LLM Evaluation

We extend the DS@GT ARC working-note submission to the Touché 2025 Retrieval-Augmented Debate task. The task has two subtasks: generating the next utterance in a simulated debate, and evaluating debate responses according to the Gricean maxims of Quantity, Quality, Relation, and Manner. The DS@GT ARC submission consisted of six leading LLMs from three providers through a retrieval-augmented prompting pipeline. We summarize the results from the working paper and explore whether multi-LLM evaluator agreement is a reliable proxy for official evaluation performance. The analysis shows that frontier LLM systems are strong response generators, and as evaluators they agree strongly within model families. However this consensus does not reliably track the official evaluation target, with the largest gap on the Quality maxim. The accompanying source code for this paper is located at https://github.com/dsgt-arc/touche-2025-rad and https://github.com/dsgt-arc/touche-2025-rad-analysis.

The key question in moral judgement is not simply whether someone chooses the "right" answer, but how they decide what matters most when moral principles conflict. Current evaluations of large language models (LLMs) remain limited: most test whether models give morally acceptable answers, match human preferences, or avoid obvious violations, rather than whether they can prioritise between competing principles when no option is morally cost-free. We introduce CORDA (Conditioned Ordering and Ranked Directive Adherence), a benchmark for evaluating hierarchical, harm-centred moral reasoning in LLMs. Building on the morality chains formalism, CORDA tests 90 moral dilemmas involving trolley-style cases, medical trade-offs, resource allocation, and human-animal-robot conflicts across four ordered ethical frameworks: Utility, Utility + Agent Harm, Dual-Process, and Dual-Process + Agent Harm. Together, these frameworks test whether models can adapt their decisions when moral priorities change. Across ten instruction-tuned models from seven providers, we find a strong deontological default, with 9 of 10 prioritising avoidance of direct personal harm over reducing overall harm. Models also perform more reliably on categorical harm-avoidance rules, such as avoiding killing, than on outcome-based comparisons, such as minimising total harm, suggesting that they recognise moral red lines more easily than they reason through competing harms. Although all models respond to explicit chain conditioning, several fail to consistently follow specified priority orderings, such as humans over animals and animals over robots. CORDA addresses a central gap in LLM moral evaluation by testing whether models can move beyond default harm-avoidant responses and apply context-specified moral priorities. Moral reliability requires more than default restraint; it requires controllability under conflict.

Language-model judges and reward models enable scalable supervision, but finite optimization can exploit evaluator errors rather than improve response quality. We characterize this failure through the covariance geometry of evaluator ensembles. For calibrated judges, the ensemble mean retains common-mode error along the all-ones direction, whereas cross-judge disagreement captures only orthogonal error. Consequently, disagreement can be high despite robust aggregation, or low while shared response-dependent errors persist. We prove that common-mode error is not identifiable from internal judge scores alone. Under a joint sub-Gaussian model, we bound best-of-K selection overstatement and target-quality regret, extending the guarantees to predictably adaptive search under conditional calibration. The resulting search terms scale as the square root of log K and are asymptotically tight for Gaussian projected errors. We further show that noisy quality proxies introduce artificial rank-one covariance without changing disagreement, and propose a bounded two-anchor Bernstein certificate for finite-search error and regret. Fixed-seed Gaussian stress tests over 120 (J, rho, K) configurations and real-model audits validate the theory while revealing the limits of disagreement-based diagnostics under increasing search pressure.

Reasoning language models increasingly use test-time compute to improve performance, but existing evaluations typically study this compute one question at a time. Yet when multiple problems share an end-to-end cost or latency constraint, models must decide how to divide limited inference compute among them. We introduce an exam-style evaluation framework for studying this setting, in which a model must distribute one shared token budget across questions with different difficulty and point values to maximize its total score. Across several open and frontier reasoning models, we find that models fail to allocate a shared budget strategically across questions of varying difficulties and values. Models behave largely as greedy sequential solvers: they prioritize questions by presentation order, front-load effort on early questions, and remain insensitive to value, with these tendencies becoming more pronounced as the number of questions grows. Explicit planning prompts spread compute more evenly but do not produce value- or difficulty-aware prioritization. The same behavioral pattern extends from mathematical to code reasoning. These findings establish global budget allocation as a distinct capability that is not captured by conventional per-question evaluation and remains a challenge for current reasoning models.

