Position-Sensitive LLM Judgments from Residual Stream Activations

The order in which candidate responses are presented can change an LLM judge's verdict. Detecting such a position flip ordinarily requires judging each pair in both orders, which doubles the number of judgments. We investigate whether residual stream activations recorded immediately before the initial verdict can predict a flip. We use nested grouped cross-validation to evaluate regularized linear probes on 534 JudgeBench pairs for three Qwen3 judges and Llama-3.1-8B. The linear probes achieve AUROCs of .621-.850 and outperform a combined baseline that uses verbalized confidence, verdict-label logits, response lengths, and the judge's initial choice by .062-.113 AUROC. Linear probes trained on JudgeBench and then frozen achieve AUROCs of .685-.853 on 1,802 MT-Bench comparisons without MT-Bench fitting or recalibration. These results show that pre-verdict activations support prediction of susceptibility to candidate order and outperform the non-activation predictors evaluated here.

Hashmath Shaik, Gnaneswar Villuri, Alex Doboli

Department of Electrical and Computer Engineering Stony Brook University Stony Brook, NY 11794, USA

LLM Evaluation LLM-as-a-Judge Bias in LLM-as-a-Judge Language Model Probing

Oct 5, 2026cs.CL

JudgeMoE: Distributional Aggregation for LLM-as-a-Judge

When an LLM judge scores an output, its score distribution retains uncertainty and disagreement information that is lost after scalar compression. We introduce JudgeMoE, a lightweight aggregator that assigns example-specific weights to cached judge score distributions and fuses them before computing a final score. A protocol study shows that score-range choice is unstable across judge--dataset settings and that soft scoring usually outperforms hard decoding. On the original 10-cell benchmark, JudgeMoE improves mean Spearman over uniform log pooling by

+0.079

. Applying the same configuration to six additional cells yields a

+0.0393

mean gain over the strongest local single judge across 16 cells, with positive differences in 12/16 cells and a one-sided Wilcoxon signed-rank

p=0.0091

. Validation-based analyses further show that the preferred aggregation method depends on the task and judge pool.

Yiqi Liu, Joseph James, Yang Wang +3

The University of Manchester · The University of Sheffield · University of Pittsburgh +1

LLM Evaluation LLM-as-a-Judge Ensemble Learning Language Model Ensembles

Oct 5, 2026cs.CL

Anatomy of LLM Sycophancy: What a Flip Rate Hides

A model under pushback can correct itself, capitulate, or hold, and one flip rate counts a correction and a capitulation alike. Using SycoLens, a modular replay protocol, we test how user pressure and evaluation settings shape measured flip rates. Each measurement is one stateless replay of an item, a committed answer, and one scripted user line in a fixed form. Every effect is read against a matched control with the line deleted. Pushback wording, committed text, answer format, boundary distance, and ground truth become factors of one instrument; earlier instruments vary one to three of them. Across eleven frontier models from three providers and about 760,000 controlled replays, which models look sycophantic depends on how the user pushes back. Lines that assert the opposite verdict and lines that challenge the answer without asserting one rank the models almost unrelatedly. Flip effects grow several-fold near a model's boundary, yet items answered identically in every screening draw still carry about half of the most-affected totals. On arithmetic tasks where the truth is known, one model re-derives and corrects itself under pressure while another abandons correct answers without written work. On the model tested, a planted derivation lowers release of the answer it argues for, true or wrong, where a bare stated value does not; the wrong answer is corrected much more often than the true one is abandoned. Under a yes/no readout the rankings come closer, entangled with a pressure-induced shift toward "no". One score per model therefore compares different behaviours across models and benchmarks. We condense these dependencies into a reporting profile; the instrument, records, and analyses will be released upon publication.

Haonan Huang

Princeton University

LLM Evaluation LLM Sycophancy

Oct 5, 2026cs.AI

GraphDecide: Benchmarking System One Models on Graph Tasks

Large language models (LLMs) are increasingly explored for graph understanding and decision-making, while System One models such as Jev select directly from supplied options. However, the capabilities of System One models on graph-related tasks remain unclear. We introduce GraphDecide, a model-independent benchmark that combines structural task profiles, matched graph-text input contrasts and heuristic-proposal controls to diagnose graph decision performance. We evaluate Jev and related choice-based models alongside language-model baselines, covering fourteen model-interface configurations. Jev's results illustrate the benchmark's central distinctions: accurate adjacency recognition does not guarantee broader structural correctness, joint graph-text input does not consistently improve prediction, and feasible construction does not establish high solution quality. Its task contracts, candidate interfaces and scoring rules support comparison across native selectors and language-model adapters. Code and aggregate results are available at https://github.com/VictorYXL/JevGraphBench.

