Reasoning Benchmark

Compositional generalization, the ability to understand and produce novel combinations of known components, remains a fundamental challenge for modern artificial intelligence. While few benchmarks exist, many focus on linguistic tasks and lack complex, explicit compositional structures. We introduce ClassicLogic, a new benchmark suite designed to evaluate an agent's ability to learn and compose problem-solving strategies. The benchmark consists of four classic logic puzzles: Sudoku, KenKen, Kakuro, and Futoshiki. Its core innovation is a hierarchical, explicit knowledge base for each game, where complex solving strategies are formally defined as compositions of simpler, foundational strategies. This structure allows for fine-grained evaluation of an agent's reasoning capabilities, from learning basic rules to applying multi-step compositional strategies to solve puzzles of increasing, mathematically validated difficulty. The open-source benchmark provides a challenging new testbed for advancing neuro-symbolic and other advanced AI reasoning systems.

Machine learning has shown speedups for numerical methods in physics applications, but integrating Python-based libraries into high-performance C++ solvers creates performance bottlenecks. We present CodeJeNN, which bridges this gap by auto-generating self-contained C++ code from trained Keras models for inference. This eliminates external dependencies through minimal inlined functions, allowing seamless integration into existing frameworks. We describe the Keras-to-C++ workflow, supported architectures, and limitations. CodeJeNN is demonstrated through inference benchmarks against Keras in eager and JIT modes and a CFD test case modeling viscosity in a hydrogen-air mixing layer, showing speedups without sacrificing accuracy.

Large Language Model (LLM) social simulations are a promising research method, but they are not yet faithful enough to be adopted widely. In this work, we investigate whether the current scaling paradigm in language modeling is likely to close these gaps, or whether simulation fidelity is orthogonal to general capabilities and therefore deserving of more research attention. We use scaling laws to study the relationship between LLMs' compute scale, general capability benchmarks, and the fidelity of social simulation in three representative sub-domains: opinion modeling, behavioral simulation, and longitudinal forecasting. Surprisingly, we discover strong compute scaling in all three settings, using a suite of 85 transformer LLMs with the Qwen3 architecture pre-trained on the DCLM web text corpus under fixed-compute budgets from $10^{18}$ to $10^{20}$ FLOPs. Then we evaluate 35 larger and more capable open-weight models up to 70B parameters, allowing us to predict downstream accuracy from loss. This reveals that the majority of behavioral and opinion simulation tasks will rapidly improve with scale, particularly when they involve populations that are well-represented in English web corpora. Longitudinal forecasting and underrepresented opinions scale more slowly, especially when they are less correlated with general knowledge and reasoning benchmarks like MMLU. In behavior simulation, scaling fails to improve model calibration with human cognitive biases like risk aversion, as well as human heuristics like learning correlated rewards from related tasks. On these tasks, even fine-tuned models fail to noticeably scale up performance from 0.5B to 8B parameters. Taken together, we conclude that scale will improve social simulations in most settings, but outliers exist, and improvements will be less reliable in low-resource domains.

Parallel thinking has enjoyed great success for boosting LLM performance on reasoning tasks without the need for any re-training. However, existing methods follow a think-first-then-decide paradigm, i.e., they first sample multiple reasoning paths, which inevitably leads to overgeneration, then prune or stop unnecessary paths to compensate. In contrast, decide-first-then-think, i.e., first identifying points that are likely to lead to desirable generations, has been underexplored so far. Following this paradigm, we propose Fork-think with confidence, that first identifies forking points using model confidence in a single seeding path, then triggers thinking, sampling multiple continuations and aggregating them for the final response. Our experiments across three models and three reasoning benchmarks show that Fork-think reduces the token consumption by up to 30% and run-time by up to 57%, while performing comparable to or better than parallel thinking. Our analysis reveals that Fork-think is able to identify forking points that are meaningful with respect to the downstream task and that sampling at later positions can lead to substantially better generations. Finally, we demonstrate how combining Fork-think with existing mechanisms such as early stopping and weighted voting can further boost the performance and perform comparably to existing state-of-the-art methods, without requiring any warm-up or offline training. Our results establish pre-determined forking as a promising research direction for efficient LLM reasoning.

Critic-free reinforcement learning with verifiable rewards (RLVR), exemplified by Group Relative Policy Optimization (GRPO), avoids training a value function (critic) and reduces memory and compute overhead relative to critic-based PPO pipelines for aligning large language models. However, GRPO-style advantage estimation depends on prompt-local (within-prompt-group) reward statistics and can be unstable. In particular, when all rollouts in a prompt group receive identical rewards, the within-group reward variance becomes zero, and group normalization yields zero advantages for that group, impeding learning in cold-start regimes with binary verifiers. We introduce BV-Blend, a critic-free framework that stabilizes advantage estimation by combining prompt-local on-policy statistics with semantic-cluster-conditioned historical moments. BV-Blend maintains EMA-tracked reward moments for each cluster, derives a confidence weight from a standard error of the mean (SEM) proxy, and uses this weight to blend historical and prompt-local baseline and variance statistics into a standardized advantage for PPO-style clipped updates. Experiments on verifiable reasoning benchmarks show that BV-Blend improves training stability and performance, and remains robust in regimes where group-normalized methods may stall.

People overthink; language models over-sample, and the extra effort can talk both into a worse answer. Reasoning systems answer a hard question by sampling it many times (test-time scaling), and the more they draw, the more often a correct answer turns up somewhere, so coverage, the fraction of problems with at least one correct try, climbs and appears to be progress. But a deployed system must return one answer, and choosing it, not knowing which try is right, is selection; selection is capped, and past a point extra samples only make the model surer of a confident mistake, even as every draw adds cost. The gap between climbing coverage and stalled selection, the identifiability gap, is the answer a model can produce but not pick. So the real question is not whether to sample but how far, and the answer is: not far. For picking an answer, the vote has already settled within a few dozen draws, the modal ceiling; for scoring a benchmark, sooner still, the correlation ceiling. Beyond that, extra draws cost compute and add nothing, and can even make the answer worse. This paper turns the cutoff into a single number, the effective number of samples, that any sampling run already reveals. The bottleneck is recognizing a right answer, not generating one.

Recent interest in multimodal large language models (MLLMs) raises a central question: can they reason over dynamic visual evidence rather than merely recognize objects or events in individual frames? This ability, which we refer to as video temporal-logical reasoning, requires models to maintain, update, and compose evidence as visual states evolve across frames. Existing video benchmarks often conflate this capability with scene complexity, static recognition, or uncontrolled temporal variation. To isolate this capability, we introduce Video-MME-Logical, a controlled benchmark organized around five temporal-logical operations: state tracking, sequential counting, temporal ordering, dynamic spatiality, and structural composition. The benchmark contains 25 fine-grained task categories generated with controlled object states, transitions, temporal dependencies, and logical compositions. It enables difficulty-controlled final-answer evaluation by varying temporal horizon and reasoning complexity, and supports intermediate-state diagnostics by verifying whether models recover the required logical reasoning trace before producing the final answer. Experiments with state-of-the-art MLLMs reveal a substantial human-model gap, especially as temporal-logical complexity increases. Supervised fine-tuning on up to 500K generated samples improves performance but remains insufficient to close the reasoning gap, positioning Video-MME-Logical as a scalable testbed for analyzing and improving temporal-logical reasoning in MLLMs.

