LLM Evaluation

Corporate bankruptcy prediction is a high-stakes financial task characterized by severe class imbalance and multi-horizon forecasting demands. Public datasets supporting it remain scarce and small: widely used free benchmarks contain between 6,000 and 80,000 company-year observations, while larger resources are behind subscription paywalls. To address this gap, we introduce V4FinBench, a benchmark of over one million company-year records from the Visegràd Group (V4) economies (2006-2021), with 131 financial and non-financial features, six prediction horizons, and a composite distress criterion jointly capturing solvency, profitability, and liquidity deterioration. V4FinBench is designed to support the evaluation of tabular and foundation-model methods under realistic class imbalance, with positive rates between 0.19% and 0.36%. We provide reference evaluations of standard tabular baselines, finetuned TabPFN, and QLoRA-finetuned Llama-3-8B. With imbalance-aware finetuning, TabPFN matches or exceeds gradient boosting at longer time horizons on both $F_1$-score and ROC-AUC. In contrast, Llama-3-8B trails gradient boosting on ROC-AUC at every horizon and is generally weaker on $F_1$-score, with the gap widening sharply beyond the immediate horizon. In an external evaluation on the American Bankruptcy Dataset, the V4FinBench-finetuned TabPFN checkpoint improves over vanilla TabPFN, suggesting that adaptation captures transferable financial-distress structure rather than only V4-specific patterns. V4FinBench is publicly released to support further evaluation and development of prediction methods on realistic financial data.

Recent advances in machine learning and large-scale biological data collections have revived the prospect of building a virtual cell, a computational model of cellular behavior that could accelerate biological discovery. One of the most compelling promises of this vision is the ability to perform in silico phenotypic screens, in which a model predicts the effects of cellular perturbations in unseen biological contexts. This task combines heterogeneous textual inputs with diverse phenotypic outputs, making it particularly well-suited to LLMs and agentic systems. Yet, no standard benchmark currently exists for this task, as existing efforts focus on narrower molecular readouts that are only indirectly aligned with the phenotypic endpoints driving many real-world drug discovery workflows. In this work, we present AssayBench, a benchmark for phenotypic screen prediction, built from 1,920 publicly available CRISPR screens spanning five broad classes of cellular phenotypes. We formulate the screen prediction task as a gene rank prediction for each screen and introduce the adjusted nDCG, a continuous metric for comparing performance across heterogeneous assays. Our extensive evaluation shows that existing methods remain far from empirically estimated performance ceilings and zero-shot generalist LLMs outperform biology-specific LLMs and trainable baselines. Optimization techniques such as fine-tuning, ensembling, and prompt optimization can further improve LLM performance on this task. Overall, AssayBench offers a practical testbed for measuring progress toward in silico phenotypic screening and, more broadly, virtual cell models.

We initiate the study of the Generalized Turing Test (GTT), a formal generalization of Turing's imitation game from humans to arbitrary interactive agents. For agents $A$ and $B$, $A$ passes the GTT against $B$ if an instance of $B$, acting as a distinguisher, cannot reliably distinguish an $A$ instructed to imitate $B$ from another instance of $B$; if so, we write $A \geq B$. We study the theoretical and empirical consequences of this idea. On the theory side, we prove sufficient conditions under which this "Turing Comparator" is transitive. We introduce natural variants with querying (the imitator can first interact with a specimen of the target), a Universal Turing Test with arbitrary distinguishers and targets, and complexity-theoretic variants that control interaction length. As a proof of concept, we evaluate the GTT and its variants across nine large language models. Remarkably, Turing Scores recover a clear model stratification consistent with standard external benchmarks despite being derived entirely from pairwise imitation games. Transcript analysis reveals that models use both stylistic signatures and substantive STEM and logic-based probes. Together, these results suggest indistinguishability could provide a meaningful signal for comparing agents, yielding an inherently adaptive form of evaluation that does not rely on fixed benchmarks.