Xianliang Yang, Yapu Zhang, Li Zhao

Microsoft · Beijing University of Technology

LLM Evaluation LLM Reasoning with Graphs

Oct 5, 2026cs.CL

Breaking Bureaucracy: Evaluating open-source LLMs for legal document review

In this paper, we evaluate open-source generative LLMs on legal Natural Language Inference (NLI). Legal inspectorial processes take place in specific domains and often deal with confidential data. This creates a need for working with local models that do not require labeled training data. We evaluate our models on the ContractNLI benchmark and two NLI4Wills datasets. We successfully reproduce the baseline for the task (Span NLI BERT) and we evaluate multiple open-source LLMs on the same task. We analyze the invalid rate of the models, and their stability across temperature settings and domains. Among the generative models, Gemma-4 26B performs the best, reaching an accuracy of 81.2%, even outperforming the supervised model on one metric. On accuracy, it is not possible to beat the supervised model with zero-shot approaches. Qwen-3.6 35B performs well on both ContractNLI and additional datasets in the legal wills domain. Our findings indicate that zero-shot, open-source, generative LLMs are a viable alternative for real-world legal NLI when no supervised data is available. Our code is available at https://github.com/fbaratov/contractnli-llms.

Farrukh Baratov, Niki van Stein, Suzan Verberne

Leiden University Leiden, the Netherlands

Zero-Shot Learning Legal NLP LLM Evaluation

Oct 5, 2026cs.AI

MS-Exam-Gen: Source-Grounded Benchmark Construction for Evaluating LLMs on Textual Multiple Sclerosis MRI Knowledge

Biomedical large language model (LLM) evaluation requires auditable assessment of narrow, evolving, source-grounded subspecialty knowledge. Multiple sclerosis MRI (MS-MRI) provides a high-stakes textual-knowledge test case because correct reasoning requires current diagnostic criteria, standardized acquisition and reporting knowledge, longitudinal monitoring concepts, lesion morphology, and recognition of difficult mimics. We present MS-Exam-Gen, a reproducible framework for constructing and auditing a text-based multiple-choice question (MCQ) benchmark for MS-MRI knowledge; it does not evaluate direct MRI image interpretation. MS-Exam-Gen targets source-grounded criteria, protocols, reporting, and differential diagnosis. The framework combines expert-source indexing, exam-oriented topic induction, evidence-grounded MCQ generation, automated quality audits, a same-family consistency screen, and empirical calibration. From a 66-source corpus indexed into 4,289 retrieval chunks, the pipeline produced a locked 3,058-item candidate benchmark spanning 16 topics and 53 subtopics. Evaluation across 12 primary LLM endpoints yielded 36,696 item-level predictions and separated performance over a 42.8-percentage-point accuracy range (89.7% to 46.9%). Across these endpoints, 25.5% of items were missed by at least four. Post-generation audits showed that refreshed construction reduced measurable answer cues, while option-order testing showed that absolute MCQ scores remain position-sensitive. Generated construction labels remain metadata rather than validated psychometric categories. Because expert adjudication and full option-order counterbalancing remain future work, MS-Exam-Gen is not a clinically certified examination. It should be interpreted as an automatically filtered, source-grounded candidate benchmark and reproducible audit workflow for item-level and topic-specific LLM evaluation.

Abdul Basit, Muhammad Abdullah Hanif, Muhammad Shafique

eBRAIN Lab, Division of Engineering, New York University Abu Dhabi (NYUAD), Abu Dhabi, UAE

LLM Evaluation Synthetic Benchmark Generation Clinical QA Evidence-Grounded Generation Neuroimaging

Oct 5, 2026cs.CL

Knowing the Rules, Applying the Rules: Evaluating Language Models on Traditional Chinese Bazi

Knowing domain rules does not guarantee applying them to a case. We study this distinction in traditional Chinese Bazi through 3,000 Chinese multiple-choice questions spanning 14 Theory and 11 Case categories. Six endpoint systems are evaluated, with primary results reported on a 2,492-item model-informed refinement. Theory accuracy exceeds Case accuracy for every system, and gaps of 16.60-29.56 percentage points remain when invalid responses are excluded. The contrast is more specific than a general case-reasoning deficit. Across six systems, Twelve Stages and Nayin reach mean accuracies of 89.10% and 88.62%, while Shensha Basics reaches 75.96%. Within Case, Luck Pillars averages 84.62%, but Career and Family Relations average only 36.98% and 38.19%. Overall rankings also conceal different category strengths. On the original 3,000 items, paired DeepSeek native/disabled comparisons associate native configurations with Theory gains of 6.53 and 12.20 points for Flash and Pro, respectively; Case changes are -3.67 and +1.27 points. These are provider-configuration associations, not isolated causal effects of reasoning. The results motivate task-specific evaluation of cultural-domain applications rather than reliance on aggregate knowledge scores. The benchmark measures agreement with a model-generated, model-verified answer key, not real-world predictive validity. Final-set results are post-selection descriptions, and incomplete provenance and expert validation constrain their interpretation.