Reasoning in multimodal large language models (MLLMs) has shown strong promise in medical imaging. However, this reasoning is usually free-form text judged only by its final answer, making it hard to interpret and verify, especially in 3D radiology, where a diagnosis should be traceable to evidence in the scan. Existing chest CT question-answering datasets compound this by reducing expert radiology reports to answer-only pairs, dropping the reasoning that links findings to conclusions and omitting the patient history clinicians rely on. As a result, reasoning-capable 3D chest CT MLLMs remain out of reach, as neither the structured supervision needed to train them nor the protocol needed to verify their reasoning yet exists. We introduce CORTEX (Clinically Organized Reasoning and sTructured EXplanation), a structured reasoning benchmark for 3D chest CT. For each question, CORTEX restores the missing reasoning as a four-stage diagnostic trace mirroring a radiologist's workflow: task understanding, visual observation, diagnostic reasoning, and answer synthesis. We generate these traces using frontier large language models with broad medical and general-domain knowledge, then filter and verify them with a stage-level evaluation protocol combining automated rubric scoring with expert radiologist review. Crucially, both the reasoning structure and evaluation rubrics are designed in close collaboration with clinicians. Built on CT-RATE, a large, publicly available chest CT dataset without reasoning annotations, CORTEX comprises 76,177 validated reasoning traces across open-ended VQA, closed-ended VQA, and report generation, providing both the structured supervision and the stage-level evaluation protocol needed to build and evaluate trustworthy reasoning models for 3D chest CT. Our dataset and evaluation code is available at https://huggingface.co/datasets/aneesurhashmi/cortex

Reasoning capability has advanced rapidly in large language models (LLMs), leading to an increasing size of key-value (KV) cache in both prefilling and decoding stages. Existing KV cache compression methods mainly rely on attention weights to estimate token importance. While attention effectively captures contextual relevance, it overlooks complementary information-theoretic signals related to predictive uncertainty and token informativeness. In this paper, we revisit token importance from a forward-looking perspective and introduce \textit{Forward Influence}, a metric that measures how compressed tokens affect future contexts. Our analysis reveals that tokens selected by attention scores mainly influence nearby contexts, whereas tokens associated with high predictive uncertainty exhibit substantially stronger influence on distant future contexts. Based on the observation, we propose \textbf{InfoKV}, an entropy-aware KV cache compression framework that incorporates information-theoretic signals. It combines token-level predictive uncertainty with layer-wise representation evolution and integrates the resulting entropy scores with attention scores during reasoning. Experiments on long-context reasoning benchmarks with Llama-3.1, Llama-3.2, and DeepSeek-R1 demonstrate that InfoKV consistently outperforms existing attention-based KV compression methods in both long prefilling and decoding scenarios.

Large language models (LLMs) reach high accuracy in mathematical reasoning, but individual traces on the same problem diverge; some arrive at the correct answer while others fail. Prior work analyzes failure at the step, chunk, or sentence level, or at tokens where failure has already occurred. Neither identifies the precise token that triggers the shift toward failure. We introduce the cliff token, a token where the token-wise potential drops significantly under an adaptive threshold that scales with the local token-wise potential, based on a one-sided two-proportion z-test. Across seven models and three mathematical reasoning benchmarks (GSM1K, MATH500, AIME 2025), cliff tokens act as failure triggers; deleting the first cliff token and resampling recovers pass@64 to 1.0, while keeping it limits recovery to between 0.71 and 1.00. We further introduce a cliff taxonomy of deterministic, uncertain, and sampled-off cliffs, defined by greedy choice and token entropy. Each type has distinct probabilistic characteristics, and the taxonomy generalizes across model scales. Finally, we validate the taxonomy via single-token preference optimization at cliff positions (Cliff-DPO). Trained on GSM8K, Cliff-DPO improves accuracy across benchmarks by up to +6.6. Optimizing at uncertain and sampled-off cliffs improves reasoning, while deterministic cliffs do not.

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a promising framework for enhancing the reasoning ability of large language models. However, much of the existing work is guided by heuristic intuition, leading to divergent algorithmic choices, even contradictory ones that nevertheless report empirical gains. To better understand this phenomenon, we conduct a theoretical analysis of RLVR updates. Our study reveals that differences in off-policy degree, determined by the number of gradient steps per rollout, substantially affect the distribution of importance sampling ratios and their clipping behavior, thereby altering which tokens dominate the update. Building on this insight, we characterize gradient expectation as the central quantity governing update dynamics and analyze the roles of token probability, advantage, and importance sampling ratio. Motivated by these findings, we propose Adaptive Clip Policy Optimization (ACPO), which adjusts clipping boundaries across token groups according to the empirical variance of their importance sampling ratios. Experiments on 3B and 7B models across diverse reasoning benchmarks, spanning mathematical problem solving, tabular QA, and logic puzzles, demonstrate that ACPO outperforms strong baselines such as DAPO and CISPO. These results demonstrate that principled, analysis-driven approaches yield more robust and effective RLVR methods. Code is available in: https://github.com/Control-derek/ACPO

Large Language Models (LLMs) achieve remarkable reasoning capabilities through reinforcement learning (RL) post-training. However, existing RL post-training commonly relies on uniform data sampling, which ignores the semantic structure of the training data and the changing capability of the training policy. To address these limitations, we propose Adaptive Data Scheduling (ADS), a dual-level data scheduling framework for pacing RL post-training that replaces uniform sampling with an adaptive distribution over semantic clusters and policy-boundary sample selection. At the cluster level, ADS organizes samples according to semantic patterns and maintains an adaptive inter-cluster distribution to solidify current training progress. At the sample level, ADS performs intra-cluster scheduling to continuously sample policy-boundary samples, which provides informative relative advantages. Experimental results across three LLMs and seven reasoning benchmarks demonstrate that ADS improves average accuracy by 5.2% over Group Relative Policy Optimization (GRPO). Notably, ADS consistently improves RL methods with different objective designs, highlighting its potential as a general data scheduling strategy for LLM RL post-training. The source code is available at: https://github.com/Richard-zrx/ADS.

We propose Keyless Attention, an attention mechanism that eliminates the key projection entirely, operating over queries and values only. This yields a Value-Only Cache that reduces KV cache memory and access overhead by exactly 50% over standard attention, while matching or exceeding standard attention's decode throughput. Beyond efficiency, we introduce Depth-$m$ Attention Factorization: standard attention computes a depth-2 factorization of the attention bilinear form, while Keyless Attention realizes a depth-$m$ instance of this family. At m=3, Keyless Attention matches the projection matrix count of standard attention via a value-space routing matrix that replaces the key projection and introduces a coupling between routing and retrieval. Experiments across five models and four architectures (GPT-2 280M, GPT-2 557M, Pythia 410M, Qwen2 1.5B, and Llama 3.2 1B) show that Keyless Attention matches or outperforms standard QKV attention on perplexity in 4 out of 5 models. On downstream zero-shot evaluation (GPT-2 557M), Keyless Attention outperforms on 4 out of 5 commonsense reasoning benchmarks, while achieving 50% KV cache reduction throughout.

Large Language Models (LLMs) have made significant progress in reasoning, particularly in deductive reasoning, which is crucial for high-stakes decision-making. As models improve, evaluation benchmarks should evolve to keep pace. However, existing benchmarks lack fine-grained control over logical complexity and struggle to balance semantic diversity with logical consistency. To address these issues, we propose QMFOL, an automated framework for generating monadic first-order logic reasoning tasks with quantifiable and controllable complexity. It constructs formal logical structures using conjunction and disjunction patterns, enabling precise control over reasoning depth, width, label types, and distractors. These structures are then translated into natural language via LLMs, with logical consistency ensured through round-trip verification using an external prover. Based on our framework, we build QMFOLBench, a benchmark comprising 2880 instances with 960 configurations across diverse logical and semantic dimensions. Evaluations on six large reasoning models (LRMs) and two LLMs show that performance degrades and computational overhead increases with rising logical complexity. Models perform better on True-labeled tasks than on False or Unknown ones, and exhibit sensitivity to semantic variation. Overall, QMFOL offers a scalable and reliable approach for constructing deductive reasoning benchmarks with controllable complexity, enabling more precise evaluation of reasoning capabilities in modern language models.

Math and science reasoning benchmarks rely on pass@k, the fraction of sampled chains that reach gold, as the canonical per-example difficulty signal. The same signal drives RL with verifiable rewards, math data curation, synthetic curricula, and verifier training. We show this proxy has a persistent blind spot on its hardest stratum: on the eight free-form math cells we test (GSM8K and MATH across four open-weight models), 10.3-22.9% of the examples that no sampling seed solves in six tries are instead solved at matched compute by a six-chain deterministic regime. These are greedy decoding plus five cheap residual-stream perturbations applied via activation grafting, while greedy alone solves at most 6% on these math cells. Recovery scales with the additional budget, across perturbations whose mechanistic distinctness we verify across all twelve cells (cross-kind fix-set Jaccard <= 0.47 in every setup). Activation grafting is used as an intervention on internal representations, not a decoding method; we use it purely as a diagnostic and diversification tool, and our recovered items show that the pass@k= 0 % stratum is structurally identifiable in the residual stream rather than that the unmodified model reaches them under ordinary inference.