Large language models (LLMs) have increasingly supported response generation grounded in user-provided knowledge spanning heterogeneous structures. However, existing benchmarks provide limited assessment of whether LLMs can faithfully perform multi-hop reasoning chains across such knowledge contexts while remaining robust to variations in their input order. We introduce TKFQA, a factuality consistency benchmark comprising 10,130 question-answering (QA) pairs grounded in tables, texts, and knowledge graphs (KGs). Each example is constructed from an explicit counterfactual reasoning chain, enabling the joint evaluation of answer correctness, reasoning-chain accuracy, and robustness to different input-order. An extensive evaluation of 14 open- and closed-source LLMs reveals that state-of-the-art models exhibit limited reasoning-chain accuracy and remain sensitive to variations in the input order of heterogeneous knowledge contexts. To address these limitations, we propose ORLF, an LLM-agnostic training framework that models cross-context topological relations through knowledge-specific latent vectors. ORLF integrates context-wise position encoding, a latent-bridge attention mask, and topological knowledge bias to preserve knowledge-specific bias and encode topological semantics. Experiments across four LLM backbones show that ORLF outperforms competitive training-free and LoRA-based baselines, improving average Exact Match and Reasoning-Chain Accuracy by 2.15% and 4.29%, respectively, while reducing order-induced performance standard deviation by 0.04% to 3.01%.

LLM benchmarks can build an organization's reputation and attract customers, but only when results are transparent and verifiable. Unverified claims that DeepSeek R1 outperformed OpenAI's o1 contributed to market panic on January 27, 2025, when Nvidia lost USD589 billion in market value. Yet vendor benchmarks often depend on an honor system. Academic reassessments and independent leaderboards have found undisclosed changes to proprietary models, contaminated training data, and selective reporting. LLM-as-a-judge methods scale evaluation by reducing human review. Studies, however, suggest that judges may show identity-aware bias, scoring an answer according to its source model rather than its quality. This bias has not been fully measured or corrected across politically sensitive, reasoning-intensive, and preference-based tasks. We examine this problem using seven verifier models: GPT-OSS 120B, Llama 3.3 70B, GLM 5.1, Qwen3 32B, DeepSeek V4 Pro, Mistral Large3, and Sarvam M. They score anonymous and identity-disclosed responses from three primary models on 58 factual, reasoning, political, and preference-based questions. Identity disclosure slightly raises scores for factual questions, moderately affects stress-reasoning tasks, and causes large changes for geopolitically sensitive topics. Notable results include GLM5.1 (+7.00 points, p = 0.0249) and Llama 3.3 70B (+1.56 points, p = 0.00). We also introduce a blockchain-based commit-reveal protocol using Autonomous Economic Agents on an Ethereum-compatible ledger. In Phase 1, each judge records a one-way hash of its score and a secret salt before candidate identities are revealed. In Phase 2, the identity and raw score are disclosed and verified on-chain. This creates a tamper-evident audit trail that separates blind evaluation from post-hoc claims and reduces the verification burden on independent researchers and leaderboard operators.

In the context of geodata, existing Large Language Models have often been studied in a homogeneous setting, which has considerably limited insights into their generalization capabilities. In this paper, we present \benchName, a comprehensive benchmark for probing LLMs on geo-related tasks. We leverage a careful selection of twelve publicly available datasets from diverse geo-related tasks and domains, and evaluate a set of LLMs on geo-spatial and temporal understanding using our benchmark. Our results show that reasoning and size have a strong impact on overall performance. GeoBenchLLM is publicly available at https://github.com/Rfr2003/GeoBenchLLM.

Test data from public benchmarks inevitably leaks into pretraining corpora, inflating evaluation scores once memorized. \textbf{Contamination mitigation evaluation} intervenes in the decoding process to suppress memorization and restore a contaminated model's genuine capability, but its prevailing metric, the \textbf{G-AP} (\textbf{G}ap of \textbf{A}ggregate \textbf{P}erformance), is flawed. Discrete correct/incorrect readouts cannot characterize per-question performance, averaging before differencing lets over- and under-suppression cancel out, and uniform per-question weighting invites strategies to push solve probabilities onto the clean model's high-frequency values. We propose \textbf{SA-PPG} (\textbf{S}tratified \textbf{A}ggregate of \textbf{P}er-question \textbf{P}robability \textbf{G}aps): estimate each question's solve probability by sampling, difference it against the clean model per question, and aggregate within groups defined by the clean model's solve probability. Existing mitigation strategies first estimate where contamination lies and then operate on the estimate, so they are only as correct as the estimate. \textbf{RailCap} instead judges contamination during generation: whenever a sample falls back onto the greedy trajectory, the next trajectory token is capped to the runner-up, accumulating suppression until the response distribution becomes sufficiently dispersed. Across multiple contaminated models and benchmarks, SA-PPG reveals that prior strategies' restoration is substantially overestimated, while RailCap attains the lowest SA-PPG.

