Large Language Model Reasoning

Fixed reasoning benchmarks evaluate canonical prompts, but semantically valid changes in presentation can still change model behavior. Studies of prompt variation can reveal such failures, but without audit they can mix genuine model errors with invalid perturbations, extraction artifacts, and unmatched search procedures. We propose an audit-constrained protocol for targeted reasoning evaluation. Prompt variants are generated from a finite component grammar, rendered deterministically, evaluated under a fixed query budget, and counted as model errors only after semantic and extraction audit. Within this protocol we instantiate Component-Adaptive Prompt Sampling (CAPS), a score-based sampler over prompt components, and compare it with equal-budget uniform component sampling under the same task bank, renderer, model interface, decoding settings, and audit procedure. Across three audited slices, the protocol identifies confirmed model-error prompt keys while excluding formatting and extraction artifacts, but matched comparisons do not show that CAPS improves audited yield or unique prompt-key discovery over uniform sampling. The contribution is methodological: targeted prompt variation can be studied under a reconstructable, reviewable, budget-matched protocol, and proxy-guided policies should be judged by audited yield rather than raw mismatch counts or selected examples alone.

Group Relative Policy Optimization (GRPO) has emerged as a promising approach for improving the reasoning capabilities of large language models. However, it struggles to effectively balance the tradeoff between exploration and exploitation during training, often resulting in suboptimal performance. Motivated by the theoretical insight that changes in entropy are governed by the covariance between token probabilities and their corresponding advantages, we propose a hyperparameter-free, covariance-weighted optimization method that dynamically down-weights extreme token-level updates via a Gaussian kernel. This approach automatically reduces the instability caused by exploration-exploitation trade-off while preserving informative learning signals. Extensive empirical evaluations show that our approach improves downstream performance across reasoning benchmarks compared with GRPO, and effectively stablizes entropy as training progresses.

Reinforcement learning with verifiable rewards helps reasoning, but GRPO-style methods stall on hard prompts where all sampled rollouts fail. SORT adds a repair update for those failures without changing rollout generation: it derives a plan from the reference solution, compares token probabilities with and without that plan, and gives higher weight to tokens that become more predictable under plan conditioning. This turns all-wrong prompts into selective, structure-aware learning signals instead of uniform imitation. Across three backbones and eight reasoning benchmarks, SORT improves over GRPO and guidance baselines, with largest gains on weaker models.

Reasoning models post-hoc rationalize answers they have already committed to internally, producing chains of reasoning theater: deliberative-looking steps that contribute nothing to correctness. This wastes inference tokens, pollutes interpretability, and obscures what the model actually computed. We introduce ProFIL (Probe-Filtered Reinforcement Learning) to reduce theater, increase chain-of-thought faithfulness, and shrink chain length in a single, drop-in extension to Group Relative Policy Optimization (GRPO). A multi-head attention probe is trained once on the frozen base model to detect post-commitment steps from internal activations alone; during GRPO, rollouts whose probe score exceeds a threshold have their advantage zeroed. Our central finding is that a probe trained on a frozen base, with verifier-derived labels and no human annotation, provides a stable signal that suppresses theater while resisting the RL-obfuscation failure mode predicted by prior work. Across four reasoning domains (GSM8K, LiveCodeBench, ToolUse, MMLU-Redux) and two model architectures (Llama-8B, Qwen-7B), ProFIL reduces post-commitment theater by 11--100%, raises faithful-fraction (e.g., +24pp on LiveCodeBench under an independent Claude 3.7 Sonnet judge), and shortens chains by 4--19%, all while preserving or improving task accuracy. ProFIL also beats a matched length-penalty GRPO baseline, isolating the gain as semantic commitment-detection rather than chain compression. Probe weights, training configurations, and rollouts are released across all four domains.