Jiulin Li, Ping Huang

Beijing Liuyi Guanhua Technology Co., Ltd. · State Key Laboratory of General Artificial Intelligence, BIGAI

LLM Evaluation Cultural Knowledge in Language Models

Oct 4, 2026cs.CL

Dataset Signatures in Human-LLM Interactions and User Modeling

Human--LLM interaction datasets shape our understanding of AI use and provide a foundation for downstream research, including training and evaluation of user models. In recent years, a growing number of datasets have sought to capture a representative picture of human--LLM interactions. But how different are the pictures these datasets provide, and what do those differences mean for research built on them? We study these questions across seven conversation datasets, spanning in-the-wild chat logs and human preference data. We begin by revisiting the dataset classification experiment of Torralba & Efros and find that neural network classifiers identify the source of a conversation from user messages alone well above chance, indicating distinctive dataset signatures. This separability persists after matching datasets on the dimensions of human-designed taxonomies, implying subtle differences that these taxonomies do not capture. We then examine the implications for user modeling: how dataset signatures propagate to the outputs of user models trained on these datasets; how dataset choice influences evaluations of user model quality and subsequent evaluations of LLM assistants paired with these user models; and how dataset classifiers can guide data selection for training user models. While each dataset is meant to capture a slice of 'real-world' interactions, our findings reveal the extent to which these slices diverge, and the consequences of those differences for research built on these foundations.

Joseph Suh, Serina Chang

University of California, Berkeley

LLM Evaluation User Modeling

Oct 4, 2026cs.LG

What Does a Harness Repair? A Preregistered Study of Visibility, Baseline Adequacy and Evaluation Defects

Harness search keeps a change to the prompts, reasoning switches, token budgets or parsers around a frozen model if the change raises a score. Such a gain can come from answers the parser could not read before, a weak comparison, or a defect in the evaluation. We preregistered a study of where these gains come from, with three small models, three benchmarks, replication and test partitions, a GEPA search arm and six evaluation defects injected one at a time, and we report all 47 primary endpoints. Turning thinking off raised accuracy over a capped thinking setting in 5 of 9 model-benchmark cells, and in each the gain came mostly from questions where the capped setting gave no readable answer. The thinking-off setting was not meaningfully worse than a rescue configuration or four GEPA-selected harnesses in 11 of 13 comparisons, and lost to the rescue on GSM8K for two models. GEPA repaired its broken starting points, but none of its selected harnesses was more accurate than the thinking-off setting. A thinking budget in the serving engine, which also allows a longer answer, lowered truncation and raised the parse rate in 6 of 9 cells. In 6 of 15 evaluable defect-model pairs, replication through the same pipeline reproduced the defect's distortion instead of revealing it. On the LongevityBench multiple-choice tasks, only the longevity-tuned model beat the strongest constant-label baseline.

Bowen Xu, Boyu Chen

Department of Biology, Stanford University, Stanford, CA, USA · Institute of Health Informatics, University College London, London, UK

LLM Evaluation Benchmark Validity

Oct 4, 2026cs.LG

GNN-CB: A Graph Neural Network Competition Benchmark for Human and LLM Evaluation

Large language models (LLMs) have demonstrated strong performance on coding and reasoning benchmarks; however, their ability to solve graph-structured machine learning problems remains largely unexplored. In particular, no benchmark currently evaluates whether LLMs can autonomously solve end-to-end Graph Neural Network (GNN) coding tasks under realistic competition settings. To address this gap, this paper introduces GNN-CB, the first competition-based benchmark for evaluating both humans and LLMs on GNN coding tasks. GNN-CB consists of 18 curated competitions spanning node-, edge-, and graph-level prediction across diverse graph categories, domains, and difficulty tiers. All submissions are evaluated through a unified automated pipeline with hidden test sets and standardized scoring. Human participants solve tasks under controlled competition constraints, while LLMs are evaluated using a frozen zero-shot prompting protocol based on a plan-then-code paradigm with bounded execute-and-repair loops. The benchmark additionally supports both non-agent and autonomous agent-based evaluation within the same protocol. Under our evaluated protocol, LLMs rarely match Human Top performance and show less stable performance across competitions. No single model dominates: a few competitions are won by LLMs, yet humans still hold the top score on most tasks. We release GNN-CB as a living benchmark with automated evaluation infrastructure, dynamic leaderboards, and reproducible execution pipelines. Beyond benchmarking, GNN-CB provides a practice-oriented resource for studying GNN implementation across progressively diverse graph-learning tasks. The benchmark and evaluation framework are publicly available at https://basiralab.github.io/GNN-CB/.

Murad Hossen, Tasneem Selim, Gurur Gamgam +22

University of Houston, Texas,USA · BASIRA Lab, Department of Computing, Imperial College London, United Kingdom · Department of Mathematics and Computer Science, Faculty of Science, Alexandria University, Alexandria, Egypt +23

LLM Evaluation Graph Neural Networks Code Generation Benchmarks Agent Benchmarks

Oct 4, 2026cs.PL

VHDL-REPOBENCH: A Repository-Level Benchmark for Evaluating Large Language Models on VHDL Design Generation