We introduce an automatically generated benchmark for predicting hidden text in technical papers. A paper supplies visible context $X$ and a hidden continuation $Y$; the evaluated model writes an auxiliary forecast string $Z$, and a separate scorer assigns next-token probability to $Y$ both with and without conditioning on $Z$. This gives a label-free test of whether $Z$ transmits information about the continuation, compared against controls where $Z$ is recent context rather than a forecast. Our main testbed is equation-suffix prediction: the predictor sees context and the first part of a displayed equation, then forecasts the rest. The task mixes surface-level arXiv/TeX text modeling with reasoning-sensitive inference; the suffix is one of many roughly equivalent continuations, so the benchmark is read statistically rather than item-by-item. On 1363 equation continuations from 138 recent physics and mathematics papers, forecasts from GPT-5.5, Opus 4.7, and GPT-5.4 nano all improve clipped likelihood over the context control under both Qwen3-8B and Kimi K2.6 scorers, distinguishing model families and reasoning-effort settings without human labels. To emulate shortcuts where $Z$ further primes the scorer rather than making a useful forecast, we also fine-tune the scorer on context-only prompts and apply it to held-out papers as a stronger control. GPT-5.5 forecasts still beat this fine-tuned control; GPT-5.4 nano forecasts do not. Longer prose/TeX continuations show positive but noisier lift over controls, concentrated near the beginning of the target. These results support cross-model likelihood scoring as a static benchmark and as a setup for probing shortcut vulnerabilities before reinforcement learning or model-selection optimization is applied.

Large language models (LLMs) can convincingly imitate human writing styles, yet it remains unclear how much stylistic information is encoded in embeddings from any language model and retained after LLM rewriting. We investigate these questions in French, using a controlled literary dataset to quantify the effect of stylistic variation via changes in embedding dispersion. We observe that embeddings reliably capture authorial stylistic features and that these signals persist after rewriting, while also exhibiting LLM-specific patterns. These analytical results offer promising directions for authorship imitation detection in the era of language models.

As artificial intelligence advances, models are not improving uniformly. Instead, progress unfolds in a jagged fashion, with capabilities growing unevenly across tasks, domains, and model scales. In this work, we examine this dynamic jaggedness through the lens of scientific idea generation. We introduce SciAidanBench, a benchmark of open-ended scientific questions designed to measure the scientific creativity of large language models (LLMs). Given a scientific question, models are asked to generate as many unique and coherent ideas as possible, with the total number of valid responses serving as a proxy for creative potential. Evaluating 19 base models across 8 providers (30 total variants including reasoning versions), we find that jaggedness manifests both across models and within models. First, in a cross-task comparison between general and scientific creativity, improvements in general creativity do not translate uniformly to scientific creativity, revealing divergent capability profiles across models. Second, at the prompt level, stronger models do not improve uniformly; instead, they exhibit high variability, with bursts of creativity on some questions and limited performance on others. Third, at the domain level, individual models display uneven strengths across scientific subfields, reflecting fragmented internal capability profiles. Finally, we show that this jaggedness can be harnessed. We explore mechanisms of inference-time compute, knowledge pooling, and brainstorming to combine models effectively and construct meta-model ensembles that outperform any single model. Our results position jaggedness not as a limitation, but as a resource, a structural feature of AI progress that, when understood and leveraged, can amplify LLM-driven scientific creativity.

This paper investigates the effectiveness of large language models (LLMs) in answering questions over datasets. We examine their performance in two scenarios: (a) directly answering questions given a dataset file as input, and (b) generating SQL queries to answer questions given the schema of a relational database. We also evaluate the impact of different prompting strategies on model performance. The study includes both state-of-the-art LLMs and smaller language models that require fewer resources and operate at lower computational and financial cost. Experiments are conducted on two datasets containing questions of varying difficulty. The results demonstrate the strong performance of large LLMs, while highlighting the limitations of smaller, more cost-efficient models. These findings contribute to a better understanding of how LLMs can be utilized in data analytics tasks and their associated limitations.

Large language models can score well on named game-theory benchmarks while failing on the same strategic computation once semantic cues are removed. We show this gap with procedurally generated zero-sum matrix games: a model that recognizes familiar games drops to 34%, 18%, and 2% success on anonymous $2{\times}2$, $3{\times}3$, and $5{\times}5$ payoff matrices. The benchmark separates semantic recall, learned approximate Nash computation, and an output-interface bottleneck that limits scale. Training only on $2{\times}2$ and $3{\times}3$ games, supervised fine-tuning raises unseen $5{\times}5$--$7{\times}7$ success from 2% to 61%, while exploitability-reward training averages 37% with high seed variance. We prove that the exploitability residual is $2$-Lipschitz in payoff perturbations, unlike discontinuous vertex-returning LP equilibrium selectors, explaining why residual training can transfer under payoff shifts even when formatting instability limits mean performance. A dominated-action padding experiment provides causal evidence: trained models solve $3{\times}3$ games embedded in much larger matrices, while random-padded controls fail and dense $12{\times}12$ games remain near failure. Procedural evaluation is therefore necessary for measuring strategic reasoning, and residual rewards expose a real but format-limited route to approximate equilibrium computation.