Reinforcement learning with verifiers (RLVR) has become a central paradigm for improving LLM reasoning, yet popular group-based optimization algorithms like GRPO often suffer from exploration collapse, where the models prematurely converge on a narrow set of high-scoring patterns, lacking the ability to explore new solutions. Recent efforts attempt to alleviate this by adding entropy regularization or diversity bonus. However, these approaches do not change the \textit{winner-takes-all} nature, where rollouts still compete for individual advantage rather than cooperating for maximizing global diversity. In this work, we propose Group Cooperative Policy Optimization (GCPO), which shifts the training paradigm from rollout competition to team cooperation. Specifically, GCPO replaces independent rollout scoring with team-level credit assignment: a rollout is rewarded by how much it contributes to the team's valid solution coverage, rather than its individual accuracy. This coverage is described as a determinant volume over reward-weighted semantic embeddings, where only correct and non-redundant rollouts contribute to this volume. During advantage estimation, GCPO redistributes the collective team reward to each single rollout according to its average marginal contribution to the team. This cooperative training paradigm routes optimization toward non-redundant correct reasoning paths. Experiments across multiple reasoning benchmarks demonstrate that GCPO significantly improves both reasoning accuracy and solution diversity over existing approaches. Code will be released at https://github.com/bradybuddiemarch/gcpo.

On-policy self-distillation has become a strong recipe for LLM reasoning, where a privileged teacher supervises the student's own rollouts while conditioning on the reference solution. A design choice shared by nearly all such methods, however, has gone unquestioned: the teacher always sees the full reference reasoning. We argue that this default itself is part of the problem and identify a teacher-side exposure mismatch: when the teacher conditions on reasoning far beyond the student's current competence, the resulting token targets become too strong to absorb. A controlled fixed-exposure sweep makes this concrete on two fronts: 1) full exposure is not reliably the best choice, and 2) student-teacher mismatch grows monotonically as the teacher sees more privileged reasoning. This motivates treating teacher exposure not as a fixed hyperparameter but as a learnable training-time control variable. We therefore propose Adaptive Teacher Exposure for Self-Distillation (ATESD). ATESD models the reveal ratio with a lightweight Beta-policy controller conditioned on compact training-state statistics, and uses one sampled exposure for a short hold window of student updates. To make this exposure controller learnable, we optimize it with a discounted learning-progress reward that scores each held decision by its effect on the student's future improvement rather than its immediate loss change, addressing the delayed credit assignment induced by on-policy distillation. Experiments on AIME 24, AIME 25, and HMMT 25 across Qwen3-{1.7B, 4B, 8B} show that ATESD consistently outperforms competitive self-distillation and RL baselines, improving over OPSD by +0.95, +2.05, and +2.33 Average@12 points respectively, and establishing adaptive teacher exposure as an effective new axis for reasoning self-distillation.

Chain-of-thought and more broadly test-time compute are known to augment the expressive capabilities of language models and have led to major innovations in reasoning. Motivated by this success, this paper explores latent chain-of-thought as well as the impact of depth and looping for time-series and tabular data. We propose a recurrent scheme in which a structured-data transformer, after an initial forward pass, compresses its query-position hidden states into feedback tokens that are appended to the input and processed again, allowing multiple rounds of latent computation before prediction. We compare CoT models against a same-depth no-CoT baseline, a deeper baseline matched to the CoT model in effective depth, and a looped transformer with weight-tied recurrence but no additional chain-of-thought tokens. Across 36 datasets in time-series forecasting and tabular prediction, latent chain-of-thought improves over the baseline on 7/9 time-series datasets (+12.63% average gain) and 23/27 tabular datasets (+3.25% average gain), with CoT models performing best on average in both settings. We also show that the benefit of CoT extends to pretrained foundation models: applying latent CoT to nanoTabPFN, a small open-source tabular foundation model, improves its performance above the much larger TabPFN-v2 on TabArena. Together, these results demonstrate that chain-of-thought is a useful axis for scaling test-time compute for structured data.

