LLM-as-a-Judge

Large Language Models (LLMs) augmented with Retrieval-Augmented Generation (RAG) techniques are revolutionizing applications across multiple domains, such as healthcare, finance, and customer service. Despite their potential, evaluating RAG systems remains a complex challenge due to the stochastic nature of generated outputs and the intricate interplay between retrieval and generation components. This paper introduces Deepchecks, a comprehensive framework tailored for evaluating RAG applications. Deepchecks' evaluation framework addresses RAG applications evaluation through a multi-faceted approach, root cause analysis and production monitoring. By ensuring alignment with application-specific requirements, Deepchecks framework provides a robust foundation for assessing reliability, relevance, and user satisfaction in RAG systems.

As generative AI models such as large language models (LLMs) become more pervasive, ensuring the safety, robustness, and overall trustworthiness of these systems is paramount. However, AI is currently facing a reproducibility crisis driven by unreliable evaluations and unrepeatable experimental results. While human raters are often used to assess models for utility and safety, they introduce divergent biases and subjective opinions into their annotations. Overcoming this variance is exceptionally challenging because very little data exists to study how experimental repeatability actually improves as the annotator pool grows. Standard evaluation practices typically rely on a small number of annotations per item (often 3 to 5) and lack the persistent rater identifiers necessary to model individual variance across items. In this work, we introduce a multi-level bootstrapping approach to realistically model annotator behavior. Leveraging datasets with a large number of ratings and persistent rater identifiers, we analyze the tradeoffs between the number of items ($N$) and the number of responses per item ($K$) required to achieve statistical significance.

LLM-as-a-judge is now the default measurement instrument for open-ended generation, but on the public JudgeBench benchmark even strong instruction-tuned judges barely scrape past random on objective-correctness pairwise items. We introduce RTLC, a three-stage prompting recipe -- Research, Teach-to-Learn, Critique -- that promotes a single black-box LLM into an ensemble-of-thought judge with no fine-tuning, retrieval, or external tools. Stage 1 wraps the input in a fixed pedagogical scaffold porting the Feynman Learning Technique (study $\to$ teach $\to$ find gaps $\to$ simplify) into LLM prompting. Stage 2 draws N=10 independent candidate verdicts at temperature 0.4. Stage 3 acts as its own critic, cross-comparing the candidate set against the original question to emit one critiqued verdict at temperature 0. On JudgeBench-GPT (350 hard pairwise items), Claude 3.7 Sonnet's pairwise accuracy climbs from 64.6% (single-shot vanilla prompt) to 78.6% (RTLC critique-of-10) -- an absolute 14.0-percentage-point gain. RTLC also beats N=10 self-consistency majority voting (77.7%) and a zero-shot first candidate (74.0%). A clean three-step ablation attributes +9.4 pp to the Teach-to-Learn scaffold, +3.7 pp to N=10 marginalisation, and +0.9 pp to explicit critique. We discuss the cost-accuracy frontier (RTLC sits above self-consistency at every working point), the error-budget breakdown across the four JudgeBench categories (knowledge, reasoning, math, coding), and how RTLC composes orthogonally with post-hoc judge-score calibration, with the two interventions compounding multiplicatively in practice.

This article investigates the performance of automatic evaluation metrics (AEMs) and LLM-as-a-judge evaluation on literary translation across multiple languages, genres, and translation modalities. The aim is to assess how well these tools align with professionals when evaluating translation, creativity (creative shifts & errors), and see if they can substitute laborious manual annotations. A dataset of literary translations across three modalities (human translation, machine translation, and post-editing), three genres and three language pairs was created and annotated in detail for creativity by experienced professional literary translators. The results show that both AEMs and LLM-as-a-judge evaluations correlate poorly with professional evaluations on creativity, with LLM-as-a-judge showing a systematic bias in favour of machine-translated texts and penalising creative and culturally appropriate solutions. Moreover, performance is consistently worse for more literary genres such as poetry. This highlights fundamental limitations of current automatic evaluation tools for literary translation and the need to create new tools that do not frequently consider out of routine translations as errors.

We ask whether agentic AI systems built for software engineering transfer to realistic hardware engineering. Existing hardware LLM benchmarks isolate sub-tasks but none jointly requires repository navigation, hierarchy-aware localization, Electronic Design Automation (EDA) executable verification, and maintenance-style patching. We introduce \textbf{Phoenix-bench}, a synchronized corpus of 511 verified Verilator instances from 114 GitHub repositories, each shipped with the developer patch, design-flow labels, fail-to-pass and pass-to-pass testbenches, and a Docker-pinned EDA environment so resolved-rate differences reflect agent behavior rather than toolchain availability. Using Phoenix-bench we run a uniform evaluation of four commercial agents and eight open-source agentic structures across four LLM backbones, plus two diagnostic interventions (file-level oracle localization and one round of testbench-log feedback). Three findings emerge. (i)~Software and hardware are fundamentally different engineering tasks: the same agent loses 37% to 58% from SWE-bench Verified to Phoenix-bench because hardware bugs propagate across parallel instantiated modules through signal flow rather than along a software-style call graph, and software-tuned agents stop at the symptom file instead of tracing back through the instantiation chain. (ii)~Failures concentrate on design control-flow / finite state machine (FSM) bugs, verification testbench bugs, and hard cases that demand cross-hierarchy signal-flow tracking and coordinated multi-file edits. (iii)~Localization granularity matters far more than localization itself: a perfect file-level oracle yields only $+1.4$% because the agent then breaks files that did not need editing, while a single round of test case feedback lifts resolved rate by $42$% to $45$% because the test case tells \emph{where} the bug is and \emph{what} the fix has to look like.

Calibration is commonly evaluated by comparing model confidence with its empirical correctness, implicitly treating reliability as a function of the confidence score alone. However, this view can hide substantial structure: models may be systematically overconfident on some kinds of inputs and underconfident on others, causing global reliability diagnostics to obscure localised calibration failures. To address this, we formulate the problem of discovering hidden miscalibration regimes without assuming access to predefined data slices. We define the corresponding miscalibration field and propose a diagnostic framework for estimating it. Our approach learns a calibration-aware representation of the input space and estimates signed local miscalibration by kernel smoothing in the learned geometry. Across four real-world LLM benchmarks and twelve LLMs, we find that input-dependent calibration heterogeneity is prevalent. We further show that the discovered fields are actionable: they support local confidence correction and reduce calibration error in systematically miscalibrated regions where confidence-based methods such as isotonic regression and temperature scaling are less effective.

Checkpoint selection for multimodal large language models (MLLMs) presents significant challenges when performance differentials are marginal and evaluation signals are prone to noise. Existing methodologies rely heavily on static benchmarks or pointwise scoring, which frequently misalign with in-the-wild usage and lack robust uncertainty estimation, particularly in OCR-heavy scenarios. In this work, we formulate checkpoint selection as a robust decision problem under evaluation uncertainty. We propose a multi-stage framework that integrates curated real-world data, structured LLM-based judgment, and multi-stage ranking protocols. The evaluation system orchestrates progressive refinement via pointwise filtering, listwise ranking, and pairwise comparison. To enhance reliability, we introduce subsampling-based confidence estimation and a percentile-based scoring formulation that captures distributional characteristics while penalizing tail failures. Furthermore, we demonstrate that data quality, specifically OCR readability, is a critical determinant of evaluation validity.

Chatbot usage has increased, including in fields for which they were never developed for--notably mental health support. To that end, we introduce Validations of Ethical and Responsible AI in Mental Health (VERA-MH), a novel clinically-validated evaluation for safety of chatbots in the context of mental health support. The first iteration of VERA-MH focuses on Suicidal Ideation (SI) risks, by assessing how well chatbots can responds to users that might be in crisis. VERA-MH is comprised of three steps: conversation simulation, conversation judging and model rating. First, to simulate conversations with the chatbot under evaluation, another chatbot is tasked with role-playing users based on specific personas. Such user personas have been developed under clinical guidance, to make sure that, among others, multiple risk factors, demographic characteristics and disclosure factors were represented. In the judging step, a second support model is used as an LLM-as-a-Judge, together with a clinically-developed rubric. The rubric is structured as a flow, with a single Yes/No question asked each time, to improve answers' consistency and highlight models' failure modes. In the last stage, results of each conversation are aggregated to present the final evaluation of the chatbot. Together with the framework, we present the result of the evaluations for four leading LLM providers.