Confirmed oncogenic microbes contribute significantly to cancer burden. Identifying novel microbial oncogenicity could yield strategies that will reduce disease burdens. However, relevant evidence is dispersed and infeasible for humans to comprehensively synthesize. LLMs may enable scalable, expert-level systematic evidence synthesis to identify microbe-cancer pairs; however, such capabilities have not yet been demonstrated. Domain experts were recruited to create a dataset to benchmark LLM performance (Gemini 2.5 Pro, Gemini 2.5 Flash, GPT-5, GPT-5 Nano) on 24 research papers using MMTV-LV and breast cancer as a case study. We devised a structured template for evidence extraction and appraisal, consisting of MCQ, Likert-scale, multi-select, and free-text question types (77 items across 24 papers). Agreement between (1) experts and (2) experts and each LLM was determined per question instance using novel metrics. LLMs were assessed by comparing inter-expert and expert-LLM agreement distributions to determine whether LLMs behaved as additional experts by increasing or maintaining inter-expert agreement. Free-text responses were further evaluated qualitatively. Across all question types, LLM responses aligned closely with experts, with GPT-5 and GPT-5 Nano achieving score distributions indistinguishable from experts. Gemini models behaved similarly but were significantly more lenient in applying microbial oncogenesis criteria. Hallucinations were rare. Methodological appraisal and identification of contradictions within full-texts were the most persistent LLM vulnerabilities. GPT-5 and GPT-5 Nano were indistinguishable from experts on structured domain research paper evaluation tasks. This supports use of LLMs for automated systematic evidence synthesis. However, methodological appraisal tasks and contradiction identification in full-texts remain weaknesses requiring strengthening.

Generative models are often evaluated through singular artifacts, whereas human creativity typically emerges through iterative generation, appraisal, and refinement. This pilot study examines whether iterative search improves LLM creativity by adapting FunSearch to recipe generation for the 2024 Pillsbury Bake-Off and evaluating outputs against human benchmarks using TTCT-based LLM evaluation. Across two experiments, we test iteration count, generator temperature, and in-loop selection-scorer model size. Results show that iterative generation-selection can produce recipes with creativity scores comparable to human benchmarks, but additional iterations alone do not improve creativity. The in-loop evaluator matters most: a smaller selection scorer yields significantly higher scores across most TTCT dimensions, while temperature has limited effects except for originality. These findings suggest that evaluator design is a first-order design variable in subjective creative search.

Although remarkable capabilities have been demonstrated by Large Language Models (LLMs) across scientific domains, feedback controller design remains underexplored. Existing benchmarks focus mainly on linear single-Degree-of-Freedom (DoF) systems and large API-hosted models, leaving performance on complex controller-design tasks and feasibility for edge deployment unclear. To address these limitations, we introduce the Complex Dynamics-to-Control Benchmark for Large Language Models (CoDyControlBench), comprising 132 system configurations across five evaluation dimensions: number of DoF, system type, coupling level, damping regime, and controller type. Six state-of-the-art LLMs were evaluated over three independent runs, including three commercial models (GPT, Gemini, and Claude) and three open-source models (GLM, DeepSeek, and Qwen). GPT achieved the highest design success rate at 94.8%, whereas Qwen showed the lowest rate at 50.0%. Across the benchmark dimensions, DoF and controller type exhibited the largest model-averaged variations in design success, with success-rate ranges of 36.3% and 17.6%, respectively, exceeding those associated with system type, coupling level, and damping regime. Comparison of GPT and Qwen showed that their performance gap arose mainly from the control-design knowledge, particularly gain selection and the use of transient-limiting mechanisms. For edge deployment, a specialized 1.5B-parameter model was developed through reasoning distillation. The reasoning-distilled model outperformed the answer-distilled and base model on CoDyControlBench, maintained stable performance across 1-6 DoFs, and achieved successful traget tracking in all three physical trials on a pneumatic-artificial-muscle-driven robotic arm. These results establish a benchmark baseline and highlight the potential of lightweight, edge-deployable controller-design models.