Reasoning benchmarks measure clinical performance on clean inputs. We evaluate the step before reasoning: retrieval over real EHR notes, where negation, temporality, and family-versus-patient attribution can flip a correct answer to a wrong one. EpiKG carries an assertion label and a temporality tag with every fact in a patient knowledge graph, then routes retrieval by question intent. ClinicalBench is a 400-question test over 43 MIMIC-IV patients across 9 assertion-sensitive categories. A 7-condition ablation tests each piece of EpiKG across six LLMs (Claude Opus 4.6, GPT-OSS 20B, MedGemma 27B, Gemma 4 31B, MedGemma 1.5 4B, Qwen 3.5 35B). Three physicians blindly adjudicated 100 paired items. The author-blind primary endpoint, leave-author-out paired exact McNemar on 50 unanimous-strict items rated by two external physicians, yields +22.0 percentage points (95 percent Newcombe CI [+5.1, +31.5], p=0.0192). The architectural novelty, intent-aware KG-RAG over a Contriever dense-RAG baseline (C2b to C4g_kw on the change-excluded n=362 endpoint), is +8.84 percentage points (paired McNemar p=1.79e-3); +12.43 percentage points under oracle intent. Sensitivities agree directionally: three-rater physician majority +24.0 percentage points (subject to single-author circularity); deterministic keyword reproducibility proxy +39.5 percentage points. Across the six models, the gain shrinks as the LLM-alone baseline rises (beta=-1.123, r=-0.921, p=0.009). With n=6 this looks more like regression to the mean than encoding substituting for model size. Physician adjudication identified 56 percent of auto-generated reference answers as defective, a methodological finding indicating that NLP-pipeline clinical-QA benchmarks require physician adjudication to be usable. ClinicalBench, the frozen evaluator, three-rater adjudication data, and the EpiKG output stack are publicly released.

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

Corruption studies, the standard tool for evaluating chain-of-thought (CoT) faithfulness, infer which steps are ``computationally important'' from accuracy loss when steps are corrupted. We show that when benchmark chains end with an explicit terminal answer line, as in GSM8K and MATH, these tests largely measure \emph{answer placement} rather than where intermediate computation is carried out. Using matched GSM8K examples, removing only the final answer statement while preserving all reasoning collapses suffix sensitivity by about $19\times$ for Qwen~2.5-3B ($N{=}300$, $p{=}0.022$). Conflicting-answer prompts, which contain correct reasoning but a wrong explicit final answer, drive accuracy to zero or near-zero at 7B across five open-weight model families; wrong-answer following is strong at 3B--7B and attenuates sharply at larger scales. Replications on MATH, within-stable comparisons at 7B, and suffix-free chains show the same pattern in different guises: corruption sensitivity tracks the location of explicit answer text, not a fixed computational depth in the reasoning. Generation-time probes indicate that final answers are rarely early-determined during generation (${<}5\%$ early commitment), yet consumption-time behavior systematically follows explicit answer text. The confound is therefore largely a readout effect when the chain is consumed. We propose a three-prerequisite protocol (question-only control, format characterization, and an all-position sweep) as a practical minimum for future corruption-based faithfulness studies.

Language models are deployed in settings that require compartmentalization: system prompts should not be disclosed, chain-of-thought reasoning is hidden from users, and sensitive data passes through shared contexts. We test whether models can keep prompted information out of their writing. We give each model a secret word with instructions not to reveal it, then ask it to write a story. A second model tries to identify the secret from the story in a binary discrimination test. The secret word never appears literally in any output, but all five frontier models we test leak it thematically -- through topic choice, imagery, and setting--6hy-at rates significantly different from chance, up to 79%. When told to actively hide the secret, models write \emph{away from} it, and this avoidance is itself detectable. The leakage is cross-model readable, scales sharply with model size within two model families, and disappears entirely for short-form writing like jokes. Giving the model a decoy concept to ``focus on instead'' partially redirects the leakage from the real secret to the decoy. Attending to a secret appears to open up an information channel that frontier LLMs cannot close, even when instructed to.

Cancer screening is a reasoning task. A radiologist observes findings, compares them to prior scans, integrates clinical context, and reaches a diagnostic conclusion confirmed by pathology. We present RadThinking, a Visual Question Answering (VQA) dataset that makes this reasoning explicit and trainable. RadThinking releases VQA pairs at three difficulty tiers. Foundation VQAs are atomic perception questions. Single-step reasoning VQAs apply one clinical rule. Compositional VQAs require multi-step chain-of-thought to reach a guideline category such as LI-RADS-5. For every compositional VQA, we release the chain of foundation VQAs that solves it. The chain follows the rules of the governing clinical reporting standard. The dataset spans 20,362 CT scans from 9,131 patients across 43 cancer groups, plus 2,077 verified healthy controls with >1-year follow-up. To our knowledge, RadThinking is the first cancer-screening VQA corpus that stratifies questions by reasoning depth and grounds compositions in clinical reporting standards. The foundation tier supplies atomic perception supervision. The compositional tier supplies chain-of-thought data and verifiable rewards for reinforcement-learning recipes such as DeepSeek-R1 and OpenAI o1. RadThinking enables systematic training and evaluation of whether AI systems can reason about cancer, not merely detect it.