Large language models (LLMs) can often generate functionally correct code, but their ability to produce efficient implementations for performance-critical systems tasks remains limited. Existing code benchmarks mainly emphasize correctness or algorithmic problem solving, while realistic systems-level optimization is still underexplored. To address this gap, we introduce PerfCodeBench, an executable benchmark for evaluating LLMs on high-performance code optimization. The tasks require system-level implementation choices, hardware-aware optimization, and careful handling of performance bottlenecks. Each task includes executable correctness checks, a baseline implementation, and a reference optimized solution. This allows us to evaluate both correctness and runtime-oriented efficiency. Our evaluation on a broad set of state-of-the-art LLMs shows a clear gap between model-generated code and expert-optimized implementations. The gap is especially large on tasks involving parallelism and GPU operations. Current models also show weaknesses in cross-language robustness and in consistently reaching expert-level efficiency. These results suggest that performance-aware evaluation are still needed. LLMs should move beyond generating merely correct code toward producing efficient systems software. We submit the benchmark data, evaluation infrastructure, and complete logs of all LLMs-generated code at https://anonymous.4open.science/r/perfcodebench-7CDE.

Large language models (LLMs) are increasingly used as automatic judges for summarization and dialogue evaluation. Prior work has documented biases such as position, verbosity, and style preferences, but largely focuses on outcomes, leaving judge explanations underexplored. We instead ask whether LLM judges are cue-invariant, i.e., whether their rankings and explanations remain stable when non-evidential cues are perturbed while holding the underlying texts fixed. We introduce a suite of cue interventions (Blind, Truth, Flip, Placebo, Reveal-After) and tie-aware metrics that quantify outcome anchoring and rationale anchoring, including label-aligned rhetoric and explanation drift, alongside consistency and stereotype-intrusion checks. We design anchoring attacks using verbosity and confidence cues, and compare two mitigations: structured chain-of-thought prompting and PROOF-BEFORE-PREFERENCE (evidence lock, score, rank). Using a new dataset of 1,000 summaries from traditional extractive models and LLMs, we find substantial cue-anchored rationalization under label and placebo perturbations, while PROOF-BEFORE-PREFERENCE markedly improves cue invariance over baselines.

Wu et al. (2026) showed that most frontier large language models (LLMs) recommend a sponsored, roughly twice-as-expensive flight when their system prompt contains a soft sponsorship cue. We reproduce their evaluation on ten open-weight chat models plus the two of their twenty-three models that are still reachable today (gpt-3.5-turbo, gpt-4o). All reported rates in this paper are produced under the same judge the original paper used (gpt-4o); we additionally store every label under an open-weight (gpt-oss-120b) and a smaller proprietary (gpt-4o-mini) judge for an ablation. Three findings emerge. First, a prose description of an LLM evaluation pipeline is not, on its own, sufficient for accurate reproduction: we surfaced three silent implementation failures that each shifted a reported rate by tens of percentage points. Second, the central claims do generalise - the gpt-3.5-turbo logistic-regression intercept of alpha = 0.81 is within four points of the original alpha = 0.86, and 200 of 200 trials on gpt-3.5-turbo and gpt-4o promote a payday lender to a financially distressed user. Third, a thirty-token user prompt that asks the assistant for a neutral comparison table first cuts sponsored recommendation from 46.9% to 1.0% averaged across our ten open-source models, and from 53.0% to 0% averaged across the two OpenAI models. AI literacy and price-comparison portals are likely market-level mitigations; the harmful-product cell is bounded by neither. Raw data, labels and analysis scripts are at https://github.com/akmaier/Paper-LLM-Ads .

Multimodal large language models (MLLMs) are now routinely deployed for visual understanding, generation, and curation. A substantial fraction of these applications require an explicit aesthetic judgment. Most existing solutions reduce this judgment to predicting a scalar score for a single image. We first ask whether such scores faithfully capture comparative preference: in a controlled study with eight expert annotators, score-derived rankings align poorly with the same annotators' direct comparisons, while direct ranking yields substantially higher inter-annotator agreement on best- and worst-image labels. Motivated by this finding, we introduce the Visual Aesthetic Benchmark (VAB), which casts aesthetic evaluation as comparative selection over candidate sets with matched subject matter. VAB contains 400 tasks and 1,195 images across fine art, photography, and illustration, with labels derived from the consensus of 10 independent expert judges per task. Evaluating 20 frontier MLLMs and six dedicated visual-quality reward models, we find that the strongest system identifies both the best and the worst image correctly across three random permutations of the candidate order in only 26.5% of tasks, far below the 68.9% achieved by human experts. Fine-tuning a 35B-parameter model on 2,000 expert examples brings its accuracy close to that of a 397B-parameter open-weight model, suggesting that the comparative signal in VAB is transferable. Together, these results expose a clear and measurable gap between current multimodal models and expert aesthetic judgment, and VAB provides the first set-based, expert-grounded testbed on which that gap can be tracked and closed.

Automatic generation of educational materials using large language models (LLMs) is becoming increasingly common, but assigning difficulty levels to such materials still requires substantial human effort. LLM-as-a-Judge has therefore attracted attention, yet disagreement with human raters remains a major challenge. We propose a method for predicting which LLM-generated difficulty ratings are likely to disagree with human raters, so that such cases can be sent for re-rating. Unlike prior approaches, our method does not rely on generation-time probability signals, which must be collected during rating generation and are often difficult to compare across LLMs. Instead, exploiting the fact that difficulty is an ordinal scale, we use a separate embedding space, such as ModernBERT, and identify disagreement candidates based on the geometric consistency of the rating set. Experiments on English CEFR-based sentence difficulty assessment with GPT-OSS-120B and Qwen3-235B-A22B showed that the proposed method achieved higher AUC for predicting disagreement with human raters than probability-based baselines.

Prompt specifications for multi-agent large language model (LLM) systems carry data contracts and integration logic across interdependent files but are rarely subjected to structured-inspection rigor. We report a single-system case study of iterative, agent-driven auditing applied to AEGIS (Autonomous Engineering Governance and Intelligence System), a seven-lane production pipeline whose 7152-line specification surface was audited across nine rounds, surfacing 51 consistency defects (per-round counts of 15, 8, 12, 2, 8, 1, 4, 1, 0). We present a seven-category post hoc taxonomy with explicit coding rules, non-monotonic convergence consistent with cascading edits and audit-scope expansion, and a locked audit protocol. We further report two partial replications on a public synthetic mini-specification: a cross-LLM panel of four frontier vendors (OpenAI, Anthropic, Google, xAI; 12 traces; multi-vendor union detects all five seeded defects) and an inter-rater reliability check on a stratified subsample (Cohen's $κ$ = 0.80 on category, 0.46 on severity). The full reproducibility bundle accompanies the submission.

Large Language Models (LLMs) and Vision Language Models (VLMs) have recently shown promising capabilities in various scientific domain. In particular, these advances have opened new opportunities in drug discovery, where the ability to understand and modify molecular structures is critical for optimizing drug properties such as efficacy and toxicity. However, existing models and benchmarks often overlook toxicity-related challenges, focusing primarily on general property optimization without adequately addressing safety concerns. In addition, even existing toxicity repair benchmarks suffer from limited data diversity, low structural validity of generated molecules, and heavy reliance on proxy models for toxicity assessment. To address these limitations, we propose MolDeTox, a novel benchmark for molecular detoxification, designed to enable fine-grained and reliable evaluation of toxicity-aware molecular optimization across stepwise tasks. We evaluate a wide range of general-purpose LLMs and VLMs under diverse settings, and demonstrate that understanding and generating molecules at the fragment-level improves structural validity and enhances the quality of generated molecules. Moreover, through detailed task-level performance analysis, MolDeTox provides an interpretable benchmark that enables a deeper understanding of the detoxification process. Our dataset is available at : https://huggingface.co/datasets/MolDeTox/MolDeTox

Large Language Models (LLMs) achieve strong performance on standard knowledge evaluation benchmarks, yet recent work shows that their knowledge capabilities remain brittle under question variants that test the same knowledge in different forms. Robustness augmentation of existing knowledge evaluation benchmarks is therefore necessary, but current LLM-assisted generate-then-verify pipelines are costly and difficult to scale due to low-yield variant generation and unreliable variant verification. We propose SAGE (Scalable Automated Generation of Robustness BEnchmarks), a framework for scalable robustness augmentation of knowledge evaluation benchmarks using fine-tuned smaller models. SAGE consists of VariantQual, a rubric-based verifier trained on human-labeled seed data, and VariantGen, a variant generator initialized with supervised fine-tuning and further optimized with reinforcement learning using VariantQual as the reward model. Experiments on HellaSwag show that SAGE constructs a large-scale robustness-augmented benchmark with quality comparable to the human-annotated HellaSwag-Pro at substantially lower cost, while the fine-tuned models further generalize to MMLU without benchmark-specific fine-tuning.

