Reasoning Benchmark

Self-evolving methods reduce the need for human-annotated trajectories by allowing tool-using agents to generate their own training data. Yet existing methods typically separate trajectory generation from evaluation, relying on static verifiers that cannot adapt to emerging failure modes or self-consistency signals that may reinforce errors shared across trajectories. Jointly adapting planning, execution, and evaluation offers a promising alternative, but introduces a fundamental coordination challenge: each component continuously changes the data or feedback used to train the others. We address this challenge with \textbf{UnifiedPlayers}, a cooperative framework comprising a Planning Player that generates tasks, an Execution Player that produces multi-turn trajectories with Python tool calls, and an Evaluation Player that constructs executable verifiers. We design role-specific rewards that coordinate the three players toward a shared learning objective under GRPO. Across two model backbones and twelve reasoning benchmarks, UnifiedPlayers outperforms the strongest prior baseline by at least 3.5% on mathematical reasoning and 3.9% on general reasoning tasks. Moreover, the learned verifier achieves 84.2% adversarial detection accuracy, while its reward signal exhibits 2.03$\times$ higher per-question variance than a self-consistency baseline, providing more discriminative verifications. These results highlight cooperation among specialized players as a promising path toward self-enhanced tool-integrated agents.

Characterizing LLM reasoning remains an open challenge, as many existing benchmarks isolate specific reasoning skills, rely on external knowledge, or are costly to extend. We introduce PetriBench, a compact, fully self-contained, and scalable benchmark for evaluating LLM reasoning over dynamic state spaces using Petri nets, a mature formalism for modeling real-world concurrent and distributed systems. PetriBench organizes reasoning into four task families varying by scope and temporal horizon, with Easy, Medium, and Hard levels generated by increasing structural complexity and evaluated against exact ground truth. Across a diverse set of proprietary and open-weight models, accuracy decreases consistently with difficulty, while harder instances expose increasingly distinct task-specific capability profiles. Additional analyses show that test-time compute improves performance but interacts differently with different reasoning tasks, and that procedural generation yields smooth scaling with structural complexity. Together, these results show that PetriBench provides a unified and extensible setting for probing the strengths, limits, and scaling behavior of LLM reasoning.

Verifier-based selection improves LLM performance by generating multiple candidate solutions and using a verifier to select the most promising one. However, existing methods typically treat verification only as a ranking step and discard its feedback once a fixed candidate pool has been evaluated. In this paper, we ask whether verification can also improve the candidate set itself. To this end, we introduce LLM-as-an-Improver and propose Verify--Repair--Reselect (VRR), which uses verification feedback to generate and reselect improved candidates. VRR retains the initial winner while conditionally generating three complementary alternatives: repaired versions of the winner and runner-up, and a solution based on a new approach. It filters invalid and duplicate candidates using only inference-time information and then reselects the final answer under the original evaluation criteria. Across diverse models and code-generation and reasoning benchmarks, VRR improves over fixed-pool verifier-based selection in many settings and can recover correct solutions even when all candidates in the initial pool are incorrect. These results highlight a broader role for LLMs as improvers: verification feedback can not only select among existing solutions but also construct stronger candidates beyond the initial pool.

Fallacy-detection benchmarks pair fallacy classes with a single "valid" or "none" class that takes everything data collection did not label as a fallacy. A detector has two jobs, deciding whether an argument is fallacious and naming which fallacy it commits, and the false-positive rate is meant to measure the first. We show that what these benchmarks actually score is scheme recognition, the ability behind the second job. Their own test sets already show it: when a classifier misses a fallacy, the error lands on "none" rather than on another fallacy type, so detection is failing while classification holds. The reason is what the valid class lacks. The negatives that separate the two jobs are correct arguments using the same argumentation scheme as a fallacy, and they are scarce: nearly absent from the four benchmarks we examined, and rare even under deliberate search. A detector is therefore never tested where recognizing a scheme and judging its use come apart, and can pass on recognition alone. We construct the missing arguments, together with a control condition from the same pipeline that differs only in scheme, so whatever generation contributes, it contributes to both. The classifier labels the scheme-matched negatives as the source fallacy, and labels the wrong-scheme negatives as the scheme they actually use 85.9% of the time and as the source type 0.4%. The classifier has learned which scheme an argument uses, not whether it uses it correctly. The over-flagging follows: a model that scores 16.6% on CoCoLoFa's own valid class flags 58.9% of the constructed arguments. The same dissociation appears in three zero-shot LLM detectors that never saw these benchmarks. We release the items as Scheme Foils. A reported false-positive rate should not be trusted as a measure of detection until the valid class has been audited for scheme-matched coverage.

Organizations increasingly use oversight loops where one large language model (LLM) audits another's outputs alongside procedural traces of claimed steps. A common concern about such LLM-as-a-judge pipelines is that detailed traces make overseers gullible. Using signal detection theory, we audit five LLM overseers on 19 compliance tasks (4,551 analyzed judgments), varying only trace detail and evidence labeling. With disconfirming evidence always visible, error detection remains near ceiling. Instead, elaborate traces shift the decision criterion toward rejection, increasing false alarms in susceptible overseers. Without option labels, human-validated reason coding shows about 60% of false alarms cite an inability to tie evidence to its option. Labels eliminate this stated reason, yet residual rejection of correct work persists in those overseers and rises with trace detail. Procedural traces thus act as governance artifacts that shape oversight decisions. AI auditors should be evaluated by their decision criterion and false-alarm behavior, alongside accuracy.

Real-world agents fundamentally require persistent non-parametric knowledge for dynamic reasoning, i.e., long-term memory and retrieval-augmented generation. While graphs have shown reliable advantages in providing structured evidence, the sparse graph representations naturally restrict machine readability and semantic density required for complex agentic workflows. Driven by this limitation, the entire industry is witnessing a paradigm shift from traditional sparse graphs to LLM Wiki, an agent-native knowledge representation that couples dense document contexts with markdown files containing multi-layered topological linkages. However, parameterizing such rich semantics is challenging to encode dense textual contexts using traditional sparse graph embeddings. Moreover, learning LLM Wiki with existing graph encoders could overwhelm distributed system overheads that hinder deployment in large-scale commercial scenarios. To this end, we propose a novel paradigm Wiki Foundation Model, i.e., WFM, tailored for scalable, agent-native representation and retrieval. Specifically, (i) we formalize a Wiki Graph schema that seamlessly bridges fine-grained structures with dense contexts, maintaining explicit topologies alongside continuous semantics; (ii) A query-conditioned attentive aggregation is tailored for rich wiki message passing and explicit attention variance regularization; (iii) We engineer an infrastructural NCCL boundary exchange protocol that hoists static partition indices and leverages fixed-shape GPU-to-GPU collectives, bypassing CPU serialization and memory copy overheads. Extensive evaluations across five long-term agent memory and multi-hop reasoning benchmarks demonstrate the remarkable performance of WFM, while achieving a 10.5 times training acceleration on distributed clusters.

LLM-based multi-agent systems can fail even when communication succeeds because agents do not correctly track their peers' roles, knowledge, or intentions. We investigate whether such inter-agent misalignment cases, labelled FC2 in MAST-Data, can be converted into functional partner-state reasoning items. ToMAS applies four explicit convertibility criteria to diagnosed execution traces. A full conversion pass over 242 eligible non-AG2 training traces produced 39 CLEAN items. In an 18-trace reliability pilot, two annotators achieved 94.4% raw agreement and Cohen's kappa = 0.92. We then used the converted items as binary rewards in a small-scale GRPO feasibility experiment with Qwen2.5-1.5B. On a 28-item held-out Magentic GAIA diagnostic, every evaluated condition exceeded the ROUGE-L threshold on the same 2 of 28 items. Post-hoc adapter checks show why: under the learning rate used, the LoRA update remained numerically negligible (max abs Delta W about 7e-6), so all conditions decode identically to the untrained checkpoint. The experiment therefore does not show a training effect and cannot establish one; it reports an executable pipeline together with two limitations that any conclusive study must address: a provenance gap between the training and evaluation items, and lexical-overlap scoring. ToMAS provides a preliminary rubric and pipeline for converting diagnosed coordination failures into trainable partner-state reasoning items and identifies the requirements for a conclusive matched-domain evaluation.