Current evaluations do not isolate whether text-only language models can originate visual concepts before image generation. Fluent visual prose can hide visual-plan failures: an answer may appear creative while repeating familiar visual clichés or failing to specify a renderable scene. We define Visual Creative Ideation (VCI) as the ability to produce textual visual plans that are useful, expressive, and population-novel, and introduce Ekphrasis, a 400-task benchmark spanning Abstraction, Combination, Transformation, and Adaptation. Ekphrasis scores anonymized pairwise comparisons with dimension-specific checklists, aggregates preferences with Bradley-Terry models, and uses Typed Idea Graphs to convert task-specific population clichés into novelty references. Across 14 language models, VCI separates usefulness, expressiveness, and novelty rather than reducing to fluency: strong models achieve similar overall scores through different profiles, and useful plans can remain visually clichéd. A cross-modal grounding study further shows that text-level VCI ordering largely survives faithful rendering and blind image-level preference judgment, supporting Ekphrasis as a measure of visual ideation beyond prose quality.

Large language models (LLMs) are trained on corpora that contain expressions of human judgment about films, books, music, and more. Yet whether LLMs systematically reproduce evaluative hierarchies remains unclear. Prior research on cultural bias in LLMs suggests competing expectations: models may mirror the popularity signals of internet texts, or may reproduce forms of prestige embedded in critical discourse. We probe this question through a study of film evaluations with eight models from four families (Anthropic, OpenAI, Alibaba, and Mistral), using a 200-film benchmark partitioned into critically acclaimed, commercially successful, and dual-legitimacy (critical acclaim + commercial success) films. Across 20,000 pairwise forced-choice comparisons per model analyzed with Bradley--Terry estimation, we observe a consistent critical acclaim orientation with all models: critically acclaimed yet commercially obscure films are selected over commercially successful yet critically unrecognized ones. This pattern grows with model scale within each family. In addition, nested OLS regression analyses show that evaluative orientation, public visibility, and popular reception distinctly help explain preferences. Adjusting for public visibility reverses the models' preference for dual-legitimacy films over critical acclaim-only films, while additionally accounting for popular reception attenuates much of the disadvantage of films with commercial success only. Finally, evaluative and recommendation-oriented prompt framings produce divergent rankings, suggesting that critical acclaim orientation may manifest indirectly in real-world LLM deployments.

LLM judge panels are a standard evaluation tool, but prior work reports highly correlated panel errors: nine judges provide roughly the effective information of two independent ones, and aggregation closes only a small fraction of the gap. A natural remedy--a signal from a different evidence source, e.g., executing a test suite--produced no distinguishable change in the panel's effective-vote count at scale (-0.04, 95% CI [-0.10, +0.02]). Aggregate dependence and conditional decision utility are different questions. Elementary majority arithmetic fixes the affected set for single-ballot substitution: only decisions with a one-vote margin can change. The empirical question is whether panel error rates rise and useful substitutions concentrate there. They do: the entire accuracy gain concentrates on these pivotal queries, where it is large (+10.4 to +23.3 percentage points across three headline configurations), and is exactly zero elsewhere. We confirm the pattern across three code benchmarks and four panel sizes (a 9-judge extension and 56 dependent subsampling checks, gain +6.5 to +16.1 percentage points). On HumanEval+/MBPP+, a majority-side replacement rule raises overall accuracy from 82.44% to 85.62% while invoking the signal on 16.2% of queries; signal-only remains stronger at 87.60%. Thus population-level dependence diagnostics and margin-stratified utility are complementary, and the affected-set characterization yields a call-reduction rule for any specified single-ballot substitution policy.

No single large language model (LLM) is optimal across all queries and budget constraints, making model routing essential for cost-effective deployment. Existing routers adopt diverse formulations and implementations, making fair comparison and extension difficult. We present a unified formulation of LLM routing as a sequential decision process characterized by five components: context encoders, model encoders, scoring functions, decision rules, and learning signals, covering single-turn, multi-turn, and personalized routing. Based on this formulation, we develop an automated pipeline for constructing routing supervision and evaluating routers jointly on response quality and inference cost. The resulting benchmark, xRouteBench, spans generic LLM, memory-augmented, vision, time-series, and personalized routing tasks. We further introduce LLMRouter, an open-source modular infrastructure with more than 16 representative routers. Our empirical study shows that learned routers outperform the strongest fixed-model baseline by 14.6% relatively, lightweight routers become more competitive under tight cost constraints, and user-conditioned routing consistently improves personalization.