Modern Vision-Language Models (VLMs) often struggle with strategic reasoning, i.e., anticipating and influencing other agents' actions, under uncertainty in competitive and cooperative settings. Real-time strategy (RTS) games can be a natural testbed for diagnosing this limitation, as they demand coordination with allies, adaptation to opponents' strategy, and long-horizon planning under partial observability. However, existing RTS benchmarks offer limited evaluation scope, lack systematic competency diagnosis, and remain fixed in the pre-designed scenario coverage. To address these limitations, we present RTSGameBench, which is built on Beyond All Reason, a large-scale RTS game with an expanded battlefield that demands broader strategy diversity than the existing testbeds. The proposed benchmark provides evaluations through diverse gameplay across various matchup structures, diagnostic assessment via mini-games, each targeting an individual strategic competency, and extensible coverage via a self-evolving generation framework that converts free-form queries into new mini-games, improving over successive cycles. Additionally, for VLMs to operate in large-scale RTS games, we provide RTSGameAgent that manages units by an FSM with agentic memory. We empirically validate that multiple state-of-the-art VLMs do not perform well when matchups demand tighter coordination, multiagent coordination and when task scale increases.

Looped architectures provide an inductive bias toward learning step-by-step procedures for tasks that require compositional reasoning. The number of effective layers reached by looping determines the quality of the solution these models find. Like deep architectures, looped architectures are prone to a signal propagation problem induced by depth as the halting decision is postponed. In this paper, we address this signal propagation issue using pre-norm layers and residual scaling. Building on these architectural modifications, we propose FPRM, a Transformer-based Fixed-Point Reasoning Model that uses fixed-point convergence as an end-to-end halting mechanism in a looped architecture. We show that fixed-point halting allows FPRM to adapt its compute to task difficulty. FPRM is effective on common reasoning benchmarks, namely Sudoku, Maze, state-tracking, and ARC-AGI.

Large language models perform increasingly well on standardized logical reasoning benchmarks, but whether this ability remains robust beyond English is unclear. We introduce ChLogic, an English--Chinese aligned benchmark that tests whether models preserve logical reasoning performance when the same latent logical structure is expressed in English and diverse Chinese surface realizations. Built from formal logical templates, the benchmark contains three data sets: (i) the General aligned set, derived from 60 General Propositions across nine template families; (ii) the Difficult aligned set, derived from 40 Difficult Problems; and (iii) the Chinese-only set, covering 15 language-specific phenomenon types. Each aligned item pairs one English reference expression with five Chinese realizations. Experiments on Qwen3, Ministral, and GLM models reveal a persistent English--Chinese performance gap. Back-translation from standard Chinese into English often improves performance on the General aligned set, but produces mixed effects on the Difficult aligned set, where Qwen3-32B and GLM-5.1 perform worse after translation. These results indicate that Chinese surface realization, translation artifacts, and model-specific behavior jointly affect multilingual logical reasoning. Overall, ChLogic provides a useful stress test for the robustness of multilingual reasoning.

Pathological images are inherently multi-scale, requiring pathologists to integrate evidence from global tissue architecture at low magnification to cellular morphology at higher magnification for accurate diagnosis. While existing pathological datasets for vision-language models (VLMs) include various scales, they often lack explicit cross-scale reasoning objectives. This limitation prevents VLMs from capturing essential cross-scale representations and learning evidence-based reasoning. To bridge this gap, we introduce the first cross-scale training and evaluation paradigm that formulates pathology interpretation as multi-magnification reasoning. However, creating such a task reveals a critical challenge: multi-image visual question answering (VQA) is prone to text-only shortcuts, which allow models to guess answers using magnification-dependent artifacts rather than visual evidence. To address this, we propose a leakage-aware curation pipeline that combines adversarial text-only screening with constraint-guided question design. Using this pipeline, we construct Scale-VQA, a high-quality benchmark with 4,685 multiple-choice questions grounded in 2,537 pathology images across multiple magnification levels. Finally, we present ScaleReasoner-R1, a model trained via reinforcement learning to optimize performance on cross-scale VQA tasks. ScaleReasoner-R1 achieves state-of-the-art performance on our cross-scale reasoning benchmark and generalizes to SOTA performance on established single-scale benchmarks. Findings suggest that even the limited cross-scale supervision can significantly improve pathological understanding. Code is available at https://github.com/iMVR-PL/ScaleReasoner-R1.

Biomedical question answering (QA) increasingly requires reasoning over interacting entities, where supporting evidence is scattered across biomedical knowledge graphs, literature documents, and web-accessible resources. However, existing biomedical QA benchmarks mainly focus on exam-style knowledge, literature comprehension, or short-range multi-hop inference, leaving source-conditioned graph reasoning and evidence topology construction underexplored. To fill this gap, we introduce BioMedHop, a multi-source graph-grounded benchmark for evaluating biomedical reasoning over structured evidence topologies. BioMedHop contains 10,045 instances across KG, document, web, and hybrid evidence settings, covering shared-neighbor matching, intersection reasoning, path-based reasoning, and counting, with option-based, open-ended, and numeric count renderings. To support this benchmark, we further propose BioWeave, a source-aware reasoning framework that retrieves biomedical KG paths, gathers supporting clues from documents and web sources, assembles them into a unified evidence graph, and verifies answers through entity-level evidence support. Comprehensive experiments show that BioWeave achieves the best overall performance among compared methods on BioMedHop, outperforming the strong hybrid baseline ToG-2 by 10.5% in the overall average. Moreover, BioWeave consistently improves different LLM backbones and enables smaller models, such as Qwen3-4B, to achieve reasoning performance comparable to GPT-4-Turbo.

Reinforcement learning with verifiable rewards (RLVR) has become a key approach for enhancing the reasoning abilities of large language models. However, RLVR often suffers from \emph{diversity collapse}: Pass@$1$ improves while high-$k$ Pass@$k$ degrades, which is viewed as a narrowing of the model's reasoning boundary. We formalize this diversity collapse through the lens of \emph{overtraining}: once a problem's contribution to the reference metric has effectively saturated, further updates no longer expand what the model can solve but still concentrate probability mass on the trajectories favored by on-policy sampling. Under a standard setup with few rollouts per problem, even a single observed success places a problem in a nearly saturated regime for high-$k$ Pass@$k$, so most updates in standard RLVR are overtraining from the boundary perspective. This perspective also suggests a reading of whether RLVR can expand the model's reasoning abilities beyond the base model: since RLVR is structurally biased against high-$k$ Pass@$k$, its aggregate decline does not by itself mean that no new reasoning gains occurred. Interventionally, restricting updates to problems with zero observed success lifts Pass@$256$ above the base model on difficult benchmarks; observationally, a non-trivial fraction of initially unsolvable problems become solvable during standard RLVR training. Building on these findings, we propose \emph{Bayesian Boundary Gating} (BBG), which redirects optimization away from overtraining by estimating each problem's marginal contribution to the reasoning boundary. Across multiple reasoning benchmarks, BBG improves average Pass@$k$ across a wide range of $k$.

Large Language Models (LLMs) have made notable progress in automated theorem proving, yet existing formal benchmarks remain limited in both mathematical coverage and difficulty. Most are concentrated in areas that are easier to formalize, such as algebra and elementary number theory, and provide limited coverage of subfields that require deeper reasoning, including mathematical analysis. To address this gap, we introduce MA-ProofBench, to the best of our knowledge, the first formal theorem-proving benchmark dedicated to Mathematical Analysis. The benchmark contains 200 formalized theorems covering 6 core topics and 27 subcategories, including measure and integration theory, complex analysis, and functional analysis. The problems are divided into two difficulty levels, an undergraduate level (Level I, 100 problems) and a Ph.D. qualifying level (Level II, 100 problems), to evaluate how well LLMs perform formal reasoning at different mathematical depths. Each problem is constructed through a human-led, LLM-assisted formalization pipeline followed by independent expert review, ensuring that the formal statements remain faithful to the original mathematics. We evaluate a range of recent general-purpose reasoning models and formal theorem provers on MA-ProofBench. However, most models perform poorly: even the best-performing model, GPT-5.5, achieves only 16% Pass@8 on Level I and 5% on Level II, while most models stay close to 0% on Level II. Further analysis identifies Mathlib hallucinations and incomplete proofs as the two dominant failure modes, while an evaluation on the natural-language version of the benchmark exposes a clear gap between informal and formal reasoning. MA-ProofBench is intended to serve as a reliable reference for tracking progress in formal mathematical reasoning in advanced domains.