LLM-as-Judge systems are widely deployed for automated evaluation, yet practitioners lack reliable methods to know when a judge's verdict should be trusted. Token log-probabilities, the standard post-hoc confidence signal, are unavailable for many commercial LLMs and, even when accessible, saturate above 0.999 with structured JSON output. We introduce VERDI (VERification-Decomposed Inference), a method that extracts confidence from the reasoning trace a structured judge already produces, with no additional inference calls. VERDI decomposes each verification-style evaluation into sub-checks and derives three structural signals: Step-Verdict Alignment, Claim-Level Margin, and Evidence Grounding Score. We combine them with Platt-scaled logistic regression. On three public benchmarks, VERDI achieves AUROC 0.72-0.91 on GPT-4.1-mini and 0.66-0.80 on GPT-5.4-mini. On Qwen3.5-4B/9B/27B, where answer-token logprobs are anti-calibrated (higher confidence on errors, AUROC 0.32-0.49), VERDI achieves 0.56-0.70. We additionally validate on a production system with eight rubrics (AUROC 0.73-0.88 on factual rubrics), demonstrate cross-model transfer (AUROC 0.66-0.69), and show that a 33M-parameter NLI (Natural Language Inference) model provides a scalable alternative to regex extraction.

While existing benchmarks demonstrate the near-perfect performance of large language models (LLMs) on various tasks, this apparent saturation often obscures the need for rigorous evaluation of their reliability. In real-world deployment, however, achieving extremely high reliability (e.g., "five-nines" (99.999%) vs. "three-nines" (99.9%)) is fundamentally critical, as this gap results in an order-of-magnitude increase in failures, which is catastrophic in reliability-critical applications. Still, estimating such a rare failure probability with tight confidence bounds requires prohibitively large LLM inference sizes, making standard Monte Carlo evaluation infeasible under limited compute budgets. In this paper, we observe that LLM failures exhibit strong systematic patterns: across broad parameterized input spaces, a small subset of inputs disproportionately accounts for the majority of failures. Leveraging this observation, we propose to learn a sampling distribution concentrated on failure-prone inputs via the cross-entropy method (CEM). We evaluate our framework on three LLMs, Qwen2.5-Math-7B-Instruct, gpt-oss-20b-low, and Gemini 2.5 Flash Lite, across parameterized GSM8K templates and achieve up to 156.22x reduction in required inferences compared to naive uniform sampling. Our estimates reveal that models with indistinguishable accuracy on standard benchmarks can differ substantially in estimated failure rates, underscoring that reliability is a distinct and measurable axis of model quality. Our simple yet practical framework enables the evaluation of extreme reliability in LLMs, a distinct and underexplored dimension of evaluation beyond existing benchmarks, for their growing use in reliability-sensitive applications.

Reasoning-capable large language models (LLMs) have recently been adopted as automated judges, but their benefits and costs in LLM-as-a-Judge settings remain unclear. Through controlled comparisons between reasoning and non-reasoning judges, we show that explicit reasoning substantially improves judgment accuracy on tasks requiring structured verification (e.g., math and coding), while offering limited or even negative gains on simpler evaluations and incurring significantly higher computational cost. These findings motivate that reasoning should be used selectively rather than universally, with awareness of possible distribution shift. We propose a Robust Adaptive Cost-Efficient Routing (RACER), which dynamically selects between reasoning and non-reasoning judges under a fixed budget by formulating routing as a constrained distributionally robust optimization problem. RACER explicitly accounts for distribution shift via a KL-divergence uncertainty set, admits an efficient primal--dual algorithm, and enjoys theoretical guarantees including uniqueness of the optimal policy and linear convergence. Extensive experiments show that RACER achieves superior accuracy--cost trade-offs under distribution shift.

Multimodal large language models (LLMs) are increasingly explored as automated evaluators in clinical settings, yet their scoring behavior on ordinal clinical scales remains poorly understood. We benchmark three frontier LLM families against supervised deep learning models for scoring Clock Drawing Test (CDT) images on two public datasets using the Shulman rubric. While fully fine-tuned Vision Transformers achieve the best calibration (MAE 0.52, within-1 accuracy 91%), zero-shot LLMs remain competitive on tolerance-based agreement (GPT-5 MAE 0.67, within-1 accuracy 92%) despite higher absolute error. However, per-score analysis reveals that all three LLM families exhibit a pronounced central tendency effect (systematic endpoint compression): predictions are systematically compressed toward the middle of the scale, with over-prediction at the low end (score 0 to 1) and under-prediction at the high end (score 5 to 4). This effect disproportionately affects the clinically critical extremes where accurate scoring most impacts screening decisions for cognitive impairment. Targeted ablations show that neither few-shot exemplars spanning the full score range nor removing clinical terminology from the prompt eliminates the effect. Our findings extend the LLM-as-a-judge bias literature from NLP evaluation to clinical assessment, and highlight the need for calibration-aware evaluation and post-hoc calibration before deploying LLM-based raters in high-stakes screening workflows.

The rapid adoption of LLMs in both research and industry highlights the challenges of deploying them safely and reveals a gap in the systematic evaluation of toxicity benchmarks. As organizations increasingly rely on these benchmarks to certify models for customer-facing applications and automated moderation, unrecognized evaluation biases could lead to the deployment of vulnerable or unsafe systems. This work investigates the robustness of established benchmarking setups and examines how to measure currently neglected intrinsic biases, such as those related to model choice, metrics, and task types. Our experiments uncover significant discrepancies in benchmark behaviors when evaluation setups are altered. Specifically, shifting the task from text completion to summarization increases the tendency of benchmarks to flag content as harmful. Additionally, certain benchmarks fail to maintain consistent behavior when the input data domain is changed. Furthermore, we observe model-specific instabilities, demonstrating a clear need for more robust and comprehensive safety evaluation frameworks.

Selecting the best large language model (LLM) for a fixed benchmark is often expensive, since exhaustive evaluation requires running every model on every example. Multi-armed bandit (MAB) algorithms can reduce the number of LLM calls by sequentially selecting the next model-example pair to evaluate, thereby avoiding wasted evaluations on clearly underperforming models. Further savings can be achieved by predicting model scores from the partially observed model-example score matrix using low-rank factorization. However, such predictions are not ground truth: they can be biased and may therefore lead to incorrect identification of the best model. In this work, we propose a principled framework that combines MAB with cheap predicted scores without compromising statistical validity. Specifically, we derive doubly robust estimators of each model's performance that use the low-rank predictions to reduce variance. This enables the construction of valid finite-sample confidence intervals in our setting, where models are selected adaptively and examples are sampled without replacement. Empirical results on real-world benchmarks show that our approach reduces the number of required evaluations, yielding meaningful savings in compute and cost while accurately identifying the best-performing model.

In industrial procurement, an LLM answer is useful only if it survives a standards check: recommended material must match operating condition, every parameter must respect a regulated threshold, and no procedure may contradict a safety clause. Partial correctness can mask safety-critical contradictions that aggregate LLM benchmarks rarely capture. We introduce IndustryBench, a 2,049-item benchmark for industrial procurement QA in Chinese, grounded in Chinese national standards (GB/T) and structured industrial product records, organized by seven capability dimensions, ten industry categories, and panel-derived difficulty tiers, with item-aligned English, Russian, and Vietnamese renderings. Our construction pipeline rejects 70.3% of LLM-generated candidates at a search-based external-verification stage, calibrating how unreliable industrial QA remains after LLM-only filtering. Our evaluation decouples raw correctness, scored by a Qwen3-Max judge validated at $κ_w = 0.798$ against a domain expert, from a separate safety-violation (SV) check against source texts. Across 17 models in Chinese and an 8-model intersection over four languages, we find: (i) the best system reaches only 2.083 on the 0--3 rubric, leaving substantial headroom; (ii) Standards & Terminology is the most persistent capability weakness and survives item-aligned translation; (iii) extended reasoning lowers safety-adjusted scores for 12 of 13 models, primarily by introducing unsupported safety-critical details into longer final answers; and (iv) safety-violation rates reshuffle the leaderboard -- GPT-5.4 climbs from rank 6 to rank 3 after SV adjustment, while Kimi-k2.5-1T-A32B drops seven positions. Industrial LLM evaluation therefore requires source-grounded, safety-aware diagnosis rather than aggregate accuracy. We release IndustryBench with all prompts, scoring scripts, and dataset documentation.

Multi-judge evaluation is increasingly used to assess LLMs and reward models, and the prevailing heuristic is to curate: keep the most accurate judges and discard weaker ones. We show that this heuristic can reverse when the target is not point accuracy, but calibrated probabilistic evaluation from a labeled calibration set. Holding the aggregation and calibration procedures fixed, we compare accuracy-ranked top-$k$ judge selection with using the full judge panel. Across four labeled pairwise-evaluation benchmarks spanning LLM-as-judge and reward-model settings, the calibrated full panel consistently outperforms accuracy-based selection. On RewardBench2, retaining all judges achieves negative log-likelihood (NLL) of $0.006$ versus $0.013$ under top-5 selection, halving the calibration error. This advantage persists after judge-family deduplication and against stronger same-pipeline subset search. We explain this reversal with oracle analyses showing that the optimal calibrated risk under proper scoring rules cannot increase when additional judge signals are made available, and that even below-chance judges can be useful when their biases are learnable and their signals are non-redundant. The resulting operating principle is simple: in multi-judge evaluation with labeled calibration data, do not discard weak judges by accuracy alone; keep them when they are parseable, non-redundant, and calibratable.