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.

Large language models (LLMs) have become capable mathematical problem-solvers, often producing correct proofs for challenging problems. However, correctness alone is not sufficient: mathematical proofs should also be clear, concise, insightful, and transferable to other problems. While this proof quality is subjective and depends on the reader and context, many of its components are concrete and broadly valued. In this work, we identify such components and introduce ProofRank, a benchmark curated from challenging mathematical competitions. ProofRank evaluates several scalable proxies of proof quality: (i) conciseness, measuring whether proofs avoid unnecessary steps; (ii) computational ease, measuring the extent to which a proof relies on tedious calculations; (iii) cognitive simplicity, measuring how accessible the used proof techniques are; (iv) diversity, measuring how varied a model's proofs for a single problem are; and (v) adaptivity, measuring whether a model can follow a specified proof technique. Across models, we find substantial differences in proof quality that are not captured by correctness-only benchmarks. We also observe significant trade-offs between proof-quality metrics and correctness, suggesting that future evaluations of mathematical reasoning should measure how useful LLM-generated proofs are.

High-quality source code documentation is vital yet often neglected, especially in critical domains like healthcare where reliability and maintainability are essential. We presented an AI powered framework that automates documentation generation from code and repositories using eight state of the art Large Language Models (LLMs), including GPT, Gemini, Qwen, and LLaMA variants. Built on the PocketFlow orchestration framework, the system applies modular pipelines and advanced prompt engineering to produce structured, context aware documentation. To ensure quality and guide model selection, we introduced a MultiLLMasJudges evaluation framework, where four independent LLMs assess outputs across nine criteria, such as Completeness, Clarity, and Faithfulness. Experiments conducted on an open-source medical physics library, demonstrated showed a 42% performance gap between top and bottom models. By combining diverse model outputs, optimized prompting, and rigorous evaluation, our approach enhances documentation quality and reduces manual effort, especially in safety critical healthcare software.

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.

Large language models (LLMs) are increasingly integrated into legal drafting and research workflows, where incorrect citations or fabricated precedents can cause serious professional harm. Existing legal benchmarks largely emphasize statutory reasoning, contract understanding, or general legal question answering, but they do not directly study a central common-law failure mode: when asked to provide case authorities without external grounding, models may return plausible-looking but incorrect citations or cases. We introduce LegalCiteBench, a benchmark for studying closed-book citation recovery, citation verification, and case matching in legal language models. LegalCiteBench contains approximately 24K evaluation instances constructed from 1,000 real U.S. judicial opinions from the Case Law Access Project. The benchmark covers five citation-centric tasks: citation retrieval, citation completion, citation error detection, case matching, and case verification and correction. Across 21 LLMs, exact citation recovery remains highly challenging in this closed-book setting: even the strongest models score below 7/100 on citation retrieval and completion. Within the evaluated models, scale and legal-domain pretraining provide limited gains and do not resolve this difficulty. Models also frequently provide concrete but incorrect or low-overlap authorities under our evaluation protocol, with Misleading Answer Rates (MAR) exceeding 94% for 20 of 21 evaluated models on retrieval-heavy tasks. A prompt-only abstention experiment shows that explicit uncertainty instructions reduce some confident fabrication but do not improve citation correctness. LegalCiteBench is intended as a diagnostic framework for studying authority generation failures, verification behavior, and abstention when external grounding is absent, incomplete, or bypassed.

Large language models (LLMs) require reliable evaluation from pre-training to test-time scaling, making evaluation a recurring rather than one-off cost. As model scales grow and target tasks increasingly demand expert annotators, both the compute and labeling costs needed for each evaluation rise rapidly. Active testing aims to alleviate this bottleneck by approximating the evaluation result from a small but informative subset of the evaluation pool. However, existing approaches primarily target classification and break down on generative tasks. We introduce a novel active testing algorithm tailored to generative tasks. Our method leverages semantic entropy from surrogate models to stratify the evaluation pool and then conducts approximate Neyman allocation based on signals extracted from these surrogates. Across multiple language and multimodal benchmarks and a range of surrogate-target model pairs, our method significantly improves on baselines and closely tracks Oracle-Neyman, delivering up to 28% MSE reduction over Uniform Sampling and an average of 22.9% budget savings.