Large Language Models (LLMs) are increasingly applied in hardware design automation, demonstrating strong potential in generating and understanding hardware description languages. However, most existing benchmarks focus on Verilog, with limited evaluation of VHDL, which remains widely used in industry and academia for FPGA and safety-critical systems. To address this gap, we introduce VHDL-REPOBENCH, a large-scale, cross-file, repository-level benchmark for assessing LLM capabilities on realistic VHDL design generation and analysis tasks. VHDL-REPOBENCH curates ~100 open-source VHDL repositories, encompassing ~2.5k VHDL files and ~500 testbenches, and provides structured problem statements, module stubs, and self-verifying testbenches. The benchmark enables comprehensive evaluation across syntax, semantic correctness, hierarchical reasoning, cross-file dependency resolution, and functional verification. We evaluate several state-of-the-art models, including GPT-4o, Llama-3-70B, Qwen2.5-72B, CodeLlama-70B, and multi-step reasoning approaches such as Reflexion and CoDes. Results reveal that while current LLMs achieve moderate line- and block-level accuracy, substantial challenges remain in multi-file reasoning, hierarchical design understanding, and specification-to-module generation. VHDL-REPOBENCH represents the first large-scale VHDL-focused benchmark and provides a valuable resource for the hardware design community to evaluate, compare, and advance LLM capabilities for practical VHDL development.

Prashanth Vijayaraghavan, Akul Malhotra, Ashutosh Jadhav +2

IBM Research San Jose, CA, USA

LLM Evaluation Code Generation Code Generation Benchmarks LLM Code Generation

Oct 4, 2026cs.CL

MemStrata: 95% and 90.91% Source-Aware Accuracy on LongMemEval-500 and LoCoMo-1540 with a Local Qwen 3.8 27B Q4_K_M Reader

An adequate conversational answer may differ from a short or incomplete benchmark reference. To measure adequacy against the recorded history we prefer source-aware grading, in which the judge checks the reference against the full source before assessing system-blinded answers; original reference-only grading is reported alongside. With a local Qwen 3.8 27B Q4_K_M reader and a 24,000-token evidence ceiling, MemStrata CL1 scores 475/500 (95.0%) on LongMemEval-S and 1,400/1,540 (90.91%) on LoCoMo categories 1-4 under source-aware GPT-5.5 adjudication, against 463/500 (92.6%) and 1,205/1,540 (78.25%) under reference-only grading of the same answers. It preserves a retrieval backbone and adds nonduplicated, dated, speaker-attributed source spans. A same-reader full-history control with about 4.7 times the evidence scores 464/500 reference-only and 470/500 (94.0%) source-aware; neither difference is decisive. Keyword-only selection at the same budget scores 425, and a matched-reader Letta arm 438. On LongMemEval-M, where the packet holds about 1.6% of each history, MemStrata CL1 scores 427/500, with losses concentrated in multi-session and temporal questions. On 300 BEAM-1M questions it outscores dense retrieval, 0.738 to 0.706 (Wilcoxon p = 0.011). A same-seed replay of unchanged requests changed 1.5-2.3% of labels. On identical packets GLM 5.3 flash is non-inferior within 3 points (462 versus 463); Muse Spark 1.3 did not show non-inferiority on 269 questions. None of four pre-registered interventions met all of its registered advancement or feasibility criteria. Signed read-side artifacts support inspection but do not regenerate the private retrieval pipeline. The superiority of source-aware grading to human adjudication is not established, and development exposure, automated-judge dependence and the absence of held-out data preclude an independent-replication or leaderboard claim.

Neeraj Yadav

Called It Inc.

LLM Evaluation Memory-Augmented Language Models RAG Evaluation

Oct 4, 2026cs.CL

When Verifiable Counts Depend on Wording: Auditing Wording Robustness in Instruction Following

Verifiable instruction-following benchmarks often express each constraint through one fixed template. We test whether scores remain stable when the operational requirement is unchanged but its wording varies. We introduce WISE, a matched evaluation suite and reporting protocol instantiated on exact word count, keyword inclusion exactly once, and an inclusive 8--12 word range. Across 100 matched tasks, up to thirteen models from seven providers, and repeated generations scored over the complete visible output, wording alone produces substantial compliance shifts. In an avoidance-family panel, five avoidance and exclusion forms fall below the positive baseline, while constructional controls also shift compliance substantially: in the nine-model control panel, compliance is 54.9% for the original positive form, 48.2% for a longer positive form, 36.7% when the target appears later, and 33.8% for AVOID1. A strict JSON-structure probe shows wording sensitivity beyond counting, with a different direction of effect. Effect sizes, failure directions, weakest forms, and model rankings vary across realizations. Under the most disruptive exclusion form, the top-ranked model changes and 24.1% of strictly ordered model pairs reverse. Human validation further shows that unanimous agreement on an exact-count interpretation can coexist with substantially different model behavior. WISE supplements conventional scores with mean and worst-form compliance, wording gaps, failure profiles, and ranking stability.