Three paradigmatic forms of inference recur across scientific reasoning: deduction, induction, and causal abduction. Reliably evaluating LLMs on these in scientific settings is currently out of reach: scientific benchmarks built on human annotations are costly and lack mechanistic ground truth, while synthetic logical-reasoning benchmarks do not resemble real scientific documents. We introduce SciR, a benchmark that combines multi-paradigm reasoning with controllable scientific rendering, anchored on three paradigmatic scientific problems. Tasks are generated from formal objects (deduction tree, inductive rule hypothesis, causal graph) to guarantee verifiable answers, then rendered into multi-document scientific discourse via per-track domain-tuned genres. The construction lets us independently vary two difficulty axes: how hard it is to extract the key information needed for inference, and how hard the principled inference itself is. We test six models. Both axes hurt every model, and their effects compound. The rendering even hurts neurosymbolic pipelines, which hand inference to a verified solver. The two axes yield a per-model extraction-vs-inference profile: for instance, reasoning models like deepseek-r1 mostly surpass non-reasoning instruct models on the inference axis. To our knowledge, SciR is the first multi-paradigm scientific-reasoning benchmark with parametric control on both extraction and inference difficulty.

Conversational shopping assistants now serve hundreds of millions of customers, yet no existing benchmark jointly evaluates the open-ended multi-turn reasoning, domain expertise, and criterion-level quality that real shopping conversations demand. Shopping reasoning is unique among language model applications. Unlike factual question answering or verifiable code generation, it requires balancing subjective preferences, budget constraints, and cross-product trade-offs across multi-turn dialogue, capabilities absent from previous e-commerce and general-purpose benchmarks. We introduce the Shopping Reasoning Bench, an expert-authored benchmark of 525 missions (232 single-turn, 293 multi-turn) with 10863 importance-weighted binary rubrics authored by retail domain experts. These criteria are organized under a taxonomy of five reasoning categories and fifteen subcategories covering diverse demands such as preference refinement, trade-off analysis, and compatibility assessment. An evaluation of nine models across three families (GPT, Claude, Gemini) shows that pass rates reach only 57--77% overall. On multi-turn missions, all models score 13--29 points lower on optional above-and-beyond criteria than on required ones, and performance degrades 4--18 points as conversations progress. These gaps show that current models handle basic shopping assistance but fall short of expert-level advice, making Shopping Reasoning Bench a challenging testbed for future shopping assistant development.

Instruction-tuned LLMs are increasingly converted into reasoning models through post-training to improve multi-step task performance. This conversion is usually optimized for reasoning accuracy, without explicitly preserving the alignment behavior of the instruction-tuned model, such as safe refusal, bias avoidance, and privacy protection. We ask: does this conversion preserve alignment? We study this question through a trustworthiness audit and find that it is not behavior-preserving by default. For a systematic analysis, we compare reasoning models produced via supervised fine-tuning, RL-based post-training, and distillation against matched instruction-tuned baselines across six trustworthiness dimensions: safety, toxicity, stereotyping and bias, machine ethics, privacy, and out-of-distribution robustness. We observe that reasoning models often improve on reasoning benchmarks but exhibit alignment regressions, including increased toxicity, amplified stereotyping, miscalibrated refusal, and contextual privacy leakage. These regressions are consistent with behavioral drift from the instruction-tuned baseline, measured by KL divergence. Overall, our results point to the broader conclusion that trustworthiness metrics are essential for evaluating reasoning models and should be reported alongside gains in reasoning capability.

Combinatorics is central to Olympiad-level mathematical problem solving, requiring deep discrete reasoning, creative constructions, and rigorous structural insight. Recent evidence suggests that even today's strongest frontier models remain uneven on Olympiad combinatorics, revealing a gap in creative mathematical reasoning. We introduce ComBench, an Olympiad-level combinatorics benchmark for evaluating and diagnosing the combinatorial reasoning capabilities of large language models. ComBench contains 100 human-annotated competition-level problems organized around two complementary settings: analysis-centric problems, which primarily require rigorous mathematical arguments, and construction-centric problems, which require explicit constructions in addition to correctness justifications. The evaluation protocol combines rubric-guided proof grading with deterministic construction verification, exposing cases where proof quality and construction validity diverge. Experiments on frontier open- and closed-source models show that ComBench is far from saturated: the strongest model reaches 65.4% overall Avg. and 75.3% overall Best@4. We further find that Rigorous Proof Reasoning and Constructive Realization are distinct capabilities: Kimi-K2.6 trails GPT-5.5 on analysis-centric proof grading but surpasses it on construction-centric Best@4, while Existence and Construction problems remain consistently hardest across representative frontier models.

Math reasoning benchmarks have proliferated, yet most lack a per-item difficulty signal grounded in actual human performance. We introduce KCSAT-ML, a decade (2014-2025) of Korean College Scholastic Ability Test (KCSAT; Suneung) mathematics: 664 problems with a 339-item core set carrying official per-item error rates from nationwide cohorts of hundreds of thousands of examinees. We pair the benchmark with Difficulty-aligned Reasoning Gain (DRG): a score-orthogonal metric that asks whether a model's mistakes concentrate on the items humans found hard, or on items humans found easy. Together they expose, across a wide range of VLMs (and LLMs via OCR), three patterns: (i) low-budget accuracy collapses on the high-human-error tail at every model size; (ii) test-time scaling (TTS) raises token use roughly linearly with cohort error rate, while accuracy gains follow a non-monotonic curve; (iii) within a single family, TTS flips between anti-scaling on the hardest items and overthinking on easier ones -- two faces of the same alignment failure. On DRG, models with near-identical accuracy can sit at near-opposite values: one model gets wrong what humans also find hard, while another solves the hardest items yet fails on items humans find easy -- a contrast that aggregate accuracy hides. Our code and dataset builder will be open-sourced at https://github.com/naver-ai/KCSAT-ML.

Spreadsheets and tables are widely used representations for structured data analysis, but effective analysis still requires substantial manual effort and domain expertise. Recent large language model (LLM) agents can automate parts of this process, but they often provide limited transparency into intermediate decisions, rely on implicit assumptions, struggle with multi-table comparison, and repeat similar workflows without adapting to a user's preferences. This paper presents TabClaw, an open-source interactive AI agent for spreadsheet manipulation and table reasoning. Users upload CSV or Excel files and issue natural-language requests; TabClaw clarifies ambiguous intent, exposes an editable execution plan, streams a ReAct-style tool-using analysis loop, dispatches specialist agents for parallel multi-table reasoning, and synthesizes findings with explicit consensus and uncertainty markers. Beyond one-off analysis, TabClaw records completed workflows, extracts persistent user memory, distills reusable skills from repeated tool-use patterns, supports package-style skill import, and upgrades skills from negative feedback. Experiments on spreadsheet manipulation and table reasoning benchmarks show that TabClaw improves executable task completion and reasoning performance while preserving an inspectable user workflow. This paper shows how TabClaw turns spreadsheets and tables into inspectable analytical workflows while gradually personalizing itself to recurring data-analysis tasks. Our code is available.

Large language models (LLMs) have shown impressive performance on diverse reasoning tasks, yet their capacity for structural reasoning in graphs remains unclear. We investigate whether LLMs can genuinely understand graph isomorphism -a fundamental problem in graph theory. While LLMs achieve near-perfect accuracy on isomorphism detection, we show this performance is illusory. When identical graphs are presented with permuted node labels, LLMs fail to identify their isomorphism. This finding suggests that LLMs exploit patterns rather than reasoning about abstract graph structure. Since permutation invariance is a fundamental requirement for valid structural reasoning, these results indicate that success on graph reasoning benchmarks should not be interpreted as evidence of genuine topological understanding.

LLMs have recently achieved strong results on formal proving benchmarks. However, existing evaluations remain heavily concentrated on competition-style problems and often fail to capture how models behave on longer, more dependency-rich mathematical developments. We introduce TheoremBench, a Lean4 benchmark designed to evaluate theorem provers beyond contest settings. The benchmark is built from nearly one hundred classical theorems and is released in two complementary forms: a plain main version containing one target theorem per instance, and a premised version that expands each theorem into a structured family of related proving tasks consisting of the main theorem together with automatically extracted supporting subtheorems. This design enables evaluation of not only whether the final theorem was proved from scratch, but also of partial progress through the internal proof structure of a theorem. Our experiments show that explicit premises substantially improve performance for Lean4-capable prover models. To provide a comprehensive evaluation, we introduce theorem-level coverage and token-efficiency metrics that expose qualitative differences in proof behavior. The results show that current provers remain strongly biased toward easy subtheorems and often solve theorems through long and inefficient tactic traces rather than compact proof plans. TheoremBench therefore provides a more fine-grained view of formal reasoning ability and highlights the importance of structural benchmark design for evaluating Lean4 theorem provers.