Cell-type-specific marker genes are fundamental to plant biology, yet existing resources primarily rely on curated databases or high-throughput studies without explicitly modeling the supporting evidence found in scientific literature. We introduce PlantMarkerBench, a multi-species benchmark for evaluating literature-grounded plant marker evidence interpretation from full-text biological papers. PlantMarkerBench is constructed using a modular curation pipeline integrating large-scale literature retrieval, hybrid search, species-aware biological grounding, structured evidence extraction, and targeted human review. The benchmark spans four plant species -- Arabidopsis, maize, rice, and tomato -- and contains 5,550 sentence-level evidence instances annotated for marker-evidence validity, evidence type, and support strength. We define two benchmark tasks: determining whether a candidate sentence provides valid marker evidence for a gene-cell-type pair, and classifying the evidence into expression, localization, function, indirect, or negative categories. We benchmark diverse open-weight and closed-source language models across species and prompting strategies. Although frontier models achieve relatively strong performance on direct expression evidence, performance drops substantially on functional, indirect, and weak-support evidence, with evidence-type confusion emerging as a dominant failure mode. Open-weight models additionally exhibit elevated false-positive rates under ambiguous biological contexts. PlantMarkerBench provides a challenging and reproducible evaluation framework for literature-grounded biological evidence attribution and supports future research on trustworthy scientific information extraction and AI-assisted plant biology.

Deploying large language models in telecommunications requires more than task accuracy. In realistic workflows, a model may inherit partially completed reasoning from a prior step, an upstream agent, or its own earlier generation, and must continue that reasoning even when it is already going wrong. We introduce TeleResilienceBench, a benchmark that quantifies this capability, which we term reasoning resilience, across seven telecom sub-domains drawn from the GSMA Open-Telco LLM suite. Instances are constructed by collecting failures from a weak generator model, truncating the flawed reasoning trace at its midpoint, and asking a target model to continue and correct it. We propose the Correct Flip Rate (CFR) as a direct measure of successful recovery and evaluate eight models spanning the Qwen3.5, Gemma4, and Nemotron-3 families. Our results show that even the strongest model achieves a macro-average CFR of only 29.1%, and scale does not reliably improve resilience within families. Nemotron-3-nano 4b outperforms all Qwen3.5 variants including the 27b model and leads the auxiliary TeleMath numerical evaluation at 23.4% CR%, offering the best resilience-to-cost ratio in the set. A difficulty-stratified analysis further reveals that existing telecom benchmark difficulty labels reflect factual specificity rather than reasoning depth, suggesting that current evaluations measure knowledge coverage more than reasoning ability.

Reconciling platform revenue with user experience in LLM advertising motivates a data-centric foundation. We introduce NaiAD, the first comprehensive dataset for LLM-native advertising comprising 58,999 carefully constructed ad-embedded responses paired with user queries. NaiAD is organized around theoretically grounded evaluation metrics that separately and comprehensively capture user and commercial utility. To mitigate the dimensional collinearity of aligned LLMs, we propose a decoupled generation pipeline that produces structurally diverse samples, ranging from responses that explicitly disentangle stakeholder utilities to responses that are uniformly strong or weak across dimensions. We further provide score labels calibrated by a Variance-Calibrated Prediction-Powered Inference (VC-PPI) framework, aligning automated scoring with human annotations. Mechanistic analyses reveal that successful ad integration relies on reasoning paths that cluster into four distinct semantic strategies. Models leveraging NaiAD internalize these strategies to simultaneously improve user and commercial utility, while enabling independent control over these distinct objectives via in-context learning. Together, these results position NaiAD as a foundational infrastructure for developing future LLM-native ad systems.

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.

While extremely powerful and versatile at various tasks, the thinking capabilities of large language models (LLMs) are often put under scrutiny as they sometimes fail to solve problems that humans can systematically solve. However, recent literature focuses on breaking LLM reasoning with increasingly complex problems, and whether an LLM is robust in simple logical reasoning remains underexplored. This paper proposes Absurd World, a benchmarking framework, to test LLMs against altered realism, where scenarios are logically coherent, and humans can easily solve the tasks. Absurd World breaks a real-world model into symbols, actions, sequences, and events, which are automatically altered to create absurd worlds where the logic to solve the tasks remains the same. It evaluates a large collection of models with simple and advanced prompting techniques, and proves that it is an effective tool to determine LLMs' ability to think logically, ignoring the patterns learned from the real world. One can use this framework to extensively test an LLM against a real-world problem to verify whether the LLM's reasoning capability is robust against variations of the task.