Large language models (LLMs) have demonstrated impressive performance in various natural language processing tasks, yet their ability to perform multi-step logical reasoning remains an open challenge. Although Chain-of-Thought prompting has improved logical reasoning by enabling models to generate intermediate steps, it lacks mechanisms to assess the coherence of these logical transitions. In this paper, we propose a novel, lightweight evaluation strategy for logical reasoning that uses query-key alignments inside transformer attention heads. By computing a single forward pass and extracting a "QK-score" from carefully chosen heads, our method reveals latent representations that reliably separate valid from invalid inferences, offering a scalable alternative to traditional ablation-based techniques. We also provide an empirical validation on multiple logical reasoning benchmarks, demonstrating improved robustness of our evaluation method against distractors and increased reasoning depth. The experiments were conducted on a diverse set of models, ranging from 1.5B to 70B parameters.

Frontier gains in language-model reasoning come from reinforcement learning on reasoning traces and are concentrated in domains with a cheap, sound verifier. We argue the field's binding constraint is the verification gap: no scalable, incorruptible reward for reasoning outside formal domains. We make four contributions. (1) Theory: in a joint-Gaussian model of best-of-N selection, verifier-gold correlation rho is the exact exchange rate between test-time compute and capability, and an unsound verifier pays a polynomial penalty N^(1/rho^2); a margin-free copula form predicts realized soundness of real LLM judges to 4% median error. (2) Demonstration: in program-synthesis testbeds with executable ground truth, including a pre-registered scaled replication, unsound verifiers lose Soundness-under-Pressure as optimization grows (0.94 to 0.32 at N=4096) while a sound verifier improves monotonically; reality-anchored settlement beats a frozen verifier under i.i.d. and adversarial pressure, driving the hacking gap from ~0.27 to ~0; soundness scales log-linearly with settled labels, with on-policy settlement ~10x more label-efficient than random labeling. With real LLM judges and unit-test execution as gold, a weak judge loses soundness under best-of-N (p<0.001), a stronger judge is more robust, and selection alone manufactures +0.53 hacking gaps from honest samples. Under real GRPO training, a frozen reward model traces the full overoptimization curve (executed reward collapses 90%) while the same model refit on a 10% settlement stream preserves 6x the executed reward. (3) Paradigm: proof-carrying cognition, where reasoning steps are typed probabilistic claims priced by a self-built world model trained only on held-out reality and settled by proper scoring rules. (4) Benchmark: we specify Soundness-under-Pressure as the headline metric for a reality-settled reasoning benchmark.

Although MMLU is widely adopted as a benchmark for calibrating general AI capabilities, we psychometrically demonstrate that its aggregate score primarily evaluates a model's factual retrieval capacity rather than its reasoning ability. By calibrating item difficulty for 1,000 open-weights language models over 14,042 MMLU test items using Item Response Theory, we show that evaluating both abilities via a single test is inherently flawed. Difficulty is then regressed on a deterministic, text-extractable framework of structural complexity. Applying a joint Wald test with subject-clustered covariances demonstrates that the MMLU conflates fundamentally separable constructs. The mapping from structural complexity to difficulty is not invariant across the benchmark's STEM and non-STEM partitions. This finding has practical consequences. Aggregate leaderboard ranks track non-STEM accuracy more closely than STEM accuracy, so selecting a Top-50 model on the aggregate for a reasoning-intensive deployment displaces roughly 22% of the STEM-appropriate choices. Furthermore, when controlling for the multiple-choice guessing floor natively inside the response model, we find that higher-ability models continue to degrade more steeply under increased reasoning depth. The MMLU aggregate therefore weights retrieval capacity and reasoning stability unequally, inadvertently favoring models optimized for retrieval. We release our deterministic framework as a reproducible auditing instrument and recommend disaggregated reporting.

Leading benchmarks for formal theorem proving with large language models are small collections drawn from competition math, such as the IMO and Putnam, that poorly represent field-specific applications. We introduce StochBench, a Lean 4 benchmark of 450 graduate stochastic-processes problems at varying abstraction levels, each paired with its natural-language source. Addressing a field underrepresented in Mathlib, it covers finite and countable Markov chains, renewal processes, random walks, martingales, stopping times, queues, Brownian motion, stochastic calculus, weak convergence, and Poisson and continuous-time Markov processes. Our Opus 4.8-based agent achieves a 34.9% proof rate (157/450) under a 15-minute per-problem limit. StochBench better represents domain-specific applied mathematics while remaining challenging for advanced provers.

Self-play is an effective paradigm for language-model self-evolution, but without guidance, solver performance can plateau or decline across rounds. Unguided methods steer question generation with signals such as difficulty, learnability, or diversity. These signals keep questions challenging and varied but do not specify which unresolved reasoning weaknesses later rounds should target. Guided methods obtain direction from external task resources, including human examples, document corpora, or specified difficulty targets, and therefore rely on task information supplied outside the self-play loop. We show that the needed direction can instead be derived from the solver's own failure history. We introduce DiagEvo, whose diagnostician extracts recurring error causes from this history and stores them in a hierarchical error-cause memory. The memory groups related causes under skill nodes and tracks each as Active or Mastered according to self-consistency on targeted questions. The challenger uses these states and recurrence counts to balance cause-targeted generation with free exploration. Double-confidence filtering retains intermediate-difficulty questions only when the most common solver answer has a clear vote lead. DiagEvo derives its curriculum from information produced during self-play, without external task resources. With the default 4B diagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine benchmarks for each of the three solvers: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. On Qwen3-8B, it reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero. Its mean accuracy across all nine benchmarks is 57.4%, 1.1 percentage points above DARC. Ablations show that the hierarchical error-cause memory and double-confidence filtering both contribute to these gains.

When a large language model fails a reasoning task, it is often assumed to lack the underlying capability. However, this conflates a genuine absence of reasoning with a late-stage output bottleneck. We observe a consistent readout gap across diverse reasoning benchmarks: hidden-state probes successfully decode correct answers even when native sequence scoring completely collapses due to structural biases. To test whether instance-specific logic survives this collapse, we introduce a diagnostic protocol using a minimal, target-label-free additive correction. Fitting just two parameters on as few as 25 unlabeled examples recovers 9--34 accuracy points for Qwen3.5 models, transferring successfully to OLMo-2-1B and Llama-3.1-8B. Crucially, these recovered decisions persist on hard instances unresolved by simple lexical overlap and significantly exceed count-preserving permutation baselines. Our results show that many apparent zero-shot reasoning deficits are expression failures masking intact internal logic, urging a narrower interpretation of benchmark evaluations.

Recent advances in Large Multimodal Models (LMMs) have greatly improved video understanding, yet their ability to reason about human-centered social situations remains limited. Existing benchmarks typically rely on videos with a single observed trajectory, making it difficult to determine whether models truly understand social dynamics or merely exploit recurring narrative patterns. We introduce SocialReasonBench, a video multiple-choice QA benchmark for evaluating socially grounded reasoning in scenarios derived from interactive narratives. Built from gameplay videos of Detroit: Become Human, the benchmark leverages branching storylines where player decisions lead to alternative social outcomes that can be checked against the game's own script, flowchart, and recorded branches. We develop a multi-agent curation pipeline that localizes socially meaningful clips, grounds answer labels in game-state signals, and generates theory-guided questions with diagnostic distractors. SocialReasonBench covers seven reasoning dimensions, including intent recognition, emotional empathy, moral dilemma, counterfactual reasoning, and causal antecedent. Experiments on contemporary LMMs show that models perform reasonably well on basic social understanding but struggle with counterfactual and causal reasoning. Further ablation and diagnostic error analyses reveal that models often depend on incomplete modality cues and fall into reasoning traps such as visual shortcuts, highlighting a gap between observable event recognition and deeper reasoning over latent social states.

Large Reasoning Models produce Long Chains-of-Thought (LCoTs) which involve breaking down the problem into smaller reasoning steps before reaching the conclusion. However, these steps often contain contradictions, unsupported inferences, or irrelevant steps, even when the final answer is correct. We propose Long Chain-of-Thought Graph Verifier (LCoT-GV), a graph-based framework that represents LCoTs as reasoning graphs. Each node in the graph represents a reasoning step and the edges encode semantic and logical relations. A Graph Attention Network is then trained to predict chain-of-thought correctness from the reasoning graph. We construct a new graph-oriented verification dataset from multiple reasoning benchmarks for question answering in various domains. The results show that our method is competitive with the most similar approaches.