Recently, language models have made rapid progress across various domains and applications. However, their capability for self-improvement, i.e., whether they are adept at recognising and correcting flaws in their own reasoning, remains dubious. In this study, we address this question by constructing a sufficiency test to rigorously examine the self-correction capabilities of small language models (SLMs). We propose a minimal three-step self-correction pipeline that collects initial SLM answers, prompts the same model to generate hints for its incorrect responses given the ground truth, and feeds the model the same question with its own feedback to refine the initial answer. We evaluate a variety of instruction-tuned and reasoning SLMs in this experimental setup on arithmetic and logical reasoning benchmarks. Our findings show that SLMs with injected hint sentences yield only a 4.4 percent gain over initial question-answering accuracy. Even though the correct answer was provided alongside the model's incorrect reasoning, the evaluated SLMs fail to understand what was missing in their reasoning and show minimal semantic difference between hints that lead to corrections and ones that do not. Furthermore, our experiments show that longer hints are positively correlated with incorrect final answers, suggesting that longer deliberation on problems can hinder the reasoning process, meaning that SLMs do not necessarily scale in performance with a larger compute budget.

Between April 1 and May 15, 2026, a group of 49 mathematicians compiled a dataset of research-level mathematics questions with known answers. Most of the work was done during the 3-day workshop Benchmarks in Leipzig with 35 participants at the Max Planck Institute for Mathematics in the Sciences in Leipzig, Germany. We present the resulting collection of 100 questions. We evaluated these questions in three stages: a single attempt by five state-of-the-art LLMs, followed by a 20-runs-per-model evaluation with three of these models, and finally a 3-run attempt with two heavy-thinking models. After Stage 1, 41 questions remained completely unsolved; after Stage 2, this count dropped to 16; and we concluded Stage 3 with only 2 unsolved questions. This demonstrates that the mathematical reasoning capabilities of LLMs are becoming impressive.

In real-world applications, models are expected to perform reliably across diverse settings. Yet, many existing multimodal benchmarks expand task types without capturing the visual diversity needed to handle open-ended visual inputs. We present WorldBench, a challenging and visually diverse reasoning benchmark to evaluate Multimodal Large Language Models (MLLMs). We build a taxonomy of thousands of visual concepts across multiple domains (e.g., living things). Guided by this taxonomy, we curate a broad collection of images from search engines and existing datasets to comprehensively represent the visual world. Through structured trial-and-error, we manually design challenging questions that frontier MLLMs fail to answer. On quantitative and human evaluations, WorldBench achieves higher visual diversity than any existing diverse benchmark. Evaluating 15 MLLMs on WorldBench reveals weaknesses in visual understanding: even the strongest model reaches only 64.0% accuracy, while some models perform marginally above chance-level. We hope our work highlights the importance of visual diversity in building multimodal benchmarks.

Post-training Small Language Models (SLMs) for reasoning typically follows an SFT-then-RL pipeline, yet existing work rarely considers what data should be learned at each stage. We argue that data strategy should be aligned with the distinct roles of SFT and RL: SFT is better suited for acquiring not-yet-mastered reasoning skills, while RL is better suited for consolidating skills that the model can already partially access. Based on this principle, we propose a difficulty-aware SFT-then-RL framework that organizes training data into stage-specific sets. For hard samples in the SFT stage, we introduce a Bridge mechanism that transforms raw teacher-generated reasoning traces into more learnable supervision for SLMs. For hard samples that remain unsolved during RL, we apply Critique Fine-Tuning by converting all-zero-reward failures into diagnostic, repair, and new reasoning trace supervision for the next SFT stage. Experiments on two SLMs across five reasoning benchmarks show that our method consistently improves over representative SFT, distillation, and RL baselines. Our results highlight the importance of coordinating data difficulty across SFT and RL for effective SLM reasoning post-training.

Low-Rank Adaptation (LoRA) is a popular parameter-efficient fine-tuning (PEFT) method that restricts weight updates to low-rank adapters, introducing a fixed low-rank inductive bias by optimizing in a low-dimensional subspace. In this work, we question whether a fixed-rank constraint is the most effective inductive bias for parameter-efficient fine-tuning. We introduce Learnable Rank LoRA (LR-LoRA), a PEFT method in which the adapter rank is learned during the training process. Instead of prescribing a uniform rank for all adapter layers, LR-LoRA allows the optimizer to determine the appropriate rank for each layer. Using this approach, we find substantial layer-wise variation in the learned ranks, with the attention and MLP layers in the transformer models exhibiting systematically different rank preferences. Across a range of language understanding and commonsense reasoning benchmarks, LR-LoRA achieves state-of-the-art performance in most settings and consistently outperforms strong PEFT baselines, demonstrating that a learnable rank provides a more flexible and effective inductive bias than fixed-rank adaptations.

AI agents can increasingly handle the mechanical tasks of financial analysis: retrieving documents, calculating formulas, updating spreadsheets. The harder, more valuable challenge is reasoning through the open-ended questions that define expert Analyst work. Existing benchmarks do not capture this class of problem, and those that attempt to evaluate open-ended reasoning rely on model-judged outputs that introduce noise and circularity. We present Hedge-Bench 1.0: a benchmark of 102 actual, on-the-job tasks grounded in the explicit reasoning traces of professional hedge fund analysts working with relevant information sources. This approach enables deterministic grading against verified expert steps. Frontier models and agents score below 16% on the benchmark. We publish the dataset and evaluation harness at github.com/Trata-Inc/trata-hedge-bench.

Agentic reinforcement learning (RL) equips large language models (LLMs) with tool-use capabilities that substantially improve reasoning on complex tasks. However, integrating external tools often destabilizes training: over-reliance on tools can induce input distribution shift, while overly conservative tool use limits effective exploration. To address this issue, we propose a unified framework TAO-RL that couples tool-aware trajectory filtering with entropy-guided exploration for efficient policy optimization. Specifically, at the data level, TAO-RL filters rollout trajectories along two criteria: discarding those where all tool invocations fail to execute, and removing those where all rollouts are either correct or incorrect, as both cases yield degenerate advantage estimates that contribute no discriminative learning signal. This joint filtering retains data that are both tool-capable and informative, establishing a high-quality training distribution. At the algorithmic level, we introduce a tool-aware entropy-guided bonus that reshapes the advantage function at post-tool-call tokens, encouraging the policy to explore more diverse reasoning paths at critical decision points. These two components are mutually reinforcing: trajectory filtering establishes a clean and informative training foundation, while entropy-guided exploration drives stronger reasoning behaviors at critical tool-interaction junctures. Extensive experiments on 7 challenging reasoning benchmarks across 3 model scales demonstrate the superiority of TAO-RL over existing methods.

Test-time reinforcement learning has emerged as a promising paradigm for enhancing the complex reasoning abilities of large language models in a completely label-free manner. Despite existing studies focusing on Pass@1 performance, optimizing Pass@k remains under-explored yet critical in label-free settings, which measures generation coverage for sustained exploration. Optimizing Pass@k in label-free setting is highly non-trivial, as directly applying the Pass@k advantage designs effective for RLVR yields unsatisfactory performance. Through in-depth empirical analysis, we discover the root causes hindering performance: pseudo-label estimations for low-confidence samples have a high probability of being incorrect, while candidate answers for high-confidence samples suffer from severe diversity collapse. To overcome these hurdles, we propose TTRL-CoCoV (Test-Time Reinforcement Learning with Confidence-Conditioned Verification), a novel confidence-adaptive framework that expands Pass@k coverage and improves Pass@1 performance. Based on our key insight that verification capability generally leads generation capability, TTRL-CoCoV employs a confidence-conditioned mechanism: for high-confidence samples, it bootstraps verifier and applies an exploration-enhancing reward to prevent diversity collapse; for low-confidence samples, it delegates pseudo-label selection to the verifier to filter incorrect pseudo-labels; and for medium-confidence samples, it bypasses verification entirely. Extensive experiments demonstrate that TTRL-CoCoV outperforms the best competing methods across 6 widely-recognized benchmarks, achieves average absolute gains of +9.8% in Pass@1 and +18.7% in Pass@16 over TTRL, and even achieves absolute Pass@1 improvements of up to +5.0% across multiple reasoning benchmarks when compared against fully supervised RL methods. Our code repository: https://github.com/shanjf666/CoCoV.