When should a language model answer directly, sample and vote, or engage in multi-agent debate? Recent work shows voting often explains much of the gain attributed to debate, while selective-debate systems activate deliberation only on uncertain examples. We ask: under a matched ceiling on generated tokens (960 per example), how much per-example routing headroom exists, and how much is recoverable from cheap pre-deliberation signals? We evaluate greedy decoding, three-sample voting, and a two-agent critique-revise debate on MuSiQue and GSM8K using Llama 3.1 8B Instruct and Ministral 3 8B Instruct. On MuSiQue, an oracle selecting the correct protocol per example gains +14.0 and +13.7 pp over the best fixed one. The best fixed protocol is model- and dataset-dependent: each (model, dataset) cell has a different winner. This headroom is hard to recover from cheap ex-ante signals. A vote-entropy threshold is the only controller that directionally beats the best fixed protocol on both models (+1.3 and +1.7 pp), though individual paired-bootstrap CIs include zero. A joint analysis (meta-analysis +1.6 pp, p=0.125; Bayesian P(both>0)=0.59) is directionally consistent but not significant. Learned controllers (LR, GBT) do not outperform the threshold. The key finding is structural: vote entropy predicts where debate is safe, not where debate is needed. High entropy sharply reduces debate backfire, but 66% of debate-helpful examples (31/47) occur when voting is unanimous but wrong. A single-prompt self-critique probe on Llama flips the answer in 127/127 unanimous cases, yielding zero mutual information with the debate-helpful label; we cannot rule out a prompt-compliance artifact, but either interpretation disqualifies the probe as a router. Recovering the remaining headroom requires behavioral probes that avoid format-compliance confounds at the 8B scale.

Tool-integrated reasoning has emerged as a promising paradigm for enhancing large language models with external computation, retrieval, and execution capabilities. However, the field still lacks a high-quality and unified evaluation benchmark, and existing TIR evaluations remain limited in dataset quality, task diversity, diagnostic comprehensiveness, and evaluation efficiency. In this work, we introduce TIDE-Bench, a holistic and efficient benchmark for evaluating TIR methods, featuring three key advantages. First, it provides diverse task settings, combining widely used mathematical reasoning and knowledge-intensive QA tasks with two newly designed tasks, namely the tool-grounded experimental design task and the dynamic interactive task, to probe models' abilities in complex tool invocation and multi-tool coordination. Second, TIDE-Bench adopts a comprehensive yet task-aware evaluation protocol, jointly measuring final answer quality, process reliability, tool-use efficiency, and inference cost across heterogeneous task settings. Third, TIDE-Bench constructs high-quality and discriminative evaluation sets by filtering low-discrimination instances from existing datasets, substantially reducing evaluation cost while focusing on more challenging samples. Extensive experiments on multiple foundation models and TIR methods reveal persistent bottlenecks in tool grounding, offering insights for future TIR research.

Evaluating whether large language models (LLMs) can recover execution-relevant program structure, rather than only produce code that passes tests, remains an open problem. Existing code benchmarks emphasize test-passing outputs, from standalone programming tasks (HumanEval, MBPP, LiveCodeBench) to repository repair (SWE-Bench); this is useful, but offers limited diagnostic signal about which program semantics a model can recover from source. We introduce TraceEval, to our knowledge the first execution-verified, multi-language benchmark for code semantic reasoning: recovering a program's runtime call structure from source code. Unlike prior call-graph benchmarks that rely on static-tool output or hand-annotated ground truth, every positive edge in TraceEval is mechanically witnessed by validation execution, eliminating annotator disagreement and label noise for observed behavior. TraceEval consists of (i) 10,583 real-world programs (2,129 test, 8,454 train) extracted from 1,600+ open-source repositories across Python, JavaScript, and Java via an LLM-assisted harness-generation pipeline with tracer validation; and (ii) a reproducible pipeline that converts any open-source repository into new verified benchmark instances. We evaluate 10 LLMs at zero-shot on the held-out test split. The strongest model, Claude-Opus-4.6, reaches an average F1 of 72.9% across the three languages. To demonstrate the train split's utility as a supervision substrate, we fine-tune the Qwen2.5-Coder family on it: lifts of up to +55.6 F1 bring tuned Qwen2.5-Coder-32B to 71.2%, within 1.7 F1 of zero-shot Claude-Opus-4.6. We release the benchmark, pipeline, baselines, and a datasheet at https://github.com/yikun-li/TraceEva