Large language models (LLMs) often produce fluent but weakly grounded conclusions when reasoning over non-interactive, long-form narratives. A central failure mode is that unsupported intermediate hypotheses can enter the reasoning trajectory and contaminate subsequent inference, especially when evidence is scattered across distant parts of the story. To address this problem, we propose EVAR, an evidence-validated hypothesis admission framework for budget-aware narrative reasoning. EVAR first compiles the narrative into an immutable evidence store of source-linked atomic claims and assigns an instance-specific inference budget from unresolved gaps and uncertainty signals. During refinement, EVAR directly proposes candidate hypotheses for unresolved gaps, constructs hypothesis-conditioned validation challenges, and verifies each candidate against the locked store before admission: supported hypotheses enter the answer-supporting state, unverifiable ones are quarantined, and contradictory ones are discarded. A sufficiency-based stopping mechanism further avoids unnecessary refinement. Experiments on NarraCrime and multiple public reasoning benchmarks show that EVAR improves both task performance and evidence faithfulness while maintaining controllable inference cost.

Continuous self-improvement requires an ever-expanding pool of self-generated, diverse, adaptive goals. For language agents, existing training environment pools (hand-curated, statically synthesized, or frozen-verifier) keep the goal distribution fixed as the learner scales. We introduce SPADE (Self-Play in Adaptive Synthetic Executable Environments), a self-play RL framework in which a single LLM plays two roles: an Environment Designer that writes complete, long-horizon training environments as executable code with an OpenAI Gym-style reset()/step() interface, and a Reasoning Agent that learns to act in them. Each is a stateful, multi-turn environment (state transitions, reward functions, and verification code), so one interface spans reasoning problems and multi-step agentic tool use. The Reasoning Agent's regret is estimated using the gap between its reward with and without privileged hints; in optimizing this regret signal the Environment Designer learns to target environments at the edge of the agent's capabilities while keeping them feasible. Through extensive experimentation, we find several components critical to success: grounding the Environment Designer on documents sampled from a large pretraining corpus, and giving it an accumulated environment memory. Scaling to 30B-parameter models, SPADE improves over the strongest fixed-environment baseline by +5.3 on average across eight held-out math, science, code, and reasoning benchmarks, and lifts the tool-use setting by +5.7 on BFCL-v4 multi-turn and +13.9 on ACEBench-Agent; on the games setting, the margin over the strongest baseline grows with model scale. By making environment design itself a learnable component, SPADE takes a concrete step toward open-ended self-improvement.

The evaluation of LLM reasoning is moving from final-answer accuracy to process-level assessment, yet existing methods still fail to capture how models plan reasoning paths and allocate reasoning resources--that is, how they organize search. Prior process-level methods focus on the coherence and redundancy of chain-of-thought (CoT), and most benchmark tasks have a single objective solvable by static capabilities such as derivation and tool use, leaving search organization unmeasured. We introduce TsuGO, a process-level reasoning benchmark for evaluating Search Efficiency in LLM reasoning through Go life-and-death problems. These problems provide closed and verifiable solution spaces with an inherent adversarial structure, making candidate generation, response checking, branch comparison, and backtracking necessary parts of reasoning rather than incidental trace patterns. By constraining the solution space, TsuGO disentangles domain knowledge from search organization, parses CoT into a structured search tree, and reports Search Efficiency together with Token Efficiency and other diagnostic metrics and visualizations. Experiments show that current LLMs remain far from stable tsumego solving: stronger models succeed by finding the correct candidate earlier and sustaining effort on productive branches, but most models still behave much closer to unguided search algorithms than to neural-guided KataGo. Longer CoT or higher Token Efficiency does not necessarily imply better search. Our results identify search organization and reasoning-resource allocation as missing dimensions in LLM reasoning evaluation.

Recent work on distillation transfers the capabilities of large models to smaller ones often by updating the latter's parameters, through teacher forcing, on-policy distillation, and related training-time methods. In this paper, we ask whether such transfer can instead occur at test time. We study strong-to-weak scaffolding: whether a stronger builder model can construct inference-time harnesses that help a weaker target model solve tasks more reliably without any parameter updates. Using four representative Theory-of-Mind benchmarks, each builder model uses 5% of the data as a validation set to iteratively refine its harness over multiple rounds, after which the finalized harness is evaluated on the full test set. Empirically, this form of test-time capability transfer is highly effective, nearly doubling average target-model performance from 0.49 to 0.91. Our analysis shows that the gains come primarily from offloading unstable model reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement, rather than from encouraging the target model to reason more extensively or sample more broadly. We further find that builder-model reasoning effort improves harness quality monotonically, platform effects are modest relative to the builder model's own capability, and weaker target models receive the largest gains. These results suggest that inference-time harness design is an important complement to conventional training-time distillation, enabling strong models to transfer cognitive structure to weaker models without retraining.

Agents deployed in enterprise settings must reason across structured APIs and document collections, yet existing benchmarks evaluate these capabilities in isolation. We introduce VAKRA (e\textbf{V}aluating \textbf{A}PI and \textbf{K}nowledge \textbf{R}etrieval \textbf{A}gents), a benchmark of over $8{,}000$ executable APIs across $62$ domains with tasks spanning three settings of increasing difficulty: diverse API interaction styles, multi-hop reasoning over structured APIs, and multi-source reasoning with natural-language tool-use policy constraints. Correctness is verified by re-executing predicted tool calls against live APIs, accommodating multiple valid paths. Using a fixed ReAct harness to isolate model capabilities from agent architecture, we evaluate frontier and open-weight models and find that even the best model achieves only 70.4% on single-hop endpoint-style tasks and drops to 50--51% on compositional APIs; performance degrades by over 50% as reasoning depth increases, and policy-constrained questions expose severe failures (as low as 2.4% on unanswerable queries). Trace analysis shows failures concentrate at language-mediated reasoning - entity disambiguation, cross-source grounding, rather than tool invocation mechanics. Code is available https://github.com/IBM/VAKRA. Dataset is available https://huggingface.co/datasets/ibm-research/VAKRA

Standard evaluation of large language models assumes stable model rankings across inference conditions. We challenge this assumption by varying the token generation budget, i.e., the maximum tokens a model may produce, across seven levels (64--4,096), evaluating four models on three reasoning benchmarks (56,476 inferences). We report four findings: (i) 3--19% of items exhibit non-monotone behavior (accuracy decreasing with more budget), even after controlling for truncation, and this phenomenon is model-specific (cross-model overlap: 6--14%). (ii) Model rankings reverse across budgets on all benchmarks ($p {<} 0.01$, McNemar). (iii) Oracle analysis reveals model complementarity up to $+27.8$pp, most pronounced at constrained budgets. (iv) A budget-aware router captures 14.1% of the oracle gap cross-domain; budget features help within-domain ($+1.6$ to $+5.7$pp) but are domain-specific and hurt transfer ($-1.2$pp). These results argue for budget-conditioned evaluation protocols.

Tool-augmented vision-language models increasingly "think with images": they call crop, zoom, or code tools and reason over the returned pixels. However, recent work using blind tests, gain decompositions, and attention analyses has shown that returned images contribute little, raising the question: if pixels do not carry the gain, what does? We hypothesize that the load-bearing signal is the structured text emitted before any returned pixel arrives: tool name, coordinates, target description, and intent. This textual scaffold encodes where to look and what to find. We introduce TextCall (call-but-no-return) to test this: it keeps the scaffold but replaces returned images with the text placeholder [Image output skipped]. Three studies support the hypothesis. (i) Non-necessity of returned pixels: across LoRA, full fine-tuning, and RL, TextCall matches or exceeds full thinking-with-images; under RL it preserves tool use at the reported checkpoint, avoiding the failure mode where, under matched settings, seeing the returned image causes the model to stop calling tools and answer directly. (ii) Sufficiency of the scaffold: on matched training queries, scaffold-only input yields equivalent accuracy to returned-image input. (iii) Component specificity: decomposing the scaffold into reasoning text and spatial code shows both components contribute, with the dominant one varying by task. Together these results support the Tool-Call Scaffold Hypothesis: in current thinking-with-images distributions, the active signal is the structured text emitted at tool-call time; the returned image is a redundant carrier. TextCall preserves accuracy while reducing latency by 29-46% and eliminating tool-execution API calls. Our claims hold for current thinking-with-images benchmarks; constructing tasks where pixels are genuinely load-bearing remains an open direction.