Scientific and technical writing depends on markup sources that must compile: LaTeX, Typst, and Markdown pipelines fail on missing delimiters, mismatched environments, broken imports, or package conflicts. Existing document-repair evaluations inject faults with ad-hoc edits that lack an empirical fault model. We present TeXFix-Bench, a multi-format benchmark for LLM-based full-source document repair grounded in a mined fault taxonomy. A Grounded-Theory study of localized hard-crash LaTeX faults from TeX Stack Exchange, GitHub commits, and package documentation (168 verified faults, dual open coding at $κ$=0.34) yields an 18-category taxonomy instantiated as DocMut: 48 AST-aware operators across three formats. A three-model cross-benchmark shows DocMut faults are 5.6-9.2 pp harder to repair than pattern-based mutations on the same seeds, and a real-error case study (88 mined human crashes, 67.0% repair success) brackets both synthetic sets from below. We construct 10,437 instances from 743 openly licensed seeds and evaluate seven LLMs under a fixed zero-shot protocol with provider-pinned routing, collecting 48,651 attempts at about USD 200 total inference cost. A complete 6,613-instance x 7-model balanced matrix confirms all rankings. A pinned engine gate yields a 27.5-point intention-to-treat compile spread (56.7-84.2%). Typst is markedly harder than LaTeX and Markdown. A restoration oracle over 28,129 compiling repairs shows that 13.6-18.5% of compiling repairs materially alter document text, and restoration rank diverges from compile rank: the model with the lowest compile rate restores content best among its successes. Compile success alone overstates repair quality. We release the taxonomy, DocMut, and all campaign artifacts.

LLMs are increasingly being applied to tasks involving institutional and political texts, but existing benchmarks evaluate them on isolated documents or single tasks. In realpolitik, negotiations are longitudinal data, where participating parties can align or argue over multiple iterations and each turn is an outcome of the previous turns, hence, understanding one turn requires tracking everything before it. We introduce TradeVerse, a benchmark built from the World Trade Organisation (WTO) specific trade concerns, where member states challenge one another and exchange arguments over multiple rounds, sometimes for years. We, in TradeVerse, reconstruct minutes of $1170$ meetings, spanning across 5 groups and $89$ product groups and define three tasks: first, the system has to analyze the longitudinal meeting records and predict the harmonized system codes (HS chapters) of the products under discussion in the particular meeting, second, we examine whether the system, upon analyzing the anonymized content of the meeting, can guess the name of the responding country and third, we ask the system to play the role of the responding country and provide the statement for the very last round. All labels are recovered directly from the proceedings, requiring no manual annotation. Our experiments highlight the challenges these tasks pose for current LLMs. To the best of our knowledge, TradeVerseis the first benchmark to investigate potential of LLMs in understanding longitudinal political trade negotiations.

False-presupposition QA (FPQA) tests LLMs on their ability to identify false presuppositions in questions and abstain or correct them rather than reinforcing false assumptions. The common approach reduces the task to prompting LLMs to extract presuppositions and fact checking each presupposition. While the performance on dedicated benchmarks keeps improving, evaluation largely focuses on questions with false presuppositions (FPQs) while ignoring the performance on ``normal'' questions (TPQs). Since many benchmarks over-represent FPQs compared to their natural occurrence, the result is that performance on these benchmarks doesn't reflect real-world QA performance. Through extensive experiments across various model families, sizes, and benchmarks, we show that methods that perform better on FPQs tend to perform worse on TPQs. Our analysis reveals this is the result of weak fact checking modules that reject also true presuppositions. We hope our findings will help guide future work toward FPQA methods that generalize well to realistic settings.

Tool use transforms LLMs into agents that act beyond their training data, and for code-capable models, programmatic tool calling extends this further by replacing rigid JSON calls with scripts that chain and parallelize naturally. However, a systematic evaluation of tools as code on an established benchmark across current and prior model generations under real-world task conditions has not been conducted. In this work, we empirically compare programmatic tool calling (PTC) to native JSON tool calling across 14 language models on BFCL v4. In the programmatic tool calling paradigm, tools are exposed as typed Python stubs that the model invokes through code, with execution and results handled in a single agent turn. Programmatic tool calling matches or exceeds native JSON tool calling in 11 of 14 models on BFCL v4, with the GPT-5.6 family achieving a 10.6% improvement over the JSON tool calling baseline. Further, it matches or outperforms baseline in 13 of 14 models under parallel fan-out, and holds stable under context rot conditions where baseline degrades 2.3% on average. Our results demonstrate that programmatic tool calling is a viable and robust alternative to JSON tool calling, with performance tracking model capability across release generations.