Large language models now achieve high final-answer accuracy on mathematical reasoning benchmarks, but accuracy alone does not capture reasoning flexibility. We introduce a strategy-level evaluation framework instantiated on 80 AMC 10/12 and AIME problems with 217 AoPS-derived reference strategy families. Model outputs are annotated for strategy identity, validity, and correctness using dual-AI coding with human adjudication. Across four frontier models, we find a pronounced decoupling between answer accuracy and strategy diversity. Under a single-solution prompt, all models achieve high accuracy (95%-100%), but under a multiple-strategy prompt they recover substantially fewer strategies than the human reference set. Gemini, DeepSeek, GPT, and Claude generate 184, 152, 151, and 110 distinct valid strategies, respectively, with the largest gaps in Geometry and Number Theory. The models collectively produce 50 benchmark-novel valid strategies, indicating both incomplete coverage of human strategies and some capacity for alternative reasoning. A repeated-run robustness check on 20 problems shows diminishing gains in discovered strategies, with the strongest model recovering only 39 of 55 AoPS-reference strategies (71%) after three runs. These findings position strategy diversity as a complementary dimension for evaluating mathematical reasoning beyond answer correctness.

Users interacting with Large Language Models (LLMs) in a multi-turn conversation routinely refine their requests or pivot to new topics. LLMs, however, often miss these topic shifts and carry over irrelevant context from previous turns, leading to inaccurate responses. In this paper, we stress-test the multi-turn understanding of LLMs and study the following two sub-tasks: (1) detecting whether the user pivots or refines in the current turn, and (2) shortlisting relevant context from previous turns. To this end, we construct synthetic benchmarks based on real-world datasets from varied domains, as to simulate context shifts of different levels of difficulty. We then evaluate the zero-shot performance of ten LLMs (open-weight, closed-source and reasoning), and demonstrate that only some reasoning and strongly instructed LLMs are accurate in detecting pivots; open-weight LLMs struggle with the task and frequently carry stale context even with explicit cues; and all models suffer from a position bias. Based on the results, we discuss key takeaways for improving long-term robustness in multi-turn capabilities for LLMs.

Large language models (LLMs) are increasingly deployed in financial contexts, raising critical concerns about reliability, alignment, and susceptibility to adversarial manipulation. While prior finance-related benchmarks assess LLMs' capabilities in stock trading, they are often restricted to small sample and fail to demonstrate LLM susceptibility to context with potential human bias. We introduce Fin-Bias (financial herding under long and uncertain financial context), a benchmark for evaluating LLM investment decision-making when faced with uncertainty and possible human-biased opinions. Fin-Bias includes 8868 long firm-specific analyst reports, including firm aspects summarized and analyzed by sophisticated analysts with investment ratings (Bullish/Neutral/Bearish) spanning from various industries. We present large language models with firm analyst reports with/without analyst investment ratings and even with 'fake' rating, to get investment ratings generated by LLMs. Our results reveal that LLMs tend to herd the explicit bias in context. We also develop a method to detect potential human opinions, which can encourage LLMs to think independently, some models even exceed human performance in predicting future stock return.

Evaluating pragmatic reasoning in large language models (LLMs) remains challenging because model behavior can vary depending on evaluation methods. Previous studies suggest that prompt-based judgments may diverge from models' internal probability distributions, raising questions about whether observed performance reflects underlying competence or task-induced behavior. This study examines this issue using scalar diversity as a graded diagnostic for pragmatic inference. Following Hu & Levy (2023), this study compares direct probability measurement and metalinguistic prompting across multiple models and experimental settings. The results show that neither evaluation method consistently outperforms the other and that pragmatic behavior varies substantially across model families, prompting strategies, and task structures. Moreover, scalar diversity gradients emerge only in specific model-condition combinations, suggesting that pragmatic reasoning in LLMs reflects an interaction between internal probabilistic representations and task-induced prompting behavior rather than a stable competence captured by a single evaluation paradigm. These findings highlight the central role of evaluation design in interpreting pragmatic abilities in LLMs.