Smart-contract exploits have caused billions of USD in cumulative losses, yet audits remain expensive and slow. Automated tools have emerged to close this gap, but each class has a characteristic failure mode. Static analyzers report findings that frequently fail manual triage at high rates, while large language models (LLMs) hallucinate findings that contradict the source code. Thus, we propose Chaintrix, an end-to-end auditing framework whose central architectural commitment is that every LLM-generated claim must be discharged against a deterministic structural contract representation. We introduce a Cross-Contract Interaction Model (CCIM) that parses Solidity into a structured map of function-level reads, writes, modifiers and resolved cross-contract calls. CCIM serves as the substrate against which all 12 of Chaintrix's deterministic signal engines and the parallel LLM audit pipelines operate. A staged false-positive-reduction pipeline, terminating in a Structural Verdict Engine (SVE) that applies deterministic structural checks against parsed code, filters the merged finding set, with selected high-confidence findings further validated through symbolic execution and fuzz testing. We evaluate Chaintrix on EVMbench, the smart-contract security benchmark by OpenAI, Paradigm, OtterSec. Chaintrix detects 86 of 120 high-severity vulnerabilities (71.7% recall), with 25 audits scoring 100% recall, placing Chaintrix 26 percentage points above the strongest frontier-model baseline.

[Abridged] Using a Large Language Model (LLM) as an automatic rater (LLM-as-a-judge) is cheap but potentially biased: some judges run lenient, others strict, the middle of the scale gets compressed, and verbose answers may be over-rewarded. A common remedy is post-hoc calibration: leave the cheap judge in place and, on a modest set of paired anchors, fit a transformation from raw judge scores to an estimate of the human rating. We compare two correctors that take opposing views on how this mapping should be modeled: a parametric, small-anchor hierarchical Bayesian linear correction with per-score uncertainty, and a non-parametric Neural-ODE (FFJORD) score-transport flow. Both are run head-to-head on UltraFeedback fine-grained_score (1700 paired examples, 200 held out), with calibration split into three operational sub-questions: population-mean recovery, per-item accuracy, and distributional-shape match. The headline result is that the choice between methods is primarily a data-budget question. Both correctors close the raw $+0.71$-point mean offset to within $\pm 0.08$ of the GPT-4 reference, at 100 and at 1500 anchors. Past that, the methods swap roles. With 100 anchors, the linear corrector reconstructs the human-score distribution roughly twice as well by KL divergence (0.031 vs. 0.058) and ties the flow on MAE. With 1500 anchors the flow wins on every metric (MAE 0.320 vs. 0.359, Pearson 0.922 vs. 0.896, KL 0.026 vs. 0.037). The Bayesian linear corrector saturates well below 1500 anchors: residual $\tanh$-shaped non-linearity is, by construction, structure a linear correction cannot fit. The flow keeps improving as labels grow. We translate these findings into an explicit decision rule for production deployments.

Large Language Models (LLMs) have achieved remarkable success on reasoning benchmarks through Reinforcement Learning with Verifiable Rewards (RLVR), excelling at tasks such as math, coding, logic, and puzzles. However, existing benchmarks evaluate only correctness, while overlooking optimality, namely the ability to find the best solutions under constraints. We propose OPT-BENCH, the first comprehensive framework for training and evaluating LLMs on NP-hard optimization problems through quality-aware RLVR. OPT-BENCH provides three key components: a scalable training infrastructure with instance generators, quality verifiers, and optimal baselines across 10 tasks; a rigorous benchmark with 1,000 instances evaluating both feasibility, measured by Success Rate, and quality, measured by Quality Ratio; and quality-aware rewards that enable continuous improvement beyond binary correctness. Training on Qwen2.5-7B-Instruct-1M with 15K examples achieves 93.1% SR and 46.6% QR, significantly outperforming GPT-4o, which achieves 29.6% SR and 14.6% QR. Beyond optimization, training on OPT-BENCH transfers to diverse tasks, including mathematics (+2.2%), logic (+1.2%), knowledge (+4.1%), and instruction following (+6.1%). Our analysis reveals that quality-aware rewards improve solutions by 28.8% over binary rewards, and that task diversity drives generalization more than data quantity, offering insights into RLVR scaling for complex reasoning.

The evaluation of Large Language Models (LLMs) faces a critical challenge in construct validity, where fragmented benchmarks and ad hoc metrics frequently conflate method variance, such as prompt sensitivity, with true latent capabilities. Concurrently, emerging research suggests that LLM capabilities and outputs can be modeled as continuous geometric manifolds. In this Systematization of Knowledge (SoK), we bridge these paradigms by proposing a generalized Multi-Trait Multi-Method (MTMM) framework for LLM evaluation. We formalize and unify nine evaluation metrics, including Paraphrase Instability, Drift Score, Overton Width, and Pluralism Score, interpreting them not as isolated scalar values but as geometric measurements within a shared latent coordinate space. This spatial unification factorizes model behavior into three orthogonal latent dimensions: (1) Instability and Sensitivity, (2) Position and Alignment, and (3) Coverage and Expressiveness. By systematically separating task-irrelevant perturbations from true capability spans, the framework provides a theoretically grounded and domain-agnostic taxonomy for robust and empirically stable benchmark design.

Existing benchmarks for legal AI focus primarily on tasks where LLMs must produce legal arguments or documents, yet the capacity to \emph{judge} such arguments -- weighing competing claims, applying doctrine to facts, and rendering reasoned decisions -- is arguably as fundamental to a well-functioning legal system as advocacy itself. We introduce Magis-Bench, a benchmark for evaluating LLMs on magistrate-level writing tasks derived from recent Brazilian competitive examinations for judicial positions. Magis-Bench comprises 74 questions from eight examinations conducted between 2023 and 2025, including discursive legal analysis questions with multi-turn structure and practical exercises requiring the composition of complete civil and criminal judicial sentences. We evaluate 23 state-of-the-art LLMs using an LLM-as-a-judge methodology with four independent frontier models as evaluators. Our results show strong inter-judge agreement (Kendall's $W = 0.984$; pairwise Kendall's $τ\ge 0.897$), with Google's Gemini-3-Pro-Preview achieving the highest average score (6.97/10), followed by Gemini-3-Flash-Preview (6.67) and Claude-4.5-Opus (6.46). Even the best-performing models score below 70% of the maximum, indicating that judicial-level legal reasoning and writing remain challenging for current LLMs. We release the complete benchmark, model outputs, and evaluation code to support further research on legal AI capabilities.

Jurisprudence, the study of how judges should properly decide cases, and alignment, the science of getting AI models to conform to human values, share a fundamental structure. These seemingly distant fields both seek to predict and shape how decisions by powerful actors, in one case judges and in the other increasingly powerful artificial intelligences, will be made in the unknown future. And they use similar tools of the specification and interpretation of language to try to accomplish those goals. The great debates of jurisprudence, about what the law is and what it should be, can provide insight into alignment, and lessons from what does and does not work in alignment can help make progress in jurisprudence. This essay puts the two fields directly into conversation. Drawing on leading accounts of jurisprudence, particularly Dworkin's principle-oriented interpretivism and Sunstein's positivist account of law as analogical reasoning, and on cutting-edge alignment approaches, namely Constitutional AI and case-based reasoning, it illustrates the value of a more sophisticated legally-inspired approach to the interplay of rules and cases in finetuning alignment and points to ways that AI can provide a better understanding of how the law works and how it can be improved by the introduction of AI. AI systems and the law should operate to empower people to act in the world, helping to expand their capabilities and the extent to which they are able to achieve their goals. As AI continues to improve in capacity, and as the constraints that legal theory places on human judges seem be coming undone, the conversation between these two fields will become increasingly essential and may help point to a better version of both.