Large language models (LLMs) increasingly support complex professional tasks, yet their capabilities in rule-intensive document review remain insufficiently evaluated. National standard documents, such as China GB/T standards, offer a representative testbed: they are lengthy, highly structured, and governed by explicit rules for scope, terminology, normative wording, and cross-section consistency. Existing benchmarks focus on domain knowledge and question answering, largely overlooking intrinsic quality review for professional documents. Such reviews rely heavily on human experts, making them costly and difficult to scale. To bridge this gap, we introduce GB/T-Bench, the first benchmark for the structured review of national standard documents. Its GB/T Review Taxonomy is a hierarchical schema covering document structure, scope alignment, normative modality, terminology consistency, and normative references, with 25 diagnosable error types. A controllable counterexample generation mechanism combines deterministic rules and constrained LLM rewriting to process 488 documents into 7,306 traceable review error instances for evaluation. We also develop a diagnosis-oriented evaluation protocol requiring exact matches on error location, review dimension, and error type, plus document-level coverage metrics. We further propose GB/T-Reviewer, a multi-agent framework that converts review knowledge into specialized skills and coordinates global inspection, targeted diagnosis, rule scanning, and result verification. Experiments with 14 mainstream LLMs reveal a substantial human-LLM gap: the strongest model achieves only 0.3280 CMCS versus 0.6640 for experts. GB/T-Reviewer raises the best CMCS to 0.5094, showing the value of structured skill coordination for rule-intensive document review. This work paves the way for trustworthy AI in standardization and other high-stakes document domains.

Large language model (LLM) benchmark evaluations are routinely used to support claims about model safety, reliability, and deployment readiness. Yet most evaluations rely on a single access modality (model APIs), perform a single run per prompt, and report accuracy as the primary outcome metric, without accounting for conditions such as web search that may have effects on model behavior in deployment. We audit these assumptions for one of the most widely-used LLMs, comparing two modalities, ChatGPT's chat UI and OpenAI's API, with and without web search enabled. We use a stratified total sample of 401 prompts from two popular benchmarks, BBQ and SafetyBench, collecting 4,812 total responses across three repeated runs per prompt. Beyond standard performance measures, we evaluate model output dimensions including response consistency, response text similarity, citation grounding, and abstention behavior. For instance, chat UI responses were less accurate than API responses on both benchmarks with search disabled. Enabling web search reduced accuracy by up to 8 percentage points, and even reversed the direction of modality performance trends for one benchmark. Repeated runs of the same prompt produced inconsistent responses in up to 21% of prompts. The two modalities also grounded answers in different citations, and abstention behavior was also inconsistent across both modalities. These results illustrate that, even within a model family, reporting only simple accuracy metrics can obscure important forms of model behavioral variation relevant to AI safety assessments. We argue that AI safety evaluations should systematically account for modality, multi-run consistency, search conditions, and response-level behaviors to better reflect how deployed AI systems behave in practice.

Large language models (LLMs) have been shown to exhibit strong Python preferences when generating project-level code, but there is currently no systematic way to measure this behaviour across new models. To bridge this gap, we introduce LangChoiceBench, a project-level code-generation benchmark for measuring Python preference, recommendation-implementation consistency, and language diversity. LangChoiceBench covers 28 projects across seven software areas where Python is often a poor default. We evaluate 25 diverse LLMs and find that Python remains heavily over-selected, recommendation-implementation consistency is low, and smaller open-weight models generally show stronger Python preference and lower language diversity. We further analyse 9,826 reasoning traces and find that most Python choices are automatic or driven primarily by ease, rather than explicit consideration of project requirements. In a smaller but important set of cases, models fabricate contextual support for choosing Python - a failure mode we call phantom evidence - or produce code that contradicts the language selected in their own reasoning.