We introduce TCS-Bench, a benchmark for evaluating Large Language Models (LLMs) on research-level Theoretical Computer Science (TCS) proof generation. TCS-Bench consists of theorem-proving tasks from papers published at top theoretical computer science venues (STOC, FOCS, and SODA). Each task provides the necessary context to derive a self-contained proof for a target result. We evaluate state-of-the-art models on this benchmark. We verify the correctness of generated proofs via a verification agent, and further benchmark the verifier against human-expert proof judgements on a set of target statements and generated proofs pairs. Our reference verifier achieves over 90% accuracy on the expert labeled set.

Reinforcement learning (RL) has become a central post-training paradigm for eliciting reasoning capabilities in large language models, yet uniform task sampling allocates compute without regard to differences in how tasks respond to optimization. Existing task-valuation methods mostly rely on snapshot-based signals such as current pass rate or reward, which estimate how solvable a task is under the current policy. However, tasks with similar current solvability can still differ substantially in how positively they respond to further training. We study this residual axis as task learnability: a regime-conditional measure of expected positive response to continued training under a fixed RL post-training regime. By analyzing per-task reward trajectories, we find that learnability is reproducible across independently sampled training contexts and predictive of downstream utility. To make this signal practical before training begins, we propose TrajVal, a lightweight probe-based estimator that approximates per-task learnability from a short probe run and two endpoint evaluations. TrajVal can be used either as a standalone static prior for task sampling or as a multiplicative prior for existing online schedulers. Experiments on mathematical and logical reasoning benchmarks across multiple model scales show that TrajVal improves data efficiency over uniform sampling and provides complementary gains when combined with online scheduling methods.

On-policy self-distillation (OPSD) improves the reasoning abilities of LLMs by internalizing privileged context into model parameters through self-distillation. Two recent research lines promote vanilla OPSD by choosing which tokens to learn from and by controlling how much privileged information the teacher receives, respectively. However, we show that each line optimizes one variable while holding the other fixed, which leads to a suboptimal solution. We argue that the two variables are coupled through the student's learning capacity: the privileged information sets the per-token divergence the teacher prescribes, while token weighting selects which of these the student must absorb. We formalize the two lines of work into a unified optimization framework, which maximizes the aggregate teacher--student divergence, subject to a budget on the aggregate learning difficulty the student can absorb. Under this modelling, we propose Unified On-Policy Self-Distillation (USD), a lightweight online algorithm to solve the Lagrangian. USD reveals that a single dual variable governs both decisions: at one price for learning difficulty, it simultaneously sets the token-selection threshold and the direction of privileged-information adjustment, keeping supervision matched to the student's evolving capacity. Through extensive experiments, USD consistently demonstrates superior performance over OPSD and token- and PI-side baselines across various model scales on various reasoning benchmarks. Code is available at https://github.com/lauvlalala/USD.

Test-time scaling is often implemented by spending more compute along one axis: sampling more solutions, extending a chain of thought, or applying a stronger evaluator. Under a fixed inference budget, these choices compete. This paper formulates test-time reasoning as a compute-allocation problem in which a system must decide whether the next unit of compute should be spent on generation, verification, or stopping. We introduce CoBa, a compute-balanced routing policy that first obtains a small set of candidates, applies cheap verification broadly, and routes uncertain or high-value candidates to stronger verification. On 3,129 example-generator evaluations spanning MATH-500, AIME 2024/2025, AMC 2023, and procedural symbolic reasoning, CoBa-Routed-Strong reaches 85.13% macro accuracy, statistically matching a self-evaluation weighted-voting proxy at 85.20% while using 49.1% fewer parameter-weighted tokens. It also matches best-of-16 majority voting within 0.01 macro-accuracy points while using 58.9% fewer parameter-weighted tokens; paired tests retain a small best-of-16 edge at substantially higher cost. Paired bootstrap tests show significant gains over single-sample decoding, while the remaining gap to the pool oracle exposes headroom for sharper routing. For local reasoning systems, test-time scaling becomes a question of where the next computation is most valuable.

Vision-Language-Action (VLA) models have recently achieved impressive performance for end-to-end autonomous driving, yet existing approaches are primarily designed for an individual single autonomous driving agent with limited support for cooperative perception, reasoning, and planning. We present Cooperative Multi-agent Unified Driving with Reasoning (CMU-Drive), a closed-loop end-to-end benchmark for evaluating cooperative autonomous driving with multiple connected autonomous vehicles (CAVs) operating in safety-critical driving scenarios with background traffic participants. We further propose Vehicle-to-Vehicle Vision-Language-Action (V2V-VLA), a cooperative VLA model that integrates cooperative driving into a single forward pass by jointly generating driving actions, future waypoints, language reasoning, and communication policies. Experiments on CMU-Drive establish the first benchmark and baseline for cooperative VLA driving and provide a foundation for future research on multi-agent, closed-loop, end-to-end cooperative autonomous driving. Our code, benchmark, and model checkpoint will be publicly released to facilitate open-source research.

As language models are increasingly used for tasks that require verifiable reasoning, reliably distinguishing sound reasoning from flawed reasoning has become an important practical problem. Recent trajectory-based methods seek this signal in layerwise residual-stream displacements, which capture how representations change while attenuating some stable, token-specific information. However, displacement omits the state from which an update originates, whereas restoring the full state risks reintroducing shortcut-prone information. We identify this trade-off and propose a three-stream detector that combines motion with two restricted views of location. A coarse region reader based on vector quantization and a fine direction reader over normalized multi-layer states. This design restores enough state context to interpret the motion without returning to full-state probing. On reasoning benchmarks unseen during training, our method improves selection accuracy by up to 12% over the displacement-only state of the art and 21% over single-layer probing baselines. Although trained only on reasoning benchmarks, it also reads factual completion and fact verification, ahead of every detector we compare against, which places the signal on correctness rather than on a kind of reasoning. Ablations further show that motion, region, and direction provide complementary signals. These results suggest that reasoning validity is better read from state-conditioned motion than from either static states or decontextualized trajectories alone.

Large language models can solve harder reasoning problems with more inference-time compute. The term "test-time scaling," however, covers several inference algorithms: extending deliberation along one trajectory, sampling completed candidates and aggregating them by voting or verification, and searching over partial states. These algorithms differ in statistical structure, compute requirements, and failure modes. Treating them as interchangeable under a scalar "budget," or reporting accuracy without specifying the inference protocol, makes results difficult to compare across studies. We study test-time scaling along three axes. First, we formalize it as budgeted inference over the implicit prefix tree of an autoregressive model and distinguish single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling. Second, we treat the full inference system as the evaluated object and separate end-to-end performance from candidate-bank diagnostics. We introduce an evaluation profile whose coordinates and simple functionals recover or bound common repeated-sampling metrics, and require compute accounting and uncertainty estimates that match the protocol. Third, we distinguish exact replay from distributional reproducibility and state the requirements for each. We also organize open-weight reasoning models by model-side and interface mechanisms. Our empirical study covers broad knowledge, symbolic reasoning, and competition mathematics, and we publicly release 1,403,520 sampled model attempts. The project website is available at https://mohsenhariri.github.io/scorio/tts. The released datasets are Trace (https://huggingface.co/datasets/harimo/scorio-trace), Lite (https://huggingface.co/datasets/harimo/scorio-lite), Math (https://huggingface.co/buckets/harimo/scorio-math), and SuperGPQA (https://huggingface.co/buckets/harimo/scorio-gpqa).

Parameter-efficient fine-tuning (PEFT) enables efficient adaptation of large language models, but existing MoE-based PEFT methods typically improve capacity by storing multiple full LoRA experts, causing adapter storage to grow linearly with the number of experts and restricting adaptation to a fixed expert pool. We ask whether MoE-based PEFT can produce instance-specific adaptations without explicitly storing a separate LoRA module for each expert. To address this gap, we propose MoEGen, an adaptation framework that shifts MoE-based PEFT from expert selection to expert-conditioned parameter generation. Instead of storing each expert as a full LoRA adapter, MoEGen represents each expert as a small learnable vector, termed an expert code. It routes each input over these vectors and uses their weighted combination to condition a lightweight hypernetwork that generates input-specific low-rank updates. This design decouples expert capacity from adapter storage while enabling instance-conditioned adaptation. Experiments on eight commonsense reasoning benchmarks show consistent improvements over strong static and MoE-based PEFT baselines across three backbones. MoEGen also performs strongly in joint medical and legal-domain adaptation.