Large language models are increasingly capable at closed-world mathematical reasoning, but research assistance also requires source-grounded use of the literature. When a proof reaches a non-trivial step, a useful assistant should determine whether the needed tool (e.g., a lemma) already exists, identify a suitable scholarly source, and verify that its assumptions align with the current proof context. To rigorously evaluate such capabilities, we introduce Re$^2$Math, a benchmark for tool-grounded retrieval from partial mathematical proofs. Each instance is built from a candidate instrumental citation in the proof of a main theorem, with hierarchical context and an optional leakage-controlled anchor hint. We also make the task source-grounded yet citation-agnostic in that any admissible theorem sufficient for the proof transition is accepted. Evaluation uses a release-frozen retrieval artifact, ensuring reproducibility, while the benchmark itself supports automatic, continual expansion with newly constructed instances. On the current benchmark test set, the best fixed-judge ToolAcc reaches 7.0%, despite substantially higher rates of source grounding, indicating that current systems often retrieve valid statements but fail to establish their applicability to the local proof step. By decoupling citation recall, grounding, and proof-gap sufficiency, Re$^2$Math transforms literature-grounded mathematical tool use into a controlled diagnostic task.

Large language models (LLMs) increasingly exhibit behaviors suggesting awareness of their evaluation context, often adapting their reasoning strategies in benchmark settings. Prior work has shown that such evaluation awareness can distort performance measurements; however, it remains unclear whether this phenomenon reflects a single behavioral artifact or a deeper internal structure within the model. We propose that LLMs maintain a decomposable space of functional metacognitive states: internal variables encoding factors such as evaluation awareness, self-assessed capability, perceived risk, computational effort allocation, audience expertise adaptation, and intentionality. Through residual stream analysis across multiple reasoning models, we demonstrate that these states are linearly decodable from internal activations and exhibit distinct layer-wise profiles. Moreover, by steering model activations along probe-derived directions, we show that each functional metacognitive state causally modulates reasoning behavior in dissociable ways, affecting verbosity, accuracy, and safety-related responses across tasks. Our findings suggest that benchmark performance reflects not only task competence but also the activation of specific functional metacognitive states. We argue that understandi ng and controlling these internal states is essential for reliable evaluation and deployment of reasoning models, and we provide a mechanistic framework for studying functional m etacognition in artificial systems. Our code and data are publicly available at https://github.com/xlands/meta-cognition.

This study proposes a quantitative framework for profiling LLM dispositions as stable, model-specific regularities in output under repeated, controlled elicitation. Using a structured narrative constraint-selection task administered across six frontier models and three instruction types, we operationalize disposition through two dimensions: "consistency", measured as cross-replication selection overlap via Jaccard similarity, and "diversity", measured as dispersion across options via the inverse Simpson index. We further introduce Narrative Landscape, a PCA-based visualization that maps each model's selection profile into a shared space for direct comparison. Results reveal a clear rigidity-exploration spectrum across model families and show that instruction types shift the geometry of selection spaces even when scalar metrics appear similar, indicating that comparable scores can mask qualitatively distinct selection topologies.

Data preparation is a central and time-consuming stage in data analysis workflows. Traditionally, commercial tools have relied on graphical user interfaces (GUIs) to simplify data preparation, allowing users to define transformations through visual operators and workflows. Recent advances in large language models (LLMs) raise the possibility of a paradigm shift toward natural language (NL)-driven data preparation, in which users can specify preparation intents in NL directly. However, it remains unclear how far current LLM-based agents are from this paradigm shift in practice. Existing code generation benchmarks do not capture key characteristics of data preparation, including ambiguous user intents, imperfect real-world data, and the need to translate code into interpretable workflows for validation. To bridge this gap, we present PrepBench, a benchmark designed to evaluate NL-driven data preparation along three core capabilities: interactive disambiguation, prep-code generation, and code-to-workflow translation. We crawl data from the Preppin' Data Challenges, and then extend it into a systematically designed benchmark. The benchmark covers diverse domains, and each task involves 3 to 18 data preparation steps. Nearly half of the tasks require over 100 lines of Python code, and the longest solutions approach 300 lines. Our evaluation shows that, despite recent progress, realizing this paradigm shift remains challenging for state-of-the-art LLMs. PrepBench provides a principled benchmark for measuring this gap and helps identify key challenges toward realizing NL-driven data preparation.