Qishi Zhan, Seoyeon Jang, Zihan Dong +3

Marquette University · University of California, San Diego · Georgia Tech +3

LLM Evaluation Instruction Following Constraint Satisfaction Language Model Robustness Generative AI Evaluation

Oct 4, 2026cs.CL

How Much Do LLM-as-a-Judge Design Choices Matter? A Systematic Comparison of Prompt Designs, Rating Scales, and Models

Researchers increasingly use Large Language Models as judges (LLM-as-a-judge) to evaluate model outputs. Yet there are no standards for how to design these judges. Typically, researchers choose the prompt, rating scale, and model intuitively. If these choices change the judge's verdicts, two studies can reach different conclusions about the same facts. To address this risk and to provide an empirical basis for judge designs, we evaluate 10 reasoning models across multiple designs on two tasks: a scalar rating of sentence sentiment and toxicity (over 500 items per category), as well as a binary accuracy classification of question-answer pairs (n=600). For the rating tasks, despite judges showing significant disagreements with the human ground truth, the practical size of differences is small enough to consider most judges reliable (mean absolute deviation of 0.11 points on a 1 - 7 scale); toxicity judges even outperform standard classifiers. Judges are also highly accurate on average (96.5%) for the accuracy classification task. However, design choices can produce shifts: changing the rating scale alone can shift measured bias by up to 0.93 points (rating task), and while accuracy levels are rarely impacted, design choices consistently impact judge leniency (classification task; leniency drop of 28.9 percentage points when using detailed prompts, and up to 56.1 percentage points when switching models). Counterintuitively, lower reasoning effort affects neither accuracy nor leniency. Across both tasks, model identity is the dominant source of variance. These findings suggest that while LLM judges are broadly trustworthy in aggregate, design choices can be meaningful sources of variance. Given the growing reliance on automated evaluation in LLM research, we intend this study as a methodological reference for designing more robust and replicable LLM-as-a-judge pipelines.

Laurène Vaugrante, Thilo Hagendorff

AI Safety Research Group, University of Stuttgart

LLM Evaluation LLM-as-a-Judge LLM Reliability

Oct 2, 2026cs.AI

Benchmarking candidate coverage and rejection policy transfer in typed decision models

Rejection policies must remain useful as candidate sets and tasks change. We compare Laya, Jev and Qwen2.5-7B-Instruct using public reference labels, testing Laya/Jev policy transfer at equal calibration budgets and all three models on artificial omission, natural retrieval misses and public out-of-scope queries. Source calibration often fails to preserve the target operating point. A Jev policy calibrated on DBpedia rejects 69.3% of covered Emotion test inputs, while an Emotion policy loses detection entirely. Retrieval exposes a different tradeoff: with ten intent candidates, Laya detects 99.0% of out-of-scope queries but rejects 48.8% of covered queries. Separating missing-answer sources reveals these costs alongside retrieval coverage. The benchmark provides shared inputs, explicit decision and failure categories, and reproducible scoring to assess rejection policies under the conditions in which they are reused. Code and benchmark artifacts are available at https://github.com/luckykevvv/Decision_Model_Benchmark.

Jiawen Lu, Tongtong Wu

Monash University

LLM Evaluation Selective Prediction OOD Detection

Oct 1, 2026cs.HC

Where LLMs Fail with Visualization DSLs

As LLMs take up the role of authoring charts using visualization domain-specific languages (DSLs), the human constraints that shaped those languages may no longer apply, as what is easy for a person is not necessarily easy for a model. To understand how LLMs might work better with DSLs, we explore where and how they fail with current DSL designs. We evaluate 10 JSON-style visualization DSLs with 41 tasks across 3 LLMs, then assess the generated specifications with JSON and rendering checks, and qualitative coding of failed cases. Analyzing how this specification generation process fails, we identify four recurring failure patterns, link each to specific DSL features, and discuss design considerations for future DSL designs.

Chang Han, Andrew McNutt, Katherine Isaacs

University of Utah

Data Visualization LLM Evaluation LLM Code Generation

Oct 1, 2026cs.LG

A Safe Prototype Is Not a Safety Direction: Reference Dependence and Prompt Confounds in Response-Safety Embeddings

Can response safety be scored by cosine similarity to the mean embedding of known-safe responses? A recent sleeper-agent detector proposes exactly this score, yet the raw positive-centroid rule is not identified: positive observations locate the safe class relative to an encoder origin, but do not determine which direction separates safe from unsafe responses. We audit the rule on two prompt-controlled, human-labeled corpora and one auxiliary jury-labeled source control, using four frozen encoders and prompt-grouped splits. On the human-labeled corpora the safe prototype reaches ROC-AUC 0.457-0.545, with two cells significantly below chance and one above, while an explicit safe-minus-unsafe reference reaches 0.588-0.738 on the same embeddings; on the jury control the prototype is inverted (0.358-0.405) and the reference reaches 0.754-0.793. At validation-calibrated 5% false-safe thresholds, the reference accepts more safe responses on PKU-SafeRLHF (0.153-0.263 versus 0.039-0.061 across encoders) and Aegis (0.189-0.291 versus 0.004-0.045), but not reliably on BeaverTails. A fully unlabeled held-out reference recovers part to most of the referenced ranking, much less when only 5% of the pool is unsafe, whereas 80-634 labeled unsafe responses recover most of it. Prompt-only ablations show that prompt-label composition can inflate uncontrolled evaluations. This is a bounded result about a raw positive centroid, not all one-class methods or safety-specialized guards. A class mean is a location, not necessarily a safety direction; a declared reference with enough unsafe mass identifies orientation.