Benchmark gains are often treated as evidence of greater LLM capability. Yet the same gain can reflect different changes in model behavior. A model may reach new answers, or produce answers that were already within reach. Aggregate scores do not distinguish these changes question by question. We establish a question-level audit under fixed budgets, temperatures, and answer formats. A question is realized when the default deployment procedure produces the correct answer. A question is reachable when a specified probe finds that answer within a fixed budget. We first test whether inference-time layer routing can expand reachability. Under a matched budget, random routes match or exceed structured search in all 43 model and task settings. Answer-blind procedures retain almost none of this gain, which instead requires access to the correct answer. We then ask why reachable answers sometimes fail to appear. Across six cases spanning 0.5B to 31B, silencing one identified MLP block repairs 68 to 92 percent of a predefined failure set. We next test whether training closes the gap by expanding reachability. In five of six matched evaluations, deployed performance rises while the reachable ceiling remains flat or falls. For DAPO, the deployed score rises by 14.7 points while the reachable ceiling falls by 13.3 points. Across the settings we audit, realization and reachability therefore do not always change together. Claims of capability expansion should report both realized performance and reachability under matched evaluation conditions. Code is available at https://github.com/LiZaiyuan0619/reachability-not-realization

Reinforcement Learning with Verifiable Rewards (RLVR) improves LLM reasoning but typically relies on ground-truth (GT) answers, limiting scalability. Voting-based label-free RLVR replace gold supervision with answer-level consensus from model samples. However, collapse arises when the same answer-level signal is used both to estimate rewards and to drive token-level policy optimization, encouraging the model to directly reinforce answer tokens rather than improve reasoning. We propose OM-GRPO, a label-free RLVR framework that decouples reward estimation from policy optimization. OM-GRPO masks gradients on the answer span while retaining answer-level rewards through a soft consensus signal, shifting optimization pressure away from answer tokens. We further introduce Contrast-Augmented Reward, which refines reward estimation via low-cost pairwise comparisons over existing trajectories without additional rollouts. Across diverse reasoning benchmarks and three LLM backbones, OM-GRPO consistently outperforms existing label-free RLVR methods and matches supervised GT-reward training with stable optimization. This stability is particularly beneficial in the Test-Time Training setting, where OM-GRPO surpasses majority voting by 4.24 points.

Scientific reasoning benchmarks typically evaluate large language models (LLMs) using final-answer accuracy. However, a correct answer does not necessarily demonstrate the reasoning capability targeted by the problem. We identify Solution Hacking, a failure mode in which an LLM reaches the correct answer through invalid shortcuts, such as numerical search, enumeration, guessing, or answer-first verification, without providing a valid task-targeted derivation. We systematically analyze this phenomenon across difficulty levels, scientific domains, and frontier models. Solution hacking increases sharply with benchmark difficulty, from 2.2% on common problems to 28.3% on Olympiad-level problems and 37.4% on HLE. Moreover, 8.2%-44.1% of answers credited as correct across frontier models are identified as hacked solutions. We further develop expert-inspired anti-hacking strategies, including an automatic judge and a test-time instruction. The results show that suppressing shortcut behavior substantially reduces reported accuracy while having a smaller effect on correct and non-hacked accuracy. These findings reveal that answer-only evaluation can overestimate the scientific reasoning capabilities of frontier LLMs.

Teaching a language model a skill it has not mastered is obstructed by three recurring difficulties: training data is scarce, ground-truth reasoning traces are usually unavailable, and models often exhibit an apparent ceiling beyond which additional data yields no further improvement. We study these difficulties in a controlled setting, fine-tuning Qwen2.5-Math-7B on competition mathematics (AIME), a task on which it initially solves only 5.6% of problems (pass@1). To address data scarcity, we introduce Question-begets-Question (QbQ), a scalable procedure in which a teacher transforms existing problems into diverse variants that probe the same underlying skills; to model the absence of oracle reasoning, we train exclusively via reinforcement learning on problem statements and final answers, never on teacher reasoning traces. Static training on such data, however, plateaus well short of the task: real-plus-synthetic augmentation and non-curriculum QbQ generated synthetic data training cap pass@1 at 12.5% and 14.5% respectively, despite large increases in data. Our central finding is that this ceiling is not intrinsic to the model. We propose a self-evolving curriculum that, each round, evaluates the current checkpoint, seeds QbQ from the problems it can mostly get right, and trains on the resulting variants; under an identical data budget, this breaks the ceiling and lifts pass@1 to 16.5% with no sign of saturation after 20 rounds. Counterintuitively, we find that models improve when trained on variants of problems they can mostly get right, and that models trained this way go on to solve harder problems never seen during training.

Graph reasoning provides a promising testbed for evaluating the reasoning ability of large language models (LLMs), as graph instances can be programmatically generated, structurally controlled, and naturally scaled to long-input settings. However, existing graph reasoning benchmarks have limited coverage of data complexity, rely heavily on manual construction, and lack unified evaluation across text-based and code-based reasoning modes. To address these limitations, we propose {\dataset}, a five-stage \textit{semi-automatic} framework for constructing complex graph reasoning benchmarks. It expands benchmark coverage along five dimensions: \textit{Graph Size}, \textit{Task Complexity}, \textit{Task Description}, \textit{Graph Loading}, and \textit{Task Source}. The framework uses an LLM-based data generator to automatically produce task descriptions, graph data, reference solutions, graph-loading scripts, question forms, and evaluation scripts, while retaining human validation at key quality-control stages. Based on it, we construct a benchmark with $202$ tasks and evaluate LLMs under text-based, code-based, and augmented reasoning settings. Experiments show that the complexity dimensions reveal model limitations that are less visible in existing benchmarks; existing fine-tuned models struggle to generalize to GraphGym, whereas retrieval-augmented methods show scenario-dependent adaptability, improving textual reasoning but not consistently improving coding reasoning. These findings suggest that ours serves as a challenging and diagnostic benchmark for graph reasoning and provides empirical guidance for future enhancement methods. Code and dataset will be published soon.

Games and simulators make valuable benchmarks by turning decisions into measurable outcomes, but many current suites under-test rules-rich tactical reasoning: the ability to choose well when geometry, timing, resources, objectives, and rule interactions all matter at once. We introduce DungeonBench, a benchmark for tactical reasoning in Dungeons & Dragons combat, built to cover the vast majority of combat-relevant 2014 System Reference Document content whose effects can be resolved by the simulator while retaining mechanics that simplified combat simulators often abstract away. At each step, DungeonBench exposes a complete tactical observation, a pending decision, and an indexed list of executable options spanning movement, attacks, spells, reactions, objectives, preparation, and scarce resources. The task is to value legal choices whose consequences depend on action economy, creature traits, battlefield geometry, timing windows, and future encounters. DungeonBench has two tracks: Encounter, which evaluates local tactical play in single fights, and Day, which links encounters through persistent hit points, spell slots, consumables, preparation, and short-rest timing, forcing policies to trade off immediate tactical advantage against future survivability. The same engine-generated decision stream supports heuristic controllers, language-model policies, learned option rankers, and masked-action reinforcement-learning agents. We evaluate frontier language-model policies on this shared decision stream. Results show that full tactical observations do not saturate the benchmark: frontier policies often win direct encounters, but linked encounter days expose failures in resource budgeting, rest timing, and rule-aware tactical discipline.