We introduce SagaQA, a long-form video benchmark for multi-hop reasoning over full-length TV series. Existing video reasoning benchmarks often emphasize local understanding of adjacent frames or clips. SagaQA addresses this gap by requiring high-level comprehension of extended multimodal narratives in entire TV shows. A distinguishing feature of SagaQA is the granularity of its reasoning steps. Our dataset necessitates long-range reasoning hops to connect information across completely different episodes. This requires models to reason over entire events and actions, demanding a deep understanding of the show's narration and progression at a multimodal level. Motivated by recent progress in agentic methods, we further study how different planning strategies handle such complex reasoning. We categorize these approaches into three classes-Parallel, Sequential, and Hybrid planners-and evaluate their ability to generate coherent and complete reasoning plans. Our results on SagaQA suggest that hybrid planners consistently produce higher-quality plans and exhibit stronger capabilities for complex, high-level narrative understanding in TV shows.

While Large Language Models (LLMs) achieve impressive performance on multi-step reasoning tasks, their reliability is persistently hindered by critical limitations such as unconstrained hallucinations and poor numerical computation. Fundamentally, these issues arise because standard models treat reasoning as a transient, one-off generation process rather than retaining and refining successful procedural logic. To address these challenges, we propose eMoT (evolving Memory-of-Thought), a unified framework that stabilizes multi-step reasoning by treating reasoning trajectories as dynamic, evolving memories rather than static templates. The framework primarily consists of three interconnected modules: (i) a memory corrosion mechanism that reinforces high-utility reasoning structures while gradually decaying less frequent ones; (ii) a symbolic anchoring engine that utilizes Python for deterministic computation, much like a human uses a calculator; and (iii) a consistency-driven refinement process that aligns neural inference with symbolic outcomes, reducing the accumulation of logical discrepancies. Across multiple reasoning benchmarks, eMoT improves accuracy and solution consistency over standard Chain-of-Thought and structured reasoning baselines.On the traditional task Game of 24, eMoT achieves 100% accuracy, surpassing the baseline by up to 17.6%. Evaluations on mathematical task GSM8K, ASDiv, SVAMP, and MGSM further show consistent gains in multi-step mathematical reasoning. In our evaluation, we achieve superior performance despite utilizing a lightweight backbone model with constrained baseline capabilities. Compared to alternative methods that rely on massively scaled models, our results demonstrate that the performance gains are fundamentally driven by the eMoT framework's reasoning control rather than sheer model size.

Recent work suggests that large language models (LLMs) can improve through self-evolution (SE), using supervision signals generated by the model itself. In this work, we ask: under a strict closed-loop setup, where the self-evolution algorithm has access only to an unlabeled prompt set and a base model, how close can internally generated supervision come to oracle-supervised training? We analyze four representative strategies in a unified offline self-evolution framework: single-round verification, multi-turn revision with feedback, iterative training, and curriculum learning. Our primary experiments use Knights and Knaves (KK) logical reasoning tasks, which provide deterministic solutions, controlled difficulty levels, and a clean testbed for easy-to-hard generalization. We first show that self-evolution consistently improves over the base model, but plateaus after excessive training compute is invested, and eventually still leaves a non-trivial gap to oracle supervision. We find that multi-turn critic-revision with large models can reach strong self-evolution performance, with Gemma 12B nearly matching oracle-supervised training. Beyond Knights and Knaves, we also evaluate self-evolution on real-world reasoning benchmarks, where gains are also modest. Overall, our results characterize when closed-loop self-evolution can help and show how internally generated supervision remains insufficient under this minimal formulation.

Self-distillation improves learning efficiency by rewriting reference answers as training data that better matches the model's own distribution. However, reference answers also introduce strong stylistic biases, causing the generative model to imitate surface forms rather than learn useful reasoning patterns. We observe that the rewriting data contains a large number of high-perplexity (PPL) tokens, coming from two distinct sources: beneficial knowledge-enhancing logical corrections, and harmful stylistic drift induced by reference imitation. Treating all such tokens equally can disrupt the base model's original distribution and degrade performance, especially on difficult reasoning tasks. To address this, we propose Distribution-Aligned Self-Distillation (DASD), which uses an answer-aware reference model to generate candidate tokens and dynamically filters them according to the base model's confidence. DASD preserves tokens that encode useful logical knowledge while suppressing distributionally misaligned style noise. Experiments on math, code, and commonsense reasoning benchmarks show that DASD consistently outperforms competitive baselines, reduces high-PPL tokens, and improves robustness across tasks of varying difficulty.

Large language models (LLMs) exhibit strong reasoning capabilities, yet most LLM-based agents are statically deployed and unable to improve through task interactions. Existing experience-driven methods often rely on memory or heuristics without enhancing the model's ability to learn, treating it as a passive executor and leading to early performance plateaus and limited long-term improvement. To address this issue, we propose MetaEvo, a two-stage framework for continual agent evolution that focuses on improving how the model learns from tasks experience, rather than solely on what it stores. MetaEvo first applies preference-based optimization to enhance the model's ability of principle abstraction, then enables the accumulation and reuse of these principles within a modular agent architecture. Experimental results on diverse reasoning benchmarks demonstrate that MetaEvo consistently outperforms strong baselines, maintains reliable improvement across iterations. These findings validate the effectiveness of meta-optimization in enabling agents to learn from experience and continually enhance their reasoning capabilities.

Machine unlearning evaluation is structurally skewed: Why-type questions, which probe causal and relational knowledge, comprise less than 0.06% of CounterFact, 0.6% of ZSRE, and less than 1.3% of TOFU, MUSE, and WMDP-Cyber. This near-zero representation means that methods that fail on causal knowledge can score highly in aggregate, and this failure is undetectable without balanced evaluation. We present 5WBENCH, a balanced 5,000-sample benchmark with 1,000 examples per 5W category (Who, What, When, Where, Why), making causal unlearning failures quantifiable for the first time. Using 5WBENCH, we show that no existing baseline simultaneously achieves high forgetting and high retention on Why-type questions: aggressive forgetting degrades retained knowledge, while conservative methods fail to forget causal facts. Why-type difficulty stems from multi-hop reasoning chains (44% of Why entries vs. less than or equal to 2% for others) and gradient dilution over 40.1-token answer spans. We present MAAT (Multi-phase Adapter-Aware Targeted Unlearning), a three-phase framework operating on LoRA adapter weights, combining gradient-projected ascent, SVD rank-dimension pruning, task vector negation, and hybrid KL-hidden-state retain repair. MAAT is the first method to simultaneously achieve high forgetting and high retention on Why-type causal knowledge, reaching a new operating point on the forget-retain Pareto frontier. We make our code publicly available.

To improve the reasoning capabilities of large language models, test-time compute is typically scaled by generating intermediate tokens before the final answer. However, this couples reasoning to autoregressive generation and thereby conflates internal computation with external communication. In contrast, human cognition can use working memory to hold and manipulate information internally without the need to externalize intermediate thoughts. Drawing on this principle, we introduce Reasoning in Memory (RiM), a latent reasoning method that replaces the autoregressive generation of reasoning steps with memory blocks. These memory blocks are fixed sequences of special tokens that unlock the working-memory capacity of large language models. Since they are fixed rather than generated, they can be processed in a single forward pass, enabling compute-efficient latent reasoning. To operationalize these memory blocks, we employ a two-stage curriculum. First, we ground them by predicting explicit reasoning steps after each memory block. Second, we discard this step-level supervision and iteratively refine the final answer after each memory block. Our experiments on reasoning benchmarks show that, across language models of different families and sizes, RiM matches or exceeds existing latent reasoning methods while avoiding the autoregressive generation of thoughts. These results demonstrate that large language models can be trained to use working memory as an effective mechanism for latent reasoning.