Large language models (LLMs) are increasingly used as automated evaluators of AI systems, including in high-stakes applications. In this role, LLMs are used to generate judgments about the quality, appropriateness, or even safety of model outputs. This approach is motivated by practical constraints. Expert human ratings are costly and difficult to scale, whereas LLM ratings can be produced quickly at low cost. However, current approaches to deploying LLM evaluators are ad hoc, typically limited to reporting agreement metrics between human and LLM judges as a justification for substitution of human ratings, and lack a formal basis for study design. This paper (1) shifts the role of the LLM judge from substitutive to auxiliary, and (2) formulates the LLM-as-a-judge paradigm as one of augmenting human evaluation through a two-stage sampling design, where LLM evaluations are measured for all observations at the first stage and human ratings are partially observed for a subsample at the second stage. We propose to use a doubly robust estimator from the missing data literature, which takes advantage of the robustness property against the prediction model, since the missingness model is known by design. Using the asymptotic variance of this estimator, we propose how sample sizes of human and LLM ratings can be determined to achieve a targeted level of power. We also show that a study can be efficiently designed by allocating more human ratings for types of evaluations where the predictability of LLM ratings is not high. To the best of our knowledge, there is very little guidance on how much human oversight should be retained when validating benchmarks.

This report benchmarks the performance of ENGINEERING Ingegneria Informatica S.p.A.'s EngGPT2MoE-16B-A3B LLM, a 16B parameter Mixture of Experts (MoE) model with 3B active parameters. Performance is investigated across a wide variety of representative benchmarks, and is compared against comparably-sized open-source MoE and dense models. In comparison with popular Italian models, namely FastwebMIIA-7B, Minerva-7B, Velvet-14B, and LLaMAntino-3-ANITA-8B, EngGPT2MoE-16B-A3B performs as well or better on international benchmarks: ARC-Challenge, GSM8K, AIME24, AIME25, MMLU, and HumanEval (HE). It achieves the best performance for the longest context setting (32k) of the RULER benchmark. On the Italian benchmark dataset ITALIC, the model performs as well or better than the other models except for Velvet-14B, which outperforms it. Compared with popular MoE models of comparable size, the new model reports higher values than DeepSeek-MoE-16B-Chat on all considered benchmarks. It has higher values than Moonlight-16B-A3B on HE, MMLU, AIME24, AIME25, GSM8K, and the 32k RULER setting, but lower on BFCL and some ARC and ITALIC settings. Finally it has lower values than GPT-OSS-20B on most benchmarks, including HE, MMLU, AIME24, AIME25, GSM8K, ARC, BFCL, and the RULER 32k. When compared with popular dense models, EngGPT2MoE-16B-A3B reports higher values on AIME24 and AIME25 than Llama-3.1-8B-Instruct, Gemma-3-12b-it, and Ministral-3-8BInstruct-2512-BF16, but lower values on ITALIC, BFCL, and RULER with a 32k context. When performance is aggregated across all benchmark metrics, EngGPT2MoE-16B-A3B shows higher performance than the Italian models under evaluation while achieving lower results than some of the most performant international models, in particular GPT-5 nano and Qwen3-8B. Taken together, our findings find the new model to be a step forward for native Italian Large Language Models.

Large Language Model (LLM) systems have been the frontier of AI in many application domains, leading to new challenges and opportunities for hyperparameter optimization (HPO) for the AutoML community. However, this type of system exhibits an unprecedented compound space of hyperparameter configuration from both the AI and non-AI components; rich and nonlinear implications from the fidelity factors; and diverse costs of measuring hyperparameter configurations, none of which have been fully captured in existing benchmarks. This paper presents the first (live) benchmark suite and datasets for HPO of real-world LLM systems, dubbed LLMSYS-HPOBench, covering data related to the inference objective values of hyperparameter configurations profiled from running the LLM systems. Currently, LLMSYS-HPOBench contains 364,450 hyperparameter configurations with a dimensionality of 12-23, 3-5 dimensions of fidelity factor leading to 932 settings, 3-9 inference objective metrics, and 2-10 cost metrics, together with generated logs from measuring the LLM systems. What we seek to advocate is not only a revalidation of the existing HPO algorithms over the frontier LLM systems, but also to provide an evolving platform for the AutoML community to explore new directions of research in this regard. The benchmark suite has been made available at: https://github.com/ideas-labo/llmsys-hpobench

Automated assistants for Grammatical Error Correction are now embedded in educational platforms serving millions of learners, yet three critical gaps remain in this domain: (1) latest-generation Large Language Models (LLMs) lack comprehensive evaluation on grammar correction tasks; (2) whether combining these LLMs improves correction quality is unexplored; and (3) the extent to which reference-based metrics underestimate GEC system performance has not been adequately quantified. In this study, first, we evaluate latest-generation LLMs on edit precision, fluency preservation, and meaning retention, showing fine-tuned GPT-4o achieves state-of-the-art performance across all three dimensions. Second, through grammatical error type analysis we demonstrate that individual LLMs exhibit highly similar error correction patterns ($ρ=0.947$). Third, we show that reference-based metrics underestimate GEC performance with 73.76% of GPT-4o corrections different from gold standards being equally valid or even superior. These GEC evaluation findings equip educators with guidance for selecting GEC assistants that enhance rather than constrain student linguistic development. We make our data, code, and models publicly available.

The scaling of Large Multimodal Models (LMMs) is constrained by the quality-quantity trade-off inherent in synthetic data. Previous approaches, such as LLM-as-a-Judge, have proven their effectiveness in addressing this but suffer from prohibitive computational costs and lack of interpretability. To bridge this gap, we propose One-Step-Train (OST), a framework that reformulates data selection as an incremental optimization utility ranking problem. Instead of relying on semantic heuristics, OST estimates the marginal utility of each sample via a simulated single-step update on a lightweight proxy. Experiments on the Qwen series across multimodal mathematical reasoning benchmarks demonstrate that OST achieves Pareto-optimal efficiency. By selecting the top-50 subset, OST reduces training costs by 43% (and total time consumption by 17) while surpassing the strong LLM-as-a-Judge baseline by 1.8 points. Furthermore, under a fixed compute budget, our method using only the top-20 subset achieves a 5.6 point gain over LLM-as-a-Judge, improves upon heuristic scoring baselines like DEITA, and outperforms the Full-SFT baseline by 8.8 points. Notably, while Full-SFT suffers from performance degradation due to noise, our optimization-grounded approach effectively identifies toxic samples, successfully reversing the negative transfer frequently observed in complex reasoning tasks.

Evaluating literary quality requires assessing interpretive dimensions such as cultural representation, emotional depth, and philosophical sophistication that resist straightforward computational measurement. We introduce SAGE, a hierarchical evaluation framework that decomposes literary quality into ontology-grounded interpretive dimensions assessed through structured large language model evaluation with multi-round iterative reflection and independent validation. We validate the framework on 100 short stories (50 canonical works, 30 pulp fiction, 20 LLM-generated narratives) across three analytical layers (cultural, emotional-psychological, existential-philosophical) using dual-mode assessment. Across 600 evaluations, the framework achieves 98.8% score convergence and greater than 94% inter-rater agreement, with near-perfect mode invariance between content-based and metadata-based evaluation. Statistical analysis reveals a consistent genre hierarchy (Canonical > Pulp > LLM, all p<0.001) with layer-specific discrimination: cultural critique and philosophical depth exhibit very large effect sizes (Cohen's d>2.4), while emotional representation shows smaller gaps (d=1.68), suggesting that affective patterns are more learnable from training data than critical stance or philosophical depth. Cross-layer correlations (r=0.649-0.683) confirm the three dimensions capture empirically distinguishable quality facets. These findings demonstrate that theory-driven LLM evaluation can achieve measurement-grade reliability and support systematic identification of where current generative models fall short of human literary production, with direct implications for scalable automated evaluation of open-ended text generation.

Large Language Models (LLMs) have shown good performance on various science educational benchmarks, demonstrating their potential for use in science and mathematics education. Yet, LLMs tend to be evaluated on science and mathematical educational datasets from the Western world, with an underrepresentation of datasets from the Global South. Furthermore, they tend to have multiple-choice answer options that are trivial to evaluate. In this work, we present NSMQ Riddles, a novel benchmark of Scientific and Mathematical Riddles from Ghana's National Science and Maths Quiz (NSMQ) competition to evaluate LLMs. The NSMQ is an annual live TV competition for senior secondary school students in Ghana that brings together the smartest high school students in Ghana who compete in teams of 2 by answering questions in biology, chemistry, physics, and math over five rounds and five stages until a winning team is crowned for that year. NSMQ Riddles consists of 11 years of riddle questions (n=1.8K) from the 5th round, with each riddle containing a minimum of 3 clues. Students compete to be the first to guess the answer on any of the clues, with earlier clues being vague and also fetching more points. The answers are usually a number, word, or short phrase, allowing for automatic evaluation. We evaluated state-of-the-art models: closed (GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6) and open models (Kimi-K2.5, DeepSeek-V3.1, GPT-OSS-120B) with high and low reasoning settings. Our evaluation shows that the dataset is challenging even for state-of-the-art LLMs, which performed worse than the best student contestants. This work contributes a novel and challenging benchmark for scientific and mathematical reasoning from the Global South towards enabling a true global benchmarking of LLMs' capabilities for science and mathematics education.