Scaling test-time computation can improve language-model reasoning, but uniform budgets waste computation on easy inputs, while verifier-guided refinement relies on external feedback. We introduce Self-Verifying Refinement (SVR), an oracle-free multi-turn reinforcement learning framework that learns to use self-verification as a compute-control policy. At each turn, the model produces a solution together with a discrete correctness verdict and a confidence score; it retains the current answer only when the verdict is Correct and confidence exceeds a threshold, and otherwise continues refinement using its own self-verification. Ground-truth correctness is used only to construct training rewards and is never exposed to the policy through refinement prompts or required at inference. SVR is trained with GRPO on fixed-horizon trajectories using rewards that promote solution correctness, calibration-aware self-verification, and stop-ready correct states; adaptive stopping is activated only at inference. On seven mathematical reasoning benchmarks with Qwen3.5-2B, SVR achieves a macro-average accuracy of 0.563 with only 2.99 inference turns on average. In the evaluated complete-system comparison, it exceeds standard GRPO, strong multi-turn baselines, and a fixed-budget oracle-guided score-feedback reference while requiring substantially fewer turns than fixed ten-turn inference. These results demonstrate that learned self-verification can serve as an effective internal control signal for answer retention and adaptive test-time compute allocation.

AI agents increasingly rely on large skill libraries, but selecting, combining, and maintaining skills remains difficult. We propose SKIMIX, a multi-agent framework in which agents with different skill portfolios collaborate through iterative refinement. SKIMIX combines embedding-based skill retrieval, submodular anti-dilution routing, and adaptive skill evolution. Across six reasoning benchmarks, multi-agent collaboration substantially improves open-ended mathematical reasoning but offers limited or negative gains on multiple-choice tasks. Agent-count scaling is non-monotonic, and most improvements arise during the first refinement round. These results show that task characteristics determine whether skill-level ensembles help and provide practical guidance for scalable agent design.

Enzyme function prediction is a hierarchical, knowledge-intensive form of protein function classification. Existing benchmarks expose an anomaly: general LLMs often get the coarse first level right, yet once asked for a complete EC number their accuracy at levels two through four drops to almost zero, while specialized models and tools stay usable. We propose EC-Reason-Bench, a training-free, diagnostic evaluation protocol built to answer two questions: why general LLMs score close to nothing on EC number prediction, and how much of that loss can be recovered without updating a single weight. We break enzyme classification ability into four orthogonal levers that can each be measured on their own: output structure, external knowledge, reasoning structure, and reasoning robustness. We test each lever with an inference-time method against a shared zero-shot baseline reproducing previously reported near-zero performance. Experiments with several strong reasoning LLMs yield four main findings. First, external knowledge is decisive and must precede reasoning: uniformly low closed-book performance rises sharply with open-book access, narrowing model gaps. Second, in closed-book settings, whether cascading and chain-of-thought help or hurt depends on a model's tendency to abstain. Third, once evidence is available the aggregate score of the best LLM setting is indistinguishable from simply voting the EC numbers of the nearest retrieved neighbors; that tie is an artifact of averaging, and it hides a large gain on adversarial evidence set against an equally large loss on multi-functional enzymes. Reasoning over evidence therefore acts as an arbiter of conflicting neighbors rather than as a source of knowledge, and no single-number leaderboard can see it. Fourth, accuracy obeys a law of homology availability.

Multilingual reasoning evaluation overwhelmingly relies on translating English benchmarks, a practice that introduces linguistic artifacts and fails to test culturally-grounded reasoning. We introduce ADAGE (Analogical Difficulty-by-design Assessment for Grounded Evaluation), a language-agnostic pipeline that combines native-speaker curation with LLM-assisted generation to construct challenging, translation-free benchmarks for abstract analogical reasoning. We validate ADAGE by constructing benchmarks for Arabic, Amharic, and Japanese. Evaluating 14 open-weight models, we find a consistent cultural reasoning gap: models that perform well on English proverb reasoning struggle substantially on all three native benchmarks, with accuracy dropping by 12--52 percentage points relative to English. We release the pipeline, all three benchmarks, and the full evaluation suite.

Chain-of-thought (CoT) prompting enables large language models (LLMs) to tackle multi-step reasoning tasks, yet the generated intermediate steps are not guaranteed to be logically sound. We present Reason Popper-ly, a neurosymbolic framework that uses inductive logic programming (ILP) to learn relation composition rules from reasoning traces and deploys them as an online verifier for step-level correction. Given an LLM-generated trace, the method checks each inferred step against the learned rule table, diagnoses the violation type, rewrites incorrect steps with symbolically derived repairs, and regenerates the remaining suffix so that the model can produce its final answer conditioned on a verified trace. We evaluate on CLUTRR, a multi-hop kinship reasoning benchmark, using five language models over reasoning chains of 2 to 10 hops. Across all models, Reason Popper-ly consistently improves terminal accuracy over standard CoT, with gains of up to 48 percentage points for small models and 15 points for frontier models on the longest chains. Compared with a fully exogenous symbolic pipeline, our method performs better on harder instances by preserving the model's successful grounding while correcting only verifiable reasoning failures. In addition, step-level ILP verification yields a fine-grained error taxonomy that provides diagnostic insight beyond final-answer accuracy.

Many organizations aim to adapt language models for internal use, both to improve performance on domain-specific tasks and to address privacy concerns around sensitive data. However, such adaptation remains non-trivial: it often requires operationally challenging fine-tuning of open-source models or ad hoc prompt optimization. We study a minimal alternative based on a simple API-level control: allowing users to bias the model's logits with a user-defined vector. We develop a black-box method for learning a single context-independent logit-bias vector, added at every decoding step, without modifying model weights or requiring gradients. Starting from a KL-regularized reinforcement learning (RL) objective, we characterize when such a fixed logit-bias vector can approximate the optimal prefix-dependent correction and derive a closed-form inverse-propensity estimator from rollouts, rewards, and token probabilities. Empirically, this simple decoding-time intervention improves over base models on mathematical and reasoning benchmarks while using far fewer trainable parameters than conventional fine-tuning. Our results suggest that learned logit bias is a lightweight mechanism for adapting language models under minimal access requirements.

Large reasoning models (LRMs) generate long reasoning traces before producing final answers. While these traces may contain useful signals for hallucination detection, harnessing them is non-trivial because long trajectories often include noisy steps that obscure the cues relevant to truthfulness assessment. In this paper, we identify two prevalent forms of reasoning noises, i.e., irrelevant steps and repetitive steps, and show that both substantially degrade hallucination detection performance. Existing confidence-based scores and naive embedding-based filtering fail to reliably separate noisy from informative steps. To address this challenge, we propose REDE, a novel learning framework for denoising reasoning traces for hallucination detection. Specifically, REDE leverages final-answer attention as an automatic supervision signal to shape the step-level representation space, yielding refined embeddings in which noisy steps can be reliably identified and filtered. REDE can be readily plugged into diverse hallucination detectors by operating on the filtered reasoning trajectory after removing noisy steps. Extensive experiments on multiple reasoning benchmarks show that REDE consistently improves detection performance over competitive baselines.

Humans distill experience into reusable abstractions, e.g., strategies and cautionary reminders, and apply them to gradually solve problems more effectively. We study whether Large Language Models (LLMs) can similarly benefit from such experiential abstractions. From LLMs' solution traces on the MATH training set, a stronger teacher or the LLMs themselves extract natural-language abstractions into a retrievable library. We explore two usage modes: (1) inference-time retrieval and (2) reinforcement learning (RL) with abstraction-augmented training prompts. Experiential abstractions improve LLM performance on mathematical and logical reasoning benchmarks. Self-extracted abstractions match teacher-extracted ones, and our abstraction usage framework can transfer to other datasets and models. These findings suggest LLMs can extract and apply experiential abstractions much as humans leverage distilled experience.

Existing adaptive inference methods for Large Language Models rely on observational heuristics, such as hidden-state similarity or activation magnitudes, to drop redundant modules. However, these correlation-based metrics often fail to capture subtle, non-linear structural computations vital for semantic accuracy. We introduce CausalGate, an intervention-guided framework for compute-efficient transformer inference. During a calibration phase, CausalGate isolates individual Attention and MLP sub-layers, zeros out their respective outputs, and measures the exact semantic damage via the Kullback-Leibler divergence of the final logit distribution. To eliminate runtime routing overhead, this structural importance hierarchy is distilled into a global set of static, lightweight scalar gates using an Exponential Moving Average smoothing objective paired with a differentiable pairwise ranking loss. Evaluated on TinyLlama-1.1B, Qwen2.5-3B, and Llama-3.1-8B across language modeling and commonsense reasoning benchmarks, CausalGate consistently outperforms prominent dynamic routing and layer-skipping baselines, translating theoretical compute savings into concrete hardware latency reductions with zero operational overhead.