Sahil Kadadekar

New York University

Prototype-Based Classification LLM Evaluation Language Model Safety Evaluation Text Embedding Evaluation Safety Filtering

Oct 1, 2026cs.CL

Which LLM to pick? Online Active Model Selection for Large Language Models

Large Language Models (LLMs) are increasingly applied to process streaming data, with practitioners relying on benchmarks to select the best model even though these signals only approximate real performance. While oracle annotations can provide reliable feedback, they are often costly and difficult to obtain at scale. To address this challenge, we propose ONLINE LLM PICKER, the first framework for active model selection for LLMs in online settings. Given an arbitrary stream of queries and a limited annotation budget, ONLINE LLM PICKER selects the most informative prompts for annotation to identify the best LLM among candidate models. Across multiple tasks including 10 datasets, for over 130 language models, we show that ONLINE LLM PICKER saves annotation cost by up to 71.67% while reliably identifying the best or near-best model for the stream. We also show that using the returned model for sequential generation on unannotated prompts across the stream reduces regret by up to a factor of 2.51x, indicating that ONLINE LLM PICKER can identify the best or near-best model well before processing all streaming prompts.

Alessandro Turrin, Patrik Okanovic, Torsten Hoefler +1

TU Delft · ETH Zurich

Model Selection LLM Evaluation Active Learning

Oct 1, 2026cs.AI

MCRI: A Four-Dimensional Framework for Analyzing and Evaluating Agent Skills

As agents evolve from single-tool systems into modular, composite architectures, skills are becoming an important mechanism for capability development and distribution. However, the academic community lacks a structured framework for systematically analyzing and evaluating skills. Drawing on information gain and behavioral constraint, we propose the four-dimensional MCRI Framework and operationalize it as MCRI-Eval, a large language model-based evaluation method. We evaluate MCRI-Eval using 63,812 public skills from the OpenClaw skill Hub, with 58,275 skill-conditioned model executions across BigCodeBench, BFCL-Fundamental, and Mind2Web. MCRI-Eval scores are positively associated with community popularity signals and achieve the highest downstream ranking agreement among the evaluated methods. MCRI-Eval also improves top-1 skill selection across all three benchmarks: compared with the strongest baseline on each benchmark, the skills selected by MCRI-Eval advance by 17.7, 22.8, and 19.6 percentile points in downstream performance rank on BigCodeBench, BFCL-Fundamental, and Mind2Web, respectively. These results indicate that MCRI-Eval provides a useful pre-execution signal for prioritizing promising skills before costly execution-based evaluation.

Zongrui Yang, Li Xintong, Runchen Xu +4

LLM Evaluation Agent Evaluation AI Agent Benchmarks Agentic AI

Oct 1, 2026cs.CL

When Does a Second Model Help? Cross-Model Review in LLM Verification

Large language models now generate code, documentation, and analyses, and are increasingly used to review such output. We ask when a second review by a different model helps. Building on the author's earlier preprints, which varied context, repetition, and role structure within one model, we test model independence in a controlled experiment: 30 artifacts with 150 planted errors, 10 review conditions, and 900 review sessions with three reviewer models from two developers. In this experiment, (1) a top-tier cross-model reviewer is not significantly different in F1 from same-model review in a fresh session (CCR), which does not establish equivalence; (2) the two find partly different errors (Jaccard 41.2%); and (3) at two review calls, one CCR plus one cross-model review matches more planted errors than two CCR reviews (56.7% vs. 42.7%; Holm-adjusted p=.006), but not significantly more than two reviews by the top-tier cross-model reviewer, so model difference and reviewer capability are not separated. A lightweight cross-model reviewer scores no higher than same-model review. Withholding requirements from the reviewer raises F1 for the two lower tiers but not the top tier, in untested point estimates whose pattern depends on how failed sessions are scored. Before analysis we audited all session records, excluding one baseline run of uncertain provenance and 14 failed calls; results with all sessions are also reported. A partial check on public detector outputs from another benchmark neither replicates nor contradicts the main comparison. Records, artifacts, and scripts are available from the author on request.

Tae-Eun Song

Daejeon Jungang Cheonggua Co., Ltd.

LLM Evaluation Language Model Error Detection Automated Code Review

Oct 1, 2026cs.CL

Generalization Is Stability, Not Accuracy: Multi-Axis Evaluation of LLMs

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.