Large Reasoning Models (LRMs) achieve strong performance on table reasoning tasks but incur substantial inference cost due to long reasoning traces. Stepwise model routing mitigates this issue by dynamically assigning reasoning steps to smaller or larger models. However, stepwise model routing for table reasoning remains underexplored. Through empirical analysis, we find that reasoning steps involving tables contain two types of tokens with distinct uncertainty distributions: table tokens grounded in table structure, such as cell values and headers, and text tokens representing surrounding natural-language reasoning. The uncertainty of both token types is correlated with the risk that the model makes an error in the next reasoning step. However, existing methods fail to model them separately, leading to suboptimal routing decisions. To address this, we propose EcoTab, a table-aware stepwise routing framework for efficient table reasoning. At each reasoning step, EcoTab separately estimates the uncertainties of table tokens and text tokens, maps them to next-step failure risks for the small model, and combines the two risks for routing. Experiments on multiple table reasoning benchmarks show that EcoTab consistently outperforms strong baselines and achieves a better balance between accuracy and efficiency.

A paraphrase-quality audit of MathCheck (ICLR 2025) detected 4 semantically incorrect paraphrases in 129 groups (3.1%); removing them drops GPT-4o from rank 2 to rank 4 and elevates Claude Haiku and DeepSeek V3 above it; these ranking changes are invisible to any single-model evaluation. Cross-model unanimity found these errors automatically (>= 3/4 models for MathCheck; >= 6/9 for our primary evaluation) for under $10; in our own dataset the same protocol found that 47% of auto-generated connective-variation paraphrases were semantically incorrect. That flaw compounds a deeper measurement gap: Claude Haiku 4.5 achieves 86% accuracy yet SCR=50%, meaning half its theorems are answered differently under semantically equivalent restatements, while aggregate accuracy across 9 models spans only 86-96% yet Semantic Consistency Rates (SCR) span 50-82% -- a 32-point gap invisible to standard benchmarks. Formally, for any target ranking over 9 frontier models there exists a weighting over paraphrase families that realizes it (No-Free-Benchmark corollary), because no model Pareto-dominates all families -- so benchmark designers who select families are implicitly choosing which model wins. FormInv supplies the audit protocol (replicated on external benchmarks at 100% recall), SCR and per-theorem Cochran's Q as primary invariance measures evaluated on 9 models across 366-811 items (on Lean4-verified theorems), and FormInvSelector for regime-aware model selection.

Hybrid-reasoning large language models (LLMs) expose explicit controls over reasoning effort, allowing users or systems to trade off answer quality against inference cost. However, existing methods for adaptive thinking-mode selection are typically evaluated under different models, datasets, and implementation assumptions, making it difficult to compare their practical behavior. We introduce HRBench, a unified evaluation framework for studying thinking-mode switching in hybrid-reasoning LLMs. HRBench organizes the design space along two axes: three switching strategy families, prompt-based selection, external routing, and speculative execution, and four training regimes, training-free, SFT, offline and online RL, yielding 12 controlled evaluation settings. We evaluate these settings across 6 LLMs, from Qwen3.5-2B to Kimi-K2.5-1.1T, and 5 reasoning benchmarks covering mathematics, science, and code, while reimplementing 12+ representative prior methods within the same pipeline. Our analysis characterizes how different switching strategies occupy distinct effectiveness-efficiency trade-off regions: prompt-based methods often provide favorable token-accuracy trade-offs, routing methods offer more stable cost reduction, and speculative methods tend to improve accuracy at higher token cost. We further find that training affects strategies differently, and that the preferred strategy varies with model scale and task domain. HRBench provides reference implementations and a unified evaluation platform to support more controlled research on efficient reasoning in hybrid-reasoning LLMs. Our data, code and repository are available at https://github.com/usail-hkust/HRBench.

On-policy distillation (OPD) transfers reasoning behavior by training a student on teacher feedback along student-generated trajectories, but standard full-rollout training ties every update to a costly completion and can over-allocate supervision to late positions with low marginal value for the current student. We revisit this assumption through the useful supervision horizon: student-induced rollouts can drift from teacher-preferred continuations, while aligned prefixes may already preserve the long-horizon OPD update direction. We propose ADWIN, an adaptive-window framework for OPD that treats rollout length as an online admissibility decision, training on short teacher-anchored prefixes while using delayed full-rollout probes to audit prefix--full alignment and adapt the next horizon with staleness control. Across math and code reasoning benchmarks in single-task, multi-task, and strong-to-weak settings, ADWIN improves the accuracy--compute trade-off over full-rollout OPD and prefix-based baselines, reducing end-to-end training cost by up to 4.1 times while achieving comparable or better accuracy.

Implicit Chain-of-Thought (CoT) reduces the inference cost of large language models by internalizing the explicit rationales. However, existing approaches typically lack alignment with explicit rationales and adaptivity to example complexity. In this work, we propose CIRF (\textit{\underline{C}hain-of-thoughts \underline{I}nto \underline{R}eusable \underline{F}unctional units}), an implicit CoT framework that performs reasoning as a dynamic sequence of discrete functional tokens. CIRF assigns a functional token to each semantically coherent reasoning unit in explicit CoT traces. The model is then fine-tuned to autoregressively generate functional tokens and their optional results, followed by the final answer. This design aligns latent reasoning with a sequence of functional units, facilitating parallel training, explicit rationale alignment, and adaptive reasoning. Extensive experiments on mathematical, symbolic, and commonsense reasoning benchmarks show that CIRF provides a favorable accuracy-latency trade-off compared with state-of-the-art implicit CoT methods. Further analyses demonstrate that CIRF constructs distinct, interpretable functional tokens, leading to consistent performance improvements.

Large Language Models (LLMs) are being applied to increasingly difficult problems and use cases. To navigate their vast solution spaces effectively, LLMs need to be creative. Yet the subjective nature of creativity and the limits of human judgment make training LLMs for creativity especially challenging. As a solution, we train LLMs on Codenames, a word-association game that exercises the two central axes of creativity, divergent and convergent thinking, while yielding objectively verifiable outcomes. This verifiability lets us bypass human judgment and train with Reinforcement Learning with Verifiable Rewards (RLVR). We train Qwen3-1.7B, 4B, and 8B models and evaluate them on ten creativity and four reasoning benchmarks. We find that the precision-diversity trade-off is scale-dependent: the 8B model prioritizes creativity over precision, while the 1.7B and 4B models gain reasoning precision at the cost of creativity. Concretely, the 8B model shows modest but consistent creativity gains (8 of 10 benchmarks) with only minor reasoning degradation, whereas the smaller models achieve substantial gains on reasoning tasks. Our study presents a scalable and effective solution to train LLMs for creativity.

Mathematical reasoning benchmarks are vital for evaluating large language models (LLMs), but many are static and repeatedly exposed through public evaluation and training pipelines, making it difficult to separate genuine reasoning from memorization. Meanwhile, manually constructing new math problems with reliable answers remains costly. We introduce ReverseMath, a scalable method for generating new math problems through answer inversion. Given a problem and its answer, ReverseMath masks a numerical value in the original problem, treats the original answer as a known condition, and rewrites the problem so that the masked value becomes the new answer. The generated problem reverses the original input-output relation, making its answer known by construction. We study ReverseMath for both evaluation and training. For evaluation, paired original/reversed problems reveal substantial behavioral shifts: models sometimes fail on reversed problems and even incorrectly output the original answer, suggesting memorization-like behavior. For training, ReverseMath provides automatically labeled reversed problems as data augmentation for reinforcement learning (RL). Experiments show that including ReverseMath-generated data improves mathematical reasoning performance across multiple benchmarks, demonstrating its value as both an analysis tool and a scalable source of verifiable training data.

Complex Query Answering (CQA) is a fundamental knowledge representation and reasoning task over incomplete knowledge graphs (KGs). Answering existential first-order queries with $k$ free variables (i.e., $\text{EFO}_k$ queries) is a crucial yet challenging problem, as it requires ranking answer tuples in $\mathcal{E}^k$, where $\mathcal{E}$ denotes the entity set of a KG. This quickly becomes intractable as $k$ grows. Consequently, existing benchmarks and methods rely on marginal rankings over individual variables; however, marginal rankings are a poor proxy for the true joint ranking of tuples. Building on neural symbolic search for $\text{EFO}_1$ queries, we propose Neural Scalable Symbolic Search (NS3), a budgeted framework that approximates joint ranking without enumerating $\mathcal{E}^k$. NS3 (i) answers marginalized sub-queries to obtain necessary candidate sets, (ii) merges multiple free variables into hypernodes whose domains are pruned and controlled by a dynamic budget $B$, and (iii) progressively reduces an $\text{EFO}_k$ query to an $\text{EFO}_{k-1}$ query over a budgeted reduced domain. Across three standard KG datasets, NS3 substantially improves joint ranking performance while retaining strong marginal accuracy. We further release a joint-ranking benchmark that extends existing $\text{EFO}_1$ datasets to $k=3$, enabling systematic evaluation of multi-variable queries. Our code is provided in https://github.com/HKUST-KnowComp/NS3_KDD2026.