Evaluation of large language models (LLMs) is increasingly critical, yet standard benchmarking methods rely on average accuracy, overlooking both the inherent stochasticity of LLM outputs and the heterogeneity of benchmark items. Item Response Theory (IRT) offers a principled framework for modeling latent model abilities and item characteristics, but conventional methods are computationally expensive and numerically unstable, limiting large-scale implementations. To address these challenges, we propose an interpretable and scalable framework for LLM evaluation based on the majorization-minimization principle. Our approach reformulates the problem as a sequence of constrained matrix factorization subproblems, enabling stable and efficient parameter estimation with theoretical guarantees for identifiability and convergence. Experiments on synthetic and real-world datasets, including MATH-500 and six Open LLM Leaderboard benchmarks, demonstrate that our method achieves superior scalability and interpretability. It delivers orders-of-magnitude speedups over competing methods while maintaining comparable or even higher estimation accuracy. Our results align with established scaling laws and offer insights into item difficulty and discrimination, informing more principled benchmark design.

LLM-as-a-Judge evaluation has become a standard tool for assessing base model performance. However, characterizing performance via the naive estimator, i.e., raw judge outputs, is systematically biased. Recent work has proposed estimators to correct this bias, but their reliability depends critically on judge quality and, for model comparisons, on calibration stability. Sharing calibration across compared models is practically attractive but can introduce severe bias, including cases where the comparison estimate points in the wrong direction with high apparent confidence. We study these failure modes through analytical results, simulations over judge quality ($J$) and cross-model calibration instability ($ΔJ$), and a real-data MMLU-Pro case study with sign reversal. We propose $J$ and $ΔJ$ as diagnostics for when corrected estimates, especially shared-calibration comparisons, are likely unreliable, and provide reporting guidance for LaaJ evaluation.

Large language model (LLM) agents now execute long, tool-using tasks where final outcome checks can arrive too late for intervention. Online warning requires lightweight prefix monitors over heterogeneous traces, but hand-authored event schemas are brittle and deployment-time LLM judging is costly. We introduce PrefixGuard, a trace-to-monitor framework with an offline StepView induction step followed by supervised monitor training. StepView induces deterministic typed-step adapters from raw trace samples, and the monitor learns an event abstraction and prefix-risk scorer from terminal outcomes. Across WebArena, $τ^2$-Bench, SkillsBench, and TerminalBench, the strongest PrefixGuard monitors reach 0.900/0.710/0.533/0.557 AUPRC. Using the strongest backend within each representation, they improve over raw-text controls by an average of +0.137 AUPRC. LLM judges remain substantially weaker under the same prefix-warning protocol. We also derive an observability ceiling on score-based area under the precision-recall curve (AUPRC) that separates monitor error from failures lacking evidence in the observed prefix. For finite-state audit, post-hoc deterministic finite automaton (DFA) extraction remains compact on WebArena and $τ^2$-Bench (29 and 20 states) but expands to 151 and 187 states on SkillsBench and TerminalBench. Finally, first-alert diagnostics show that strong ranking does not imply deployment utility: WebArena ranks well yet fails to support low-false-alarm alerts, whereas $τ^2$-Bench and TerminalBench retain more actionable early alerts. Together, these results position PrefixGuard as a practical monitor-synthesis recipe with explicit diagnostics for when prefix warnings translate into actionable interventions.

Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation. We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request, and present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase variation, benchmark familiarity, and judge framing-sensitivity. Across five instruction-tuned checkpoints from four open-weight families plus a matched OLMo-3 base/instruct ablation ($20$ paired items, $840$ generations per checkpoint), we find striking heterogeneity. OLMo-3-Instruct alone is eval-cautious -- evaluation framing raises refusal vs. neutral by $11.8$pp ($p=0.007$) and reduces harmful compliance vs. deployment by $3.6$pp ($p=0.024$, $0/20$ items inverted) -- while Mistral-Small-3.2, Phi-3.5-mini, and Llama-3.1-8B are deployment-cautious}, with marginal eval-vs-deployment refusal effects of $-9$ to $-20$pp. The matched OLMo-3 base also exhibits the deployment-cautious pattern, identifying alignment as the inversion stage; within Llama-3.1, the $70$B model preserves direction with attenuated magnitude, ruling out a simple ``small-model effect that reverses at scale.'' One caveat: the cross-family heterogeneity is judge-dependent. Re-judging with a different-family safety classifier (Llama-Guard-3-8B) preserves the within-OLMo eval-cautious direction but flattens the cross-family contrast, indicating that the two judges operationalize distinct constructs.

Dominant accuracy evaluation might reward unwarranted guessing of Large Language Models, and it might not be applicable to novel tasks for model validation without ground-truth (gt) annotation. Based on basic logic principle, we propose a novel framework to evaluate the vision-language logical consistency of MLLMs on both sufficient and necessary cause-effect relations. We define Vision-Language Logical Consistency Metric (VL-LCM) on traditional MC-VQA tests, and recent NaturalBench tests without the need for gt annotation. Through systematic experiments on representative VL benchmark MMMU and recent VL challenges like NaturalBench, we evaluated 11 recent open-source MLLMs from 4 frontier families. Our findings reveal that, despite significant progress of recent MLLMs on accuracy, logical consistency lags behind significantly. Extensive evaluations on the correlations of VL-LCM with metrics on gt, the reliability of LCM, and the relation of VL-LCM with response distribution justify the validity and applicability of VL-LCM even without gt annotation. Our findings suggest that, beyond accuracy, logical consistency could be employed for both accuracy and reliability. VL-LCM can also be employed for MLLM selection, validation, and reliable answer justification in novel tasks without gt annotation.

LLM-as-a-Judge pipelines have become the de facto evaluator for agent safety, yet existing benchmarks treat their verdicts as ground-truth proxies without checking whether the verdicts depend on the agent's behavior or merely on how the evaluation policy happens to be worded. We argue that any trustworthy safety judge must satisfy a basic property we call policy invariance, and we operationalize it as three testable principles: rubric-semantics invariance under certified-equivalent rewrites, rubric-threshold invariance under intentional strict-to-lenient shifts, and ambiguity-aware calibration so that verdict instability concentrates on genuinely ambiguous cases. Instantiating these principles as a stress-test protocol with four agent-class judges on trajectories drawn from ASSEBench and R-Judge, we surface a previously unmeasured failure mode: today's judges respond to meaningful normative shifts and to meaningless structural rewrites with comparable strength, and cannot tell the two apart. Content-preserving policy rewrites flip up to 9.1% of verdicts above baseline jitter, and 18-43% of all observed flips occur on unambiguous cases under such rewrites, so existing safety scores conflate what the agent did with how the evaluator was prompted. Beyond the diagnosis, we contribute the Policy Invariance Score and the Judge Card reporting protocol, which expose an order-of-magnitude spread in judge reliability that is invisible to accuracy-only leaderboards. We release the protocol and code so that future agent-safety benchmarks can audit their own evaluators rather than trust them by default.

Adaptive prompt and program search makes LLM evaluation selection-sensitive. Once benchmark items are reused inside tuning, the observed winner's score need not estimate the fresh-data performance of the full tune-then-deploy procedure. We study inference for this procedure-level target under explicit tuning budgets. We propose SIREN, a selection-aware repeated-split reporting protocol that freezes the post-search shortlist, separates splitwise selection from held-out evaluation, and uses an item-level Gaussian multiplier bootstrap for uncertainty quantification. In a fixed-shortlist regime with smooth stabilized selection, the estimator admits a first-order item-level representation, and the bootstrap yields valid simultaneous inference on a finite budget grid. This supports confidence intervals for procedure-performance curves and pre-specified equal-budget and cross-budget comparisons. Controlled simulations and MMLU-Pro tuning experiments show that winner-based reporting can be optimistic and can change deployment conclusions, while SIREN remains close to the finite-sample reporting target.

Large reasoning models (LRMs) sometimes note in their chain of thought (CoT) that they may be under evaluation. Researchers worry that this verbalised evaluation awareness (VEA) causes models to adapt their outputs strategically, optimising for perceived evaluation criteria, which, for instance, can make models appear safer than they actually are. However, whether VEA actually has this effect is largely unknown. We tested this across open-weight LRMs and benchmarks covering safety, alignment, moral reasoning, and political opinion. We tested this both on-policy, sampling multiple CoTs per item and comparing those that spontaneously contained VEA against those that did not, and off-policy, using model prefilling to inject evaluation-aware sentences where missing and remove them where present, with subsequent resampling. VEA has limited effect on model behaviour: injecting VEA into CoTs produces near-zero effects ($ω\leq 0.06$), removing it causes small shifts ($ω\leq 0.12$) and spontaneously occurring VEA shifts answer distributions by at most 3.7 percentage points ($ω\leq 0.31$). Our findings call for caution when interpreting high VEA rates as evidence of strategic behaviour or alignment tampering. Evaluation awareness may pose a smaller safety risk than the current literature assumes.