Nagham Omar, Mahmoud Jabarin, Maya Rozenshtein +3

Technion – Israel Institute of Technology

Prompt Sensitivity LLM Evaluation Language Model Generation Evaluation Language Model Robustness

Oct 1, 2026cs.CL

What Wins a Vote? Formatting, Length, and Lexical Diversity in the French Compar:IA LLM Arena

LLM arenas turn pairwise human preferences into model rankings. Those preferences may reflect how an answer is presented as well as what it says. We take a stylometric approach to 137,293 decisive French-language votes from the July 2026 Compar:IA release; the primary formatting analysis includes 137,113 battles across 116 models, and the joint estimates use the 127,092 battles with all required measurements. For each battle, we reconstruct the response visible when the user voted. We then compare the raw ranking with rankings adjusted for formatting, length, readability, vocabulary variety, and sentence structure. Presentation is associated with winning, but length, bold text, and lists tend to occur together, making their individual contributions hard to separate. Across the measured features, two associations change least across specifications: bold usage (+11.0% win odds per standard deviation in the joint model) and moving-average type-token ratio (MATTR), a measure of vocabulary variety that is less sensitive to answer length (+16.8%). The bold association is substantially smaller in observed multi-turn conversations, whereas the MATTR association changes little; because users choose whether to continue, this difference is descriptive rather than causal. The full adjustment moves 36 of 116 models by at least ten ranks. Yet comparisons with external benchmarks do not show that adjusted rankings better measure capability. We therefore recommend publishing raw and adjusted rankings side by side as a transparent sensitivity analysis.

Simonas Zilinskas, Maayeesha Farzana, Christophe Benavent

Compar:IA · Université Paris Dauphine-PSL

Oct 1, 2026cs.CL

Evaluating the Robustness of Japanese LLMs to IME-Related and Typographical Errors

Large language models (LLMs) have achieved strong performance across various natural language processing tasks. However, their robustness to typographical errors remains underexplored, particularly in Japanese, where text input involves multiple writing systems and IME-based conversion. In this study, we evaluate the robustness of Japanese LLMs against realistic Japanese-specific typos. We introduce five typo categories: Character Transposition, Character Replacement, Homophone Conversion, Japanese IME Conversion, and Full-Width Conversion. These perturbations are applied to three Japanese benchmark datasets (JMMLU, JCommonsenseQA, and JamC-QA), and eleven Japanese and multilingual LLMs are evaluated. The results show that Character Transposition and Character Replacement typos consistently reduce accuracy across benchmarks, whereas IME Conversion, Full-Width Conversion, and Homophone Conversion have relatively limited impact. These findings reveal that current Japanese LLMs remain vulnerable to realistic Japanese typing errors, particularly those that substantially distort the original input, highlighting the importance of robustness evaluation in practical input environments.

Ryota Mibayashi, Hiroaki Ohshima

Kobe University, Hyogo, Japan · University of Hyogo, Hyogo, Japan

LLM Evaluation Language Modeling Language Model Robustness

Oct 1, 2026cs.LG

How Much Can Language Models Gain from Test-Time Computation?

How much can test-time computation improve a language model, and at what cost? Test-time scaling is widely proposed as a substitute for larger models, but existing comparisons mostly evaluate one domain at a time and rarely charge selection to the budget. We introduce SELF-POT, a benchmark and evaluation framework that measures the test-time potential of a model across competition mathematics, competitive programming, and agentic workflows. SELF-POT separates candidate coverage from final accuracy on static tasks, tracks correctness transitions under revision, and measures protocol completion alongside task success in agentic environments. Under a unified budget rule, it compares Direct inference with parallel sampling and self-revision under fixed multiples of the Direct budget, and charges every model call, including selection and critique, in dollars. This design supports two kinds of comparison: the gain a model obtains from additional inference, and a lower-cost model with additional inference against a stronger model. Across five low-cost reasoning models on 350 sealed tasks, with Claude Opus 5.5 Direct as the reference, the returns depend on the domain, the selection rule, and failure handling. When we replay the retained programming candidate pools, public-example selection raises correct submissions from 376 to 453 of 500 scheduled cells while saving 12-49% of logical API cost across models, and simply retaining an available candidate when judging fails recovers 61 submissions at unchanged cost. On identical mathematics pools, judging with fallback yields 186 correct submissions versus 182 for voting, while voting saves 12-21% of logical API cost. These controlled replays show how selection and failure handling change the gains realized from the same generated candidates, and they quantify the marginal value of a model judge.

Bangji Yang, Jingyuan Li, Jiajun Fan +8

University of Illinois at Urbana-Champaign · University of Washington · Tsinghua University

LLM Evaluation Test-Time Scaling

Oct 1, 2026cs.AI

Evaluating LLM-Generated Preference Distributions

Large Language Models (LLMs) are increasingly used as probabilistic generators for simulation, synthetic data generation, and decision support in settings where real-world data are unavailable. Yet, the structure and reliability of the distributions they produce remain understudied. Here, we systematically analyze LLM-generated distributions of preferences for air travel, restaurants, and consumer products. Encouragingly, all models considered in our analysis exhibit self-coherence, with the most probable outcomes stabilizing rapidly under repeated sampling. At the same time, we observe substantial discordance across both model families and scales, with little consensus even among their most probable outcomes. These patterns hold across nine open-weight models, three choice domains, and show robustness under temperature changes, greedy decoding, and perturbations of prompt and ordering. Our findings indicate that outcomes are influenced more by the choice of model than by the wording of the prompt, challenging the common assumption that sufficiently capable LLMs produce similar preference distributions when used as stand-ins for survey respondents.