We systematically investigate all 25 public ARC-AGI-3 games and find that every one is reachable through non-intelligent strategies: 10 in a single blind step, 5 after one probing action, 1 via repeated ACTION1 presses, 1 via diverse exploration, and 8 via single repeated actions with sufficient budget (50-200 steps). A library-level null-coordinate vulnerability additionally bypasses 18 games in 1 step. This benchmark critique implies the public evaluation set cannot discriminate intelligent exploration from trivial heuristics - the private 55-game evaluation is the only genuine intelligence test. Against this backdrop, we present AERA (Adaptive Epistemic Reasoning Agent), a three-phase (EXPLORE / VERIFY / PLAN) agent achieving RHAE=0.2116 (4/25 solved) on these 25 games with Qwen2.5-0.5B, while random and no-explore baselines score 0.0000. We formalise AERA through a Speed--Depth trade-off framework: under a convexity assumption (proved for a class of environments in the Appendix), RHAE's quadratic form emerges as a second-order penalty for deviating from the Pareto frontier between action efficiency and information gain. Contributions: (i) a benchmark validity analysis showing that current interactive reasoning benchmarks fail to measure the exploration they claim to require, and (ii) the EXPLORE-before-PLAN framework and model-capability x exploration interaction. The linked code track entry achieves RHAE=0.30 on the full 55-game private evaluation. Code: CC0.

High-quality data is scarce in large language model (LLM) training, yet how to schedule its use jointly with training dynamics lacks theoretical guidance. We extend functional scaling laws by incorporating a data-quality dimension, and solve the joint data-quality and batch-size scheduling problem in asymptotic closed form. The solution reveals two regimes and a dual role of high-quality data. In the noise-limited regime, high-quality data should be used as a signal amplifier: lowering the batch size converts cleaner data into more signal without amplifying noise. In the signal-limited regime, it should be used as a noise suppressor: late placement reduces terminal noise without sacrificing signal accumulation. Existing curriculum-style pipelines primarily exploit the second role by placing cleaner data late, but miss the first role because conventional decay schedules reduce update intensity exactly when high-quality data becomes available. Guided by this, we propose Drop-Stable-Rampup for LLM midtraining: upon the quality transition, drop the batch size, hold it stable to accumulate signal, then ramp up to suppress terminal noise. On a 15B Mixture-of-Experts model midtrained on 108B tokens, Drop-Stable-Rampup improves average accuracy over Warmup-Stable-Decay (WSD) by +1.70 and over Cosine-decay by +2.98, with particularly large gains on mathematical reasoning benchmarks such as GSM8K (+4.23) and MATH (+2.80).

Multi-agent LLM systems improve reasoning by combining outputs from multiple agents, but interaction-heavy methods can introduce error propagation and high communication overhead. When agents exchange raw responses or reasoning traces, incorrect intermediate reasoning may be adopted and amplified, leading to confident but wrong consensus; multi-round communication also increases token consumption, latency, and inference cost. In this paper, we propose a controlled-communication coordination framework named DarkForest. DarkForest first keeps agents independent, so each agent produces an answer without seeing the others' outputs. It then parses the raw responses into structured candidate records, groups semantically equivalent candidates into clusters, and estimates a calibrated belief distribution over these clusters using agent reliability, confidence, parse quality, support-pattern reliability, and independence corrections. A coordinator receives only policy-permitted evidence from this belief state with controlled communication. Experiments on six reasoning benchmarks show that DarkForest achieves leading overall quality, improves the strongest baseline by up to 30.7% on benchmark metrics, and reduces token consumption by up to $6.5\times$ compared with communication-heavy baselines.

Test-time scaling improves language model reasoning by spending additional compute to explore multiple solution trajectories. The key challenge is to maximize accuracy while minimizing the total number of generated tokens during reasoning. Recent PRM-guided methods score intermediate prefixes to steer this search, but most are frontier-only: they keep only the current active prefixes and irreversibly prune or resample away the rest using noisy PRM scores. This can cause premature commitment, diversity collapse, and the loss of prefixes that still admit correct continuations. We introduce stochastic backtracking over a persistent pool of historical prefixes, allowing test-time compute to revisit previously generated states instead of only expanding the current frontier. To make this efficient, we propose two complementary mechanisms. Subpool Selection strengthens greedy PRM-guided search by applying Top-N selection within random subpools, giving historical prefixes a chance to bypass over-scored frontier candidates. Power Backtrack Sequential Monte Carlo extends SMC-style resampling to the persistent pool using powered PRM scores and mixture-corrected weights. Across mathematical reasoning benchmarks and model scales, our methods consistently achieve higher accuracy per token count, and the same level of accuracy using only a fraction of the token count in comparison to strong PRM-guided baselines, demonstrating that persistent-pool stochastic backtracking provides a simple and effective way to improve the accuracy-token trade-off in test-time scaling.

Reasoning-enabled LLMs perform strongly on medical reasoning benchmarks, but it remains unclear whether these gains transfer to structured clinical documentation; we investigate this question using SOAP note generation from clinical dialogue in a source-aware benchmark spanning OMI Health, ACI-Bench, and PriMock57. We evaluate GPT-5.4, DeepSeek-V4-Flash, and Gemma-4-E4B in a controlled 2x2 design that independently toggles provider-native reasoning and same-source retrieval-augmented generation (RAG). Outputs are assessed using seven automatic metrics alongside two reference-aware LLM judges. Both evaluation approaches agree that a non-reasoning GPT-5.4 configuration achieves the highest overall quality, while DeepSeek-V4-Flash performs best among reasoning-enabled configurations. Enabling reasoning significantly degrades GPT-5.4 performance across all three datasets, whereas same-source RAG yields smaller, model-dependent improvements. Overall, the findings indicate that stronger reasoning capability should not be assumed to improve fidelity-sensitive SOAP note generation without dedicated, task-specific evaluation.

Despite remarkable progress on reasoning benchmarks, current LLM evaluation practice remains anchored to final-answer correctness, providing limited insight into how models reason, how reliably they behave under contextual variation, or how efficiently they reach conclusions. This paper proposes a unified multi-dimensional framework for measuring LLM reasoning quality from a behavioral perspective, operationalizing six theoretically grounded dimensions rooted in cognitive science: Correctness (CQ), Consistency (CS), Robustness (RS), Local Logical Coherence (LS), Efficiency (ES), and Stability (SS). The framework introduces deployment-aware aggregation, enabling context-specific model selection beyond accuracy-based leaderboards. Experiments across multiple LLMs and benchmarks reveal behaviors systematically concealed by single-metric evaluation, including the orthogonality of local logical coherence and correctness, deployment-context-dependent ranking inversions, and non-trivial dimensional profiles in small locally-deployed models. Discriminant validity analysis confirms that the proposed dimensions capture largely non-redundant signals. The resulting pipeline provides a foundation for diagnosing LLM reasoning behavior across deployment contexts, with domain-specific validation as a direction for future work.

Standard accuracy on binary reasoning benchmarks hides critical failure modes: prior collapse, inconsistency under paraphrase, and inability to reason about parameter-dependent dynamics. We present ChaosBench-Logic v2, a 40,886-question benchmark over 165 dynamical systems with 27 FOL predicates and 78 axiom edges, together with CARE (Calibration- and Adversarial-Robust Evaluation), a protocol that surfaces these pathologies. Evaluating 14 models, we find that regime-transition reasoning remains near random (MCC = 0.05) even for frontier models, whereas FOL deduction with given premises reaches MCC = 0.52. Per-family decomposition shows that the proprietary-model advantage concentrates on cross-indicator (+0.40) and consistency tasks, while open-source Qwen 2.5-32B dominates indicator diagnostics (0.91 vs. 0.45). Two models exhibit negative MCC on bifurcation questions, confirmed as systematic anti-correlation via confusion-matrix analysis.