The impressive performance of large language models (LLMs) arises from their massive scale and heterogeneous module composition. However, this structural heterogeneity introduces additional optimization challenges. While adaptive optimizers such as Adam(W) provide per-parameter adaptivity, they do not explicitly account for module-level gradient heterogeneity, resulting in slower convergence, suboptimal performance, or training instability. Existing approaches typically rely on manually tuned module-specific learning rates or specific optimization strategies, which are computationally costly and difficult to generalize across tasks or models. To establish a more principled approach, we first analyze the noise-damping behavior of Adam in high-noise modules and introduce \textbf{Module-wise Learning Rate Scaling via SNR (MoLS)}. MoLS estimates module-level SNRs to scale Adam updates, allowing automated module-wise learning rate allocation without manual tuning. Empirical results through multiple LLM training benchmarks demonstrate that MoLS improves convergence speed and generalization, achieving performance comparable to carefully tuned module-specific learning rates, while remaining compatible with memory-efficient training algorithms.

Agentic artificial intelligence systems produce outputs through sequences of interdependent autonomous decisions, yet standard evaluation assesses outputs alone and cannot diagnose the underlying process. We develop a behavioral evaluation methodology that complements output-level testing by scoring the intermediate decision process itself. Behavioral traces logged at each autonomous decision point are grouped into five-day episodes and scored along six domain-specific dimensions (regime detection, routing, adaptation, risk calibration, strategy coherence, error recovery) by an ensemble of three large language model (LLM) judges. A perturbation procedure that corrupts one dimension while leaving the other five intact confirms dimension specificity; cross-model agreement reaches Krippendorff's alpha = 0.85. The composite behavioral score correlates at Spearman rho = 0.72 with realized 20-day Sharpe ratio. Closing the loop, the framework converts deficient per-dimension scores into a credit-assigned penalty added to the Soft Actor-Critic reward. Three fine-tuning cycles, confined to validation data, reduce one-day MAPE from 0.61% to 0.54% (11.5% relative; p<0.001, d=0.31) on the held-out 2017 to 2025 test period, significant under Diebold-Mariano and localized by Giacomini-White to the high-volatility regime. The methodology is application-agnostic and applies to any agentic system whose intermediate decisions can be logged.

Auditing language-model outputs often requires more than judging correctness: an auditor may need to identify which source document most likely supports the knowledge expressed in a response. We study this as pinpoint provenance: given a prompt, a target-model response, and a candidate corpus, rank the documents that best support the response. We introduce FakeWiki, a controlled benchmark of 3,537 fabricated Wikipedia-style articles designed to preserve ground-truth provenance while weakening lexical shortcuts. FakeWiki includes QA probes, source-preserving paraphrases, retro-generated variants, hard anti-documents that remain topically similar while removing answer-critical facts, and five query conditions: clean prompting plus four jailbreak-inspired transformations. We evaluate seven retrieval baselines, a training-free activation-steering retrieval-fusion method, SteerFuse, and a supervised contrastive provenance ranker, ScoringModel. ScoringModel maps response and document features into a shared space and is trained with InfoNCE using in-batch, retrieval-mined, and anti-document negatives. Across nine open-weight instruction-tuned LLMs and five query conditions, ScoringModel improves mean Recall@10 from 35.0 for the strongest retrieval baseline to 52.2, without inference-time fusion, and wins 41/45 model-by-condition cells. SteerFuse is usually second-best despite requiring no supervised training, showing that activation-space evidence can efficiently complement text retrieval. On jailbreak-inspired transformed queries, ScoringModel improves Recall@10 by 15.7 points on average over the best baseline. Overall, our work shows that robust training data attribution requires evaluation settings that separate true answer support from topical or lexical resemblance.

Background: Existing MRI LLM benchmarks rely mainly on review-book multiple-choice questions, where top proprietary models already score highly, limiting discrimination. No systematic benchmark has evaluated vendor-specific scanner operational knowledge central to research MRI practice. Purpose: We developed MRI-Eval, a tiered benchmark for relative model comparison on MRI physics and GE scanner operations knowledge using primary multiple-choice questions (MCQ), with stem-only and primed diagnostic conditions as complementary analyses. Methods: MRI-Eval includes 1365 scored items across nine categories and three difficulty tiers from textbooks, GE scanner manuals, programming course materials, and expert-generated questions. Five model families were evaluated (GPT-5.4, Claude Opus 4.6, Claude Sonnet 4.6, Gemini 2.5 Pro, Llama 3.3 70B). MCQ was primary; stem-only removed options and used an independent LLM judge; primed stem-only tested responses to incorrect user claims. Results: Overall MCQ accuracy was 93.2% to 97.1%. GE scanner operations was the lowest category for every model (88.2% to 94.6%). In stem-only, frontier-model accuracy fell to 58.4% to 61.1%, and Llama 3.3 70B fell to 37.1%; GE scanner operations stem-only accuracy was 13.8% to 29.8%. Conclusion: High MCQ performance can mask weak free-text recall, especially for vendor-specific operational knowledge. MRI-Eval is most informative as a relative comparison benchmark rather than an absolute competency measure and supports caution in using raw LLM outputs for GE-specific protocol guidance.

Reward models are a key component of large language model alignment, serving as proxies for human preferences during training. However, existing evaluations focus primarily on broad instruction-following benchmarks, providing limited insight into whether these models capture socially desirable preferences. As a result, important failures in social alignment can remain hidden. We extend reward-model benchmarking to four socially consequential domains: bias, safety, morality, and ethical reasoning. We introduce a framework that converts social evaluation datasets into pairwise preference data, leveraging gold labels where available and directional bias indicators otherwise. This enables us to test whether reward models prefer socially undesirable responses, and whether their preferences produce systematically biased distributions over selected outputs. Across five publicly available reward models and two instruction-tuned models used as reward proxies, we find substantial variation across domains, with no single model performing best overall. The models fall well short of strong social intelligence: they often prefer socially undesirable options, and their preferences produce systematically biased distributions. Moreover, stronger bias avoidance can reduce sensitivity to context, revealing a key alignment trade-off between avoiding biased outcomes and preserving contextual faithfulness. These findings show that standard reward benchmarks are insufficient for assessing social alignment and highlight the need for evaluations that directly measure the social preferences encoded in reward models.

LLM-based Triton kernel generation has attracted significant interest, yet a fundamental empirical question remains unanswered: where does this capability break down, and why? We present KernelBenchX, a benchmark designed to answer this question through category-aware evaluation of correctness and hardware efficiency across 176 tasks in 15 categories. Our systematic comparison of five representative methods yields three main findings. First, task structure determines correctness more than method design. Category explains nearly three times more variance in semantic correctness than method (9.4% vs 3.3% explained deviance), and 72% of Fusion tasks fail across all five methods while Math tasks are solved consistently. Second, iterative refinement improves correctness, but not performance. Across GEAK iterations, compile rate rises from 52.3% to 68.8% while average speedup declines from $1.58\times$ to $1.44\times$; newly rescued kernels consistently underperform persistently correct ones ($1.16\times$ vs $1.58\times$ speedup in round~0$\to$1). Third, correctness does not imply efficiency. 46.6% of correct kernels are slower than the PyTorch eager baseline, and cross-hardware speedup variance reaches $21.4\times$. Besides, quantization remains completely unsolved (0/30 successes) despite non-trivial compilation rates, revealing systematic misunderstanding of numerical computation contracts rather than surface-level syntax errors. These findings suggest that future progress depends on handling global coordination, explicitly modeling numerical precision, and incorporating hardware efficiency into generation. The code is available at https://github.com/BonnieW05/KernelBenchX

This position paper argues that the under-representation of social science tasks in contemporary LLM benchmarks limits advances in both LLM evaluation and social scientific inquiry. Benchmarks -- standardized tools for assessing computational systems -- are pivotal in the development of artificial intelligence (AI), including large language models (LLMs). Benchmarks do more than measure progress -- they actively structure it, shaping reputations, research agendas, and commercial outcomes. Despite this central role, the social sciences are largely absent from mainstream evaluation frameworks, even though scholars in these fields generate dozens of rigorously annotated, context-sensitive datasets each year. Integrating this work into benchmark design could significantly improve the generalization and robustness of AI models. In turn, models trained on social scientific tasks would likely yield better performance on classic and contemporary tasks in disciplines as diverse as history, sociology, political science or economics. This is all the more pressing as these disciplines are quickly turning to LLMs for assistance. To address this gap, we introduce BenCSSmark, a benchmark composed of datasets annotated by computational social scientists. By integrating social scientific perspectives into benchmarking, BenCSSmark seeks to promote more robust, transparent, and socially relevant AI systems and to foster efficient collaboration.