Reinforcement learning with verifiable rewards (RLVR) improves reasoning in large language models. Yet, typical RLVR approaches fail on difficult problems: when a model cannot generate any correct solutions, it receives \textit{zero} learning signal. Providing privileged guidance during training, such as solution prefixes, can help overcome this learning cliff by steering the model towards {correct solutions with non-zero reward}. {We call these rollouts \textit{off-context}: they are generated from a training prompt that contains privileged guidance, while the target objective is defined by the original prompt without that guidance.} {We introduce} Off-Context GRPO (OC-GRPO), a minimally modified variant of GRPO that uses guided rollouts but applies an importance-corrected objective to steer the update back toward the original unguided objective, avoiding the mismatch that destabilizes uncorrected guided training. Empirically, our algorithm achieves a 3.9% absolute improvement (13.8% relative gain) over vanilla GRPO on average across standard mathematical reasoning benchmarks with negligible additional cost.

Large language models (LLMs) have demonstrated remarkable capabilities in language understanding, reasoning, and world knowledge. As embodied agents become increasingly capable, there is a growing demand for compact models that can serve as an on-device brain, preserving the broad general intelligence of LLMs while enabling effective high-level interaction with embodied environments. Existing approaches, however, often prioritize either general-purpose intelligence or specialized embodied capabilities, making it challenging to satisfy both requirements within a single model. We present \textbf{Athena-Brain-8B}, an 8B LLM designed to serve as an on-device brain for embodied intelligence for embodied intelligence. Through a multi-stage post-training pipeline consisting of General Supervised Fine-Tuning, General Reinforcement Learning, Embodied Expert training, and Model Merge, Athena-Brain-8B maintains strong general capabilities while acquiring strong high-level embodied interaction capabilities and generating concise responses for efficient embodied interaction. Experimental results demonstrate the effectiveness of Athena across both general and embodied evaluations. Compared with the corresponding Qwen3-8B thinking model, Athena-Brain-8B achieves comparable performance on general language and reasoning benchmarks while generating substantially shorter responses. On in-domain embodied benchmarks, Athena-Brain-8B consistently outperforms models of similar scale and surpasses several substantially larger frontier models evaluated zero-shot, demonstrating that compact language models can effectively integrate strong general intelligence with embodied capabilities.

With the rapid advancement of large language models (LLMs), modern systems not only possess strong foundational capabilities and extensive knowledge, but can also solve complex problems via long, multi-step reasoning. However, as reasoning traces become longer, LLMs may produce a substantial amount of hallucinated content during the reasoning process, which is often difficult to detect. In this work, we conduct a fine-grained analysis of hallucinations arising in LLM reasoning and find that the reasoning traces are particularly prone to Context-Sensitive Factual Hallucinations: cases where the model actually has the relevant knowledge, yet makes factual errors due to contextual interference during reasoning. To address this issue, we propose Step-level Self-Consistency Group Relative Policy Optimization (SSC-GRPO), which assigns step-level rewards to reasoning traces by computing self-consistency scores of individual steps across multiple rollouts. Compared with prior methods, SSC-GRPO achieves state-of-the-art performance on both mathematical reasoning benchmarks and hallucination leaderboards. Our results offer a new perspective for detecting and mitigating hallucinations in the reasoning process of large language models.

Introducing Relay-Bench, an unsaturated, holistic, text-only benchmark that measures LLMs' ability to complete an assortment of tasks from distinct domains in a single prompt. The leading model, GPT-5.5 (xHigh), scores 43.3%. The test set entirely consists of composite problems: groups of single-domain subproblems that are strung together into challenges that require reasoning across multiple domains in combination. Many of these problems then have layers of complexity added through prompt encoding and deliberate context bloat. Domains tested include visual reasoning, coding, math, information extraction (with a focus on web search), problem-solving, general knowledge, and data analysis. No restrictions are imposed outside of the model harness, and models are explicitly encouraged to leverage code-execution, web searches, and all available tools. All problems are composed of two to thirteen subproblems and do not require multi-modal input or output.

We propose the AIMO Interpretability Challenge, a competition on distinguishing robust from spurious reasoning in frontier mathematical language models based on the models' internal mechanisms. The challenge is motivated by a central limitation of standard reasoning benchmarks: strong final-answer accuracy does not reveal whether a model relies on stable reasoning mechanisms or exploits brittle reasoning shortcuts. Building on AI Mathematical Olympiad (AIMO) problems and submissions, together with resources from the Fields Model Initiative, the competition will provide (1) newly-published olympiad-level math reasoning problems and their symbolic representations, allowing generation of novel functional variants, (2) access to frontier reasoning models, and (3) assessments of models' adversarial robustness on these problems. Participants will use these resources, along with our computing infrastructure support, to develop methods for identifying which models solve problems robustly. Our competition will also create a new, open robustness benchmark and baseline systems, aiming to provide a lasting foundation for standard benchmarking in mathematical reasoning and interpretability. Scientifically, the competition connects interpretability and generalization research around a central question in AI research: can we determine if, and to what extent, the decision-making of frontier AI models is generalizable and thus, reliable?

Large language models (LLMs) have achieved remarkable performance on high-school and olympiad-style mathematics, yet their capabilities on advanced mathematics remain poorly understood. Existing benchmarks, however, fall short in both scope and evaluation granularity: they provide limited disciplinary coverage and often rely on final-answer correctness or coarse judgments, leaving the validity of the reasoning process inadequately assessed. To bridge this gap, we introduce AdvancedMathBench, a benchmark suite designed to evaluate advanced mathematical reasoning capabilities. Its core proof-generation benchmark, ProverBench, contains 296 problems spanning undergraduate and doctoral qualifying-exam levels. To provide reliable evaluation of the proofs, we develop a dedicated automatic verification pipeline trained on large-scale expert annotations to produce both correctness verdicts and fine-grained assessments of proof errors, which exhibits strong agreement with human experts on held-out proof trajectories. We further introduce VerifierBench, consisting of 888 model-generated proof trajectories paired with expert ground truth, to evaluate whether models can correctly judge proof validity and provide sound verification rationales. Experiments show that AdvancedMathBench remains challenging for frontier models. On proof generation, the best-performing model, GPT-5.5-xhigh, achieves only 75.8 and 66.1 on the UGD and QE splits, respectively, indicating substantial room for improvement on advanced mathematical proof construction. On proof verification, the best model attains a Balanced F1 of only 65.1, and models generally exhibit low true negative rates, suggesting that critical error detection remains a major bottleneck.

Large Language Models (LLMs) have achieved remarkable success in complex reasoning tasks through Chain-of-Thought (CoT) prompting. However, these models often exhibit "computational overthinking," generating redundant reasoning steps that increase latency and cost without improving accuracy. Recent studies suggest that CoT trajectories can be significantly pruned, yet existing methods often rely on forcing a static thinking budget, heuristic filtering, sub-optimal early exit via classification, or expensive re-training. In this paper, we introduce OS-Pruner, a lightweight plug-in framework that formulates chain-of-thought pruning as an optimal stopping problem. Given a reasoning prefix, OS-Pruner learns whether further reasoning is worth its token cost by optimizing an explicit utility that trades off final-answer accuracy against generated length. Our novel formulation enables the model to dynamically assess the sufficient point of termination for a reasoning chain. OS-Pruner is designed to be lightweight during both training and inference, and to provide users with fine-grained control over the reasoning-effort vs. accuracy trade-off. On diverse reasoning benchmarks and base models, OS-Pruner achieves 20-60% reduction in generation length with minimal accuracy sacrifice.

Reinforcement learning (RL) has significantly enhanced the reasoning capabilities of large language models (LLMs), yet the training process remains notoriously fragile. In this work, we investigate a critical source of this instability: over-optimization, where models exploit training heuristics at the expense of generalizable reasoning. While reverse KL regularization is the standard defense against such degradation, our analysis reveals that it is often insufficient in this regime, as it fails to ensure comprehensive coverage of the reference distribution. To address this, we propose ARMOR (Anchor Rollout and Mixed Optimization for RL), a framework that shifts the paradigm from passive penalty to active sample stabilization. ARMOR comprises two key components: (1) Anchor Rollout, which leverages off-policy data from the reference policy to preserve established solution patterns; and (2) Mixed Optimization, which reformulates the policy objective to enable controlled exploration without relying on auxiliary losses. Extensive experiments on reasoning benchmarks validate that ARMOR effectively mitigates validation collapse, enabling sustained performance improvements over extended training horizons.