Fan Huang, Minsuk Kim, C. Tyler Diggans +1

Indiana University Bloomington Bloomington, IN, USA · Massachusetts Institute of Technology Cambridge, MA, USA · Root Dynamix LLC Sewickley, PA, USA

LLM Evaluation Synthetic Data Generation LLM Reliability User Preference Modeling

Oct 1, 2026cs.CL

A Citation-Grounded Benchmark for Trustworthy Earnings Call Transcript Analysis with Large Language Models

Large language models (LLMs) have been increasingly used for financial document analysis, including earnings call transcripts (ECTs). Beyond generating standalone claims, users increasingly prefer grounded analyses that pair claims with verifiable citations from source documents to enable independent validation. However, evaluating such analytical claims typically requires extensive expert annotation, which is costly and difficult to scale, and real-world financial analysis commonly involves long context-question-answer triplets, further increasing task complexity. To address these challenges and benchmark the current landscape of grounded analysis by LLMs, we propose a numeric evidence evaluation method that enables groundedness assessment without reliance on expert annotation. We also introduce an automated dataset construction pipeline and construct ECTs-100 from the top 100 constituents of the S&P 500 to support benchmark of both groundedness and correctness. In addition, we examine conscious incompetence, a practical failure mode in financial analysis in which LLMs must detect when available evidence is insufficient and refrain from producing unsupported hallucinations. Empirical results show that LLMs perform well in groundedness but face notable limitations in correctness, with informational insufficiency presenting an additional challenge.

Yingzhu Zhao, Vlad Pandelea, Han Yuan +4

Global Decision Science, American Express · Singapore Decision Science Center of Excellence, American Express, 1 Marina Boulevard, 018989, Singapore.

LLM Evaluation Financial QA Evidence-Grounded Generation LLM Abstention

Sep 30, 2026cs.CL

VERITYGATE: A Four-Gate Schema-Level Faithfulness Framework and Paired Benchmark for Grounded LLM Narrations over Structured Evidence

Fluent LLM explanations may not follow the evidence from a structured system. We present VERITYGATE, a four-gate checker for declared evidence IDs, entities, numbers, and claim types. It checks a fixed schema; it does not verify every fact in the prose. At r=0 and r=1, we test 900 instances per setting (450 grounded-ungrounded pairs) with GPT-4o-mini, Llama-3.3-70B, and Claude Sonnet 4.6. Under this schema-level contract and before repair, 80.3% of mini claims and 47.9% of Sonnet claims fail. These are verifier rejection rates, not prose-hallucination rates. One repair pass raises claim survival from 19.7% to 28.0% for mini and from 52.1% to 54.3% for Sonnet. Verified claims per example change by +0.14 for mini, -0.71 for Llama, and -0.47 for Sonnet, so survival and output volume must be reported together. A second Sonnet pass gives no clear gain. At r=1, Gate 4 covers 97.0%, 98.7%, and 100% of failing claims for mini, Llama, and Sonnet. Small human studies support the rules but show gaps between schema checks and correct prose. A domain-specific GPT-4o judge test shows an order effect, so it is only a usefulness check. We release the code and data.

Sachin Gupta

Independent Researcher San Jose, California, USA

LLM Evaluation LLM Grounding Factual Consistency Evaluation Evidence-Grounded Generation

Sep 30, 2026cs.CL

How Divergence Becomes Decision Flips in Compressed Language Models

Compression reports summarize how far a compressed language model moved from the dense one, usually by a KL divergence; a deployment that relies on the dense model's outputs needs to know how many of its decisions changed. We show that total variation, not KL, answers this directly. Across 802 compressed and perturbed copies of 19 open models on five corpora and nine mechanically unrelated perturbation families, the rate at which the arg-max token changes (the \emph{flip rate}) tracks total variation at a ratio with median

1.05

, with no fitted constant. KL converts into flips only through its square root and a factor that varies fourfold across models and corpora, because KL averages over tokens before the root is taken; first-order statistics averaged per token, such as Hellinger distance, avoid this, but reports rarely give them. As a result, of two compressors reported on different models and corpora whose flip rates differ by at least

10%

, KL assigns the smaller divergence to the one that changes more decisions in

11%

of cases, total variation in

1%

. Two pre-registered tests mark the limits: on a held-out code corpus the ratio held for all eight models while three predictions about KL each failed for half of them or more, and on three new models with real kernels it stayed in its band for 37 of 38 checkpoints but fell below one on code for two models. In vLLM speculative decoding, total variation measured under teacher forcing predicts greedy draft acceptance with a mean relative error of

1.1

--

2.4%

, without the task-specific calibration that KL needs.

Beatriz Almeida Felicio

Instituto de Informática (INF) Universidade Federal de Goiás (UFG)

LLM Evaluation LLM Compression Distribution Matching