Epitopes determine where antibodies bind antigens and shape downstream therapeutic properties such as functional blockade and escape resistance, making epitope understanding central to antibody drug discovery. Although large language models (LLMs) have shown strong biomedical reasoning ability, it remains unclear whether they can infer epitope information directly from antigen and antibody sequences. Existing epitope resources typically focus on isolated prediction tasks or rely on specialized structural settings, while general protein benchmarks do not evaluate epitope-centered decisions across the antibody development workflow. To address this gap, we introduce EpiBench, a closed-book, sequence-based, and automatically scorable benchmark for evaluating epitope reasoning in LLMs. EpiBench contains 1,609 curated samples grounded in structural antibody--antigen contacts, curated functional B-cell assays, and deep mutational scanning escape measurements. It covers five connected tasks: targetable region discovery, antibody-conditioned epitope identification, epitope binning, functional epitope assessment, and antibody escape assessment, with controlled sampling to reduce shortcut-based evaluation artifacts. We evaluate nine general-purpose LLMs and analyze their behavior through task-specific baselines, antigen length stratification, explicit-reasoning comparison, and failure-mode inspection. The results show that current LLMs capture partial epitope-related signals but remain limited in antibody-specific sequence grounding, long-context residue localization, and biologically grounded reasoning. Therefore, EpiBench provides a diagnostic testbed for measuring and improving sequence-aware biomedical LLMs toward reliable LLM-assisted antibody discovery.

Large Language Models (LLMs) are often used as evaluators of text quality, known as LLM-as-a-Judge, which can outperform conventional automatic evaluation metrics that rely on reference texts. However, LLM evaluators tend to generate particular scores regardless of the context of the evaluated text, which is known as scoring bias. This study proposes a novel method to mitigate this scoring bias. An LLM is instructed to randomly generate number tokens, and the latent numerical bias of the LLM is identified by measuring the deviation of the observed distribution of numbers from the uniform distribution. A definition of a downstream task, for which an LLM evaluator is used, is added to the prompts for random number generation to measure task-specific latent number bias. In the evaluation by an LLM, the token generation probabilities for a given input are rectified considering the LLM's latent number bias. Results of the experiment on four different tasks, evaluation of LLM alignment, evaluation of summarization, Semantic Textual Similarity, and Semantic Textual Relatedness, demonstrate that our proposed method outperforms the baselines, including an LLM without debiasing and previous calibration methods. In addition, it is confirmed that scoring bias varies across LLMs, tasks, and score ranges, indicating the importance of measuring latent number bias as the case may be.

Anaphors are expressions that refer to other expressions, called antecedents. The process of connecting the two is called resolution. Cognitive science has identified multiple factors that affect the speed and success of anaphor resolution, including discourse structure, situation-model properties, and semantic factors. Here, we investigate whether these factors also affect anaphor resolution in five Large Language Models (LLMs) with open weights: GPT-2-XL, Llama-3.1-8B, Pythia-12B, Mistral-7B, and Mistral-24B. To model processing difficulty, we adopt the standard linking hypothesis that relates human reading times to model surprisal at the anaphor. As a second behavioral measure, we compare model accuracy to human accuracy on comprehension questions probing the antecedents of anaphors. The results show selective cognitive alignment: some LLMs exhibit human-like sensitivity to discourse prominence and distance-based factors in anaphor resolution, while showing weaker or absent sensitivity to semantic interference effects. These findings delimit the conditions under which LLMs approximate human anaphor resolution.

When a large language model (LLM) writes Harry Potter fanfiction, it reliably produces fundamental elements of the Hogwarts universe, such as recognizable places and characters. Human-written Harry Potter fanfictions, however, typically include these fundamentals and much more, incorporating stylistically irregular content and relationship-diverse plotlines. This gap between LLM and human writing has been noted across a variety of domains. LLMs tend to produce "average" writing, while human writing contains more diverse content that covers a broader distribution. Existing work has shown the existence of this distributional "gap", but no work has proposed a systematic way to measure it. Our paper proposes a human-grounded framework that uses the empirical distribution of human writing on a topic to measure the distributional breadth of LLM-generated content on that same topic. We propose two metrics, LLM Coverage (LLM-Cov) and In-Boundary Rate (IBR), that separate the plausibility of LLM content from its distributional breadth. Across ideation and narrative tasks, we find that current LLMs produce plausible but narrow content that concentrates near the center of the human response space. Our framework can enable researchers to better assess the distributional breadth of LLM-authored content, which we term its "cultural reach".