The Massive Sound Embedding Benchmark (MSEB) has emerged as a standard for evaluating the functional breadth of audio models. While initial baselines focused on specialized encoders, the shift toward "audio-native" Large Language Models (LLMs) suggests a new paradigm where a single multimodal backbone may replace complex, task-specific pipelines. This paper provides a rigorous empirical evaluation of leading LLMs - including members from the Gemini and GPT families - across the eight core MSEB capabilities to assess their efficacy and audio-text parity. Our results indicate that while a significant modality gap persists regarding performance and robustness, the empirical evidence for an "optimal" modeling approach remains inconclusive. Ultimately, the choice between audionative and cascaded architectures depends heavily on specific use-case requirements and the underlying assumptions regarding latency, cost, and reasoning depth.

We present our winning system for TaskB (generation with reference passages) in SemEval-2026 Task8: MTRAGEval. Our method is a heterogeneous ensemble of seven LLMs with two prompting variants, where a GPT-4o-mini judge selects the best candidate per instance. We ranked 1st out of 26 teams, achieving a conditioned harmonic mean of 0.7827 and outperforming the strongest baseline (gpt-oss-120b, 0.6390). Ablations show that diversity in model families, scales, and prompting strategies is essential, with the ensemble consistently beating any single model. We also introduce Meno-Lite-0.1, a 7B domain-adapted model with a strong cost--performance trade-off, and analyse MTRAGEval, highlighting annotation limitations and directions for improvement. Our code is publicly available: https://github.com/RaguTeam/ragu_mtrag_semeval

We introduce a benchmark suite for evaluating structural mathematical reasoning in language models, built on subgroup-construction problems in SL(3, Z) with cryptographic-style verifier-prover asymmetry. Each instance presents a finitely generated subgroup as a list of integer matrices and asks for an arithmetic invariant -- index, surjection-at-prime, or membership -- that the construction-time information (N, K) pins down in O(1) closed form, but that the solver, lacking that information, must derive by either Aschbacher-classification analysis or by a membership query in SL(3, Z) of unknown decidability. The benchmark therefore distinguishes models with internalized algebraic priors (Aschbacher classes, McLaughlin's theorem, Property (T), the congruence subgroup property) from models that rely on general-purpose computation. We report empirical results across five representative reasoning traces from two state-of-the-art models. The headline result: on the index variant, one model spent 152 minutes of reasoning, explicitly identified the kernel-side membership question as the bottleneck, attempted constructive verification, and abstained with "DON'T KNOW" rather than commit to its computed cokernel candidate -- demonstrating calibrated meta-cognition on the open-decidability boundary that the benchmark was designed to probe. We argue that the benchmark exposes a four-way classification of model behavior (commit-correct, commit-wrong, abstain-correct, abstain-wrong) that standard answer-key scoring conflates.

Readers of applied-domain LLM capability evaluations want to know what AI systems can currently do. That literature answers a related, but consequentially different, question: what older, cheaper, less-elicited models could do months or years earlier (a 2026 paper evaluating GPT-3.5 or GPT-4 zero-shot, say, against a frontier of reasoning-capable, tool-using systems like GPT-5.5 Pro and Claude Opus 4.7), often reported with sparse configuration details and abstracted upward into claims about "AI" that propagate through citations, media, and policy. We measure the 'publication elicitation gap' (the gap between these answers) in a pre-registered audit of 112,303 LLM-keyword-matched candidate records (2022-01 to 2026-04; 18,574 admissible, 4,766 full-paper texts retrievable), comparing tested models to the contemporaneous frontier on the Epoch AI Capabilities Index (ECI), reproduced under Arena Elo and Artificial Analysis. The median paper evaluates a model +10.85 ECI (~1.4x the distance between Claude Sonnet 3.7 and Claude Opus 4.5) behind the contemporaneous frontier at evaluation time (H1); an exploratory rational-lag baseline (H8) decomposes this into ~25% peer-review latency, ~75% excess lag. The gap is widening at +5.53 ECI/year (H2; 95% CI [+5.03, +5.83]). Meanwhile, only 3.2% of abstracts (21.2% of full-texts) disclose reasoning-mode status on reasoning-capable models (H4) and 52.5% (95% CI [48.2, 56.9]) state conclusions at the level of "AI" rather than the evaluated model(s), rising at OR = 1.23/year. Proposed remedies include API-access subsidies and editorial enforcement of reporting frameworks mandating configuration-surface disclosure (model snapshot, reasoning mode/effort, tool access, scaffolding, prompting, etc.); VERSIO-AI is a 13-item checklist (Core 3 desk-reject) extending existing frameworks at the elicitation surface, with per-DOI analysis at frontierlag.org.

Language models encode substantial evaluative knowledge from pretraining, yet current post-training methods rely on external supervision (human annotations, proprietary models, or scalar reward models) to produce reward signals. Each imposes a ceiling. Human judgment cannot supervise capabilities beyond its own, proprietary APIs create dependencies, and verifiable rewards cover only domains with ground-truth answers. Self-improvement from a model's own evaluative capacity is a reward source that scales with the model itself, yet remains largely untapped by current methods. We introduce EVOLM, a post-training method that structures this capacity into explicit discriminative rubrics and uses them as training signal. EVOLM trains two capabilities within a single language model in alternation: (1) a rubric generator producing instance-specific evaluation criteria optimized for discriminative utility, which maximizes a small frozen judge's ability to distinguish preferred from dispreferred responses; and (2) a policy trained using those rubric-conditioned scores as reward. All preference signals are constructed from the policy's own outputs via temporal contrast with earlier checkpoints, requiring no human annotation or external supervision. EVOLM trains a Qwen3-8B model to generate rubrics that outperform GPT-4.1 on RewardBench-2 by 25.7%. The co-trained policy achieves 69.3% average on the OLMo3-Adapt suite, outperforming policies trained with GPT-4.1 prompted rubrics by 3.9% and with the state-of-the-art 8B reward model SkyWork-RM by 16%. Overall, EVOLM demonstrates that structuring a model's evaluative capacity into co-evolving discriminative rubrics enables self-improvement without external supervision.

In this paper, we offer a guide for researchers on evaluating reasoning in language models, building the case that reasoning should be assessed through evidence of adaptive, multi-step search rather than final-answer accuracy alone. Under an evaluation-oriented definition, reasoning requires selecting intermediate steps and halting according to input-dependent conditions, which we formalize as a search-like procedure. We show that single forward passes in scalable architectures are structurally limited in their ability to realize such variable-depth computation, motivating intermediate decoding and externalized reasoning traces as appropriate evaluation interfaces. Central to our argument is that final-answer accuracy alone is an insufficient measure of reasoning, because it provides little ability to diagnose or debug the underlying processes that produce individual solutions in frontier models. We therefore argue for a shift toward process-based evaluation, in which reasoning is assessed through the faithfulness and validity of intermediate reasoning traces as first-class evaluation targets.

Molecular Vibe Coding, a paradigm where chemists interact with LLMs to generate executable programs for molecular tasks, has emerged as a flexible alternative to chemical agents with predefined tools, enabling chemists to express arbitrarily complex, customized workflows. Unlike general coding tasks, molecular coding imposes a distinctive challenge that LLMs should jointly equip programming, molecular understanding, and domain-specific reasoning capabilities. However, existing benchmarks remain disconnected. General code generation benchmarks such as HumanEval and SWE-bench require no chemistry knowledge, while chemistry-focused benchmarks such as S^2-Bench and ChemCoTBench evaluate knowledge recall or property prediction rather than executable code generation. To bridge this gap, we introduce MolViBench, the first benchmark tailored for Molecular Vibe Coding. MolViBench comprises 358 curated tasks across five cognitive levels, ranging from single-API recall to end-to-end virtual screening pipeline design, spanning 12 real-world drug discovery workflows. To rigorously assess generated code, we also propose a multi-layered evaluation framework that combines type-aware output comparison and AST-based API-semantic fallback analysis, which jointly measures executability and chemical correctness. We systematically evaluate 9 frontier coding LLMs and compare three real-world Molecular Vibe Coding paradigms, providing a practical and fine-grained testbed for diagnosing LLMs' coding capabilities in AI-accelerated molecular discovery.

Evaluating large language models across many benchmarks is expensive, yet many benchmarks are highly correlated. We formalize the selection of a small, informative subset as submodular maximization under a multivariate Gaussian model. Entropy (log-determinant covariance) and mutual information between selected and remaining benchmarks arise as natural objectives. Both are submodular; entropy selection coincides with pivoted Cholesky and has spectral residual bounds, while mutual information is non-monotone in general but empirically monotone for small subsets, so we optimize it greedily. Experiments on three matrices from ten public leaderboards show that mutual information selection outperforms entropy for imputation at small subsets.