Agentic retrieval-augmented generation (RAG) extends static RAG by allowing language models to iteratively reason, generate search queries, retrieve evidence, and predict answers. However, it remains challenging for models to decide when to retrieve, whether to use lexical matching or semantic similarity, and how to control context granularity to prevent irrelevant tokens from interfering with agent reasoning. In this paper, we introduce GRASP, a reinforcement learning (RL) framework for training agents to adaptively coordinate complementary retrieval tools during multi-step reasoning. GRASP provides the agent with semantic search, keyword search, and paragraph-reading actions, enabling it to retrieve sentence-level evidence and expand further context only when needed. We train the policy with a reward that jointly accounts for answer accuracy, grounded reading, complementary search, and turn efficiency. Experiments on multi-hop reasoning benchmarks show that GRASP improves both retrieval recall and downstream question answering performance compared with single-step retrieval, prompting-based agentic RAG, and RL-based retrieval baselines. Qualitative and ablation analyses show that the learned policy develops interpretable skimming and scanning behavior: it uses semantic search for broad exploration, paragraph reading for local verification, and keyword search for entity-specific evidence. These results suggest that learning to coordinate retrieval signals and context granularity is critical for agent's correct reasoning.

Large language models (LLMs) excel at generating long chains of thought, but long reasoning traces are often verbose and memory-inefficient. In this work, we introduce Structured Thoughts, a framework that organizes reasoning into alternating <try> and <outcome> blocks: <try> captures exploratory scratch work, while <outcome> contains the distilled conclusion of that step. We construct a dataset of structured thoughts by segmenting reasoning traces into <try> blocks and prompting an LLM to summarize each step into its corresponding <outcome>. Fine-tuning pretrained foundation models on this reformatted data produces models that adopt the structured reasoning style, leading to performance gains of up to 8.08% on reasoning benchmarks compared to standard SFT. The explicit structure also enables context pruning: after each <try>/<outcome> pair, the <try> can be pruned, allowing the model to retain conclusions without keeping the full scratch work in the context. A proof-of-concept pruning implementation achieves an average of 85% memory / context savings with an 8.67% performance drop across mathematical tasks.

Capable language models hold what a character believes apart from what is true: told "Anna believes the cup is blue; in reality it is red," they answer blue about Anna and red about the world. Where in the computation does that separation live? We show it rests on two separable mechanisms at two positions. A generic value slot binds the attributed value. A router at the query position selects which frame, the character's belief or reality, a query reads out. Two routes fill the slot: an asserted belief, whose value the text supplies, binds in directly; a derived belief, whose value must be inferred from what the character could see, arrives by a visibility-gated lookback. A subspace trained on either route steers the other, and only the derived route depends on described visibility. The slot itself carries no belief-reality tag: intervening on it moves a reality readout as strongly as a belief one. The separation lives instead in a dissociated pair of routing subspaces, which flip a query between frames without injecting the donor's value. These results hold across three architectures, on stimuli de-confounded against theory-of-mind-benchmark shortcuts; the behavior itself emerges between 3B and 7B across five model families. This paper develops the single belief-reality axis in depth; a companion paper shows the same slot-and-router format is shared across the other non-actual contexts a sentence can open (counterfactual, fictional, temporal).

Large language models produce chain-of-thought (CoT) reasoning that appears logically sound yet may not genuinely depend on its stated premises. We introduce interventional grounding audits, a black-box, step-level test of premise dependency: we intervene on a single premise by substituting its target predicate with a fresh symbol, re-run the model, and check whether each reasoning step's normalized conclusion (canonical predicate form) changes. We evaluate on ProntoQA, a synthetic multi-hop deductive reasoning benchmark with gold proof trees, where step-level premise dependencies are known. Applied to 50 ProntoQA problems with GPT-4o, our method achieves F1 = 0.806 on detecting proof-tree dependencies (F1 = 0.885 on predicate-determining dependencies; Recall = 100%), significantly outperforming a self-consistency baseline (F1 = 0.343; 95% bootstrap CIs non-overlapping). We further identify that 66% of correctly-solved problems contain at least one aligned step insensitive to a direct proof-tree dependency under consistent substitution -- all involving entity-introduction premises, a documented blind spot of the consistent-substitution evaluator -- a "right answer, wrong reasoning" signal invisible to passive methods. All audit certificates, raw outputs, and reproduction scripts are available in a public GitHub repository, and we discuss scope limits beyond formal, parsable benchmarks.

Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades, indicating that models still struggle to identify and use the evidence most relevant to a question. A promising way to improve long-context utilization is test-time training (TTT), which treats the test context as a training example for instance-specific parameter adaptation. However, applying TTT to the entire long context is prohibitively expensive, while adapting on randomly sampled spans introduces severe noise. Because most spans in a long context are irrelevant to the specific question, training on them may even degrade the base model's performance. Our preliminary study shows that TTT is highly sensitive to training-span quality: on LongBench-v2, TTT on randomly sampled spans hurts performance, whereas TTT on oracle spans substantially improves it. Motivated by this, we propose a simple method, Self-Guided TTT (S-TTT): before adaptation, the model identifies the evidence spans it should learn from, and the standard language-modeling training objective is applied only to those selected spans. On two challenging long-context reasoning benchmarks, LongBench-v2 and LongBench-Pro, S-TTT improves accuracy for both Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, achieving up to a 15% relative improvement.

Scientific ideas rarely start from a blank page. They inherit mechanisms, repair known limitations, and recombine pieces of earlier work, much like biological genomes. Current benchmarks still say little about whether AI systems can follow this inheritance structure. We present IdeaGene-Bench (IG-Bench), a benchmark for scientific lineage reasoning and lineage-grounded idea generation. IG-Bench is organized around the IdeaGene framework: each paper or proposal is represented as a set of minimal, typed, evidence-grounded Idea Genome objects, and a GenomeDiff aligns these objects to record inheritance, mutation, loss, external import, and novel insertion under six operational evolutionary dynamics. The benchmark contains 1,961 golden lineage traces, 1,085 curated Idea Genome objects, and 920 pairwise GenomeDiff records across 10 scientific domains. It supports two evaluations. IG-Exam (42 task types, 1,029 instances) tests closed-form lineage reasoning across Idea Genome abstraction, inheritance tracing, evolutionary reasoning, and lineage verification. IG-Arena evaluates generation with a lineage-conditioned Population-Evolution Score(PES), asking whether a proposal can be inserted as a coherent descendant of a given lineage population: it should inherit the right Idea Genome objects, vary meaningfully from nearby work, and offer selection value for future research. Experiments on 14 LLM-based scientists expose a compositional bottleneck. The strongest system reaches only 27.3% exact accuracy on lineage reasoning, and structured lineage context reshuffles system rankings rather than helping every participant uniformly.

Mathematical reasoning benchmarks typically provide all facts needed to solve each problem, while interactive benchmarks often mix reasoning with tools, retrieval, and long-horizon dialogue. We introduce MIRA-Math, a benchmark for a narrower diagnostic capability: solving mathematical problems whose full latent state has a unique answer, but whose solver-facing view is missing exactly one necessary atomic fact. The solver must request the missing information in natural language under a strict budget and then integrate the returned fact into an exact final answer. A fixed constrained LLM responder sees only the dataset-provided atomic fact and must either offer the quoted fact when the request matches it, or decline otherwise. Thus, instance generation, typed hint specifications, validation, and final-answer verification are deterministic, while request metrics are measured under a fixed LLM-mediated responder channel. MIRA-Math contains 2{,}310 generated instances from 22 typed mathematical families spanning algebra, probability, linear systems, discrete structures, signal processing, Markov chains, circuits, interpolation, and numerical boundary-value problems. Experiments across frontier and small models show that request success and final-answer accuracy are separable: models may ask for the right fact yet fail the downstream computation, or fail before obtaining the canonical hint. We release generators, verifiers, prompts, run metadata, and dataset documentation to support reproducible evaluation of minimal information requesting in mathematical reasoning.