Large language models (LLMs) can generate text that resembles a mathematical proof, but resemblance does not establish correctness. A formal proof checker verifies whether each proof step follows established logical rules. Coq bases its rules on the Calculus of Inductive Constructions, a logical framework that defines which proof steps the system may accept. This pilot study evaluated six open-weight LLMs on the same 100 theorems from CoqStoq, a benchmark derived from real Coq projects. Each LLM received one attempt per theorem with the temperature set to 0, and Coq checked every proposed proof in the theorem's original project environment. We counted a proof as successful only if the Coq kernel accepted it. Gemma 4 verified 12 of 100 theorems, Llama 3.3 verified 8, and DeepSeek Coder V2 Lite verified 1. Qwen 3.5, Mistral Small 3.1, and GPT-OSS verified none. The 21 successful model-theorem results covered 15 distinct theorems, 11 of which were not solved by a baseline of standard Coq tactics. All verified theorems had short or medium human-written reference proofs; no model verified a theorem with a long reference proof. Because the proof-length analysis was exploratory, this pattern does not establish that proof length caused the difference. For the three models with at least one success, the total generation cost per verified proof ranged from 741 to 36,193 output tokens, 14.9 to 178.0 seconds, and 0.0167 to 0.2000 aggregate GPU hours. We could not calculate these ratios for models with no verified proofs. Across 600 attempts, the models produced 21 kernel-verified proofs, giving an overall success rate of 3.5%. The study reports descriptive differences among the models but does not statistically test whether one model outperforms another. Therefore, the results do not establish a universal ranking of the six models.

Large language models (LLMs) are increasingly used as AI tutors, but a correct answer is not always a pedagogically appropriate one. In classroom learning, effective help depends not only on correctness, but also on whether a response matches the learner's current foundation, the course sequence, and the timing of concept introduction. Existing evaluations focus mainly on answer quality, leaving this instructional fit under-measured. We present the Pedagogical Suitability Index (PSI), a composite metric of six theory-informed sub-scores that evaluates how well LLM-generated tutoring responses align with learner readiness and curricular progression, and we further use PSI as a structured feedback signal for response improvement. We evaluate four LLM tutors (ChatGPT, Gemini, Gemma4, and Qwen3) across 240 scenario-based evaluations using paired standard and defective prompts, then apply a PSI-guided regeneration protocol to 62 weak-performing cases. Baseline differences across the four tested models were modest overall (PSI range: 0.557 to 0.638), and open-weight and closed models did not exhibit a clear separation in pedagogical fit. Under the tested prompt perturbations, overall PSI remained largely stable (Delta = -0.002), though sub-score trade-offs emerged. More importantly, PSI-guided feedback substantially improved weak-performing cases: 51 of 62 cases improved (82.3%). Focused manual evaluation of the 62 PSI-selected weak cases provides initial evidence that the identified weaknesses are instructionally meaningful and that many PSI-guided regenerations correspond to human-judged improvement. These results suggest that learner- and curriculum-aware alignment may matter more for effective tutoring than model category alone, and that such alignment is both measurable and improvable.

LLM judges are often asked to extract criteria and evidence before choosing between candidate answers. This workflow assumes that the intermediate record preserves the information needed for a later verdict. For reasoning-capable models, visible field order does not reveal internal decision order, so we test an observable alternative: persist the evidence in one call and make it the exclusive input to the next. Across 24,000 judgments over HelpSteer3, FeedbackQA, and CoVal, we compare standard pairwise judging, structured one-call judging, two-call evidence locking, and three-call pointwise locking with Claude Sonnet 4.5 and GPT-5. Evidence locking reduces agreement with released human preferences by 4 to 6 percentage points and increases answer-order inconsistency by 8 to 10 points relative to structured one-call judging. Pointwise locking is also harmful, while structured evidence elicitation remains close to standard judging. The result holds for both judges and all three datasets. Persisted evidence can support auditability, but it should not replace the source answers at decision time.

Giving an LLM judge more compute does not necessarily make it check more requirements. When one call must return many verdicts, some decisions become weakly grounded in the evidence, even when that call receives the same token or tool budget as a panel of separate calls. Across expert-graded research replications, legal work, and clinical-trial assessments, agreement with experts falls as the number of verdicts per call grows. We identify sharding as the intervention that mitigates this failure in model-based oversight. Sharding partitions the requirements into smaller groups, assigns each group to a separate call, and aggregates the verdicts. Against a single call with the panel's full budget, sharding improves agreement while holding the model, evidence, total budget, and per-decision budget fixed. Overall, we find that a sharded weaker judge can outperform a more capable holistic judge and match that judge even when the latter receives the panel's full budget. Additionally, we find that sharding exhibits robustness against adversaries. A best-of-N adversary can hold the underlying work fixed, vary only its presentation, and increase an overloaded judge's acceptance of genuinely unmet criteria severalfold. Wherever sharding reduces baseline error, it removes this adversarial advantage, keeping over-acceptance low even as the adversary's search widens. Sharding does not address attacks that persuade the judge separately on each criterion rather than exploiting overload. In that setting, we find that debate-style opposition on top of sharding withstands such adaptive re-optimization.