Multi-agent systems (MAS) assume that collaborating inherently improves Large Language Model (LLM) reasoning. We challenge this by demonstrating that simulated social pressure triggers an algorithmic Bystander Effect,'' inducing severe cognitive loafing. By evaluating 22,500 deterministic trajectories across 3 dataset contexts (GAIA, SWE-bench, Multi-Challenge) with 3 state-of-the-art (SOTA) models, we semantically audit internal reasoning traces against external outputs. We formalize the \textit{Interaction Depth Limit} ($D_L$), the exact plurality threshold where an agent's logical sovereignty collapses into social compliance. Crucially, we uncover the \textit{Sovereignty Gap}: models frequently compute the correct derivation internally but suffer Alignment Hallucinations'' -- actively subjugating empirical evidence to sycophantically appease a simulated swarm. We prove that multi-agent social load is strictly non-commutative; the "brand" identity of the ``Lead Anchor'' auditor disproportionately dictates the swarm's integrity. These findings expose architectural vulnerabilities, proving that unstructured multi-agent topologies can degrade independent reasoning.

Deep learning has become the dominant approach for creating high capacity, scalable models across diverse data modalities. However, because these models rely on a large number of learned parameters, tightly couple feature extraction with task objectives, and often lack explicit reasoning mechanisms, it is difficult for humans to understand how they arrive at their predictions. Understanding what representations emerge and why they arise from the training data remains an open challenge. We introduce Deep Arguing, a novel neurosymbolic approach that integrates deep learning with argumentation construction and reasoning for interpretable classification with different data modalities. In our approach deep neural networks construct an argumentation structure wherein data points support their assigned label and attack different ones. Using differentiable argumentation semantics for reasoning, the model is trained end-to-end to jointly learn feature representation and argumentative interactions. This results in argumentation structures providing faithful case-based explanations for predictions. Structure constraints over the argumentation graph guide learning, improving both interpretability and predictive performance. Experiments with tabular and imaging datasets show that Deep Arguing achieves performance competitive with standard baselines whilst offering interpretable argumentative reasoning.

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

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

Graph reasoning agents operating from natural-language inputs must solve a coupled problem: they must reconstruct a structured graph instance from text, decide whether existing computational assets are sufficient, interact with tools under a strict execution protocol, and satisfy an external verifier that checks structured correctness rather than textual plausibility. Existing approaches usually improve either the instruction side or the tool side in isolation, which leaves unclear what should be updated after failure. We propose EGL-SCA, a verifier-centric dual-space framework that models a graph reasoning agent using two collaborative components: an instruction-side policy space for reasoning strategies, and a tool-side program space for executable algorithmic tools. Our central mechanism is structural credit assignment, which maps trajectory evidence to conditional updates, precisely routing failures to either prompt optimization or tool synthesis and repair. To provide sufficient learning signals for dual-space adaptation, we introduce a training distribution stratified by task family, coupled with a Pareto-style retention strategy to balance success, generality, and parsimony. Experiments on four graph reasoning benchmarks show that EGL-SCA achieves a state-of-the-art 92.0% average success rate. By effectively co-evolving instructions and tools, our framework significantly outperforms both pure-prompting and fixed-toolbox baselines.

Test-time scaling has become an effective paradigm for improving the reasoning ability of large language models by allocating additional computation during inference. Recent structured approaches have further advanced this paradigm by organizing inference across multiple trajectories, refinement rounds, and verification-based feedback. However, existing structured test-time scaling methods either weakly coordinate parallel reasoning trajectories or rely on noisy historical information without explicitly deciding what should be retained and reused, limiting their ability to balance exploration and exploitation. In this work, we propose TMAS, a framework for scaling test-time compute via multi-agent synergy. TMAS organizes inference as a collaborative process among specialized agents, enabling structured information flow across agents, trajectories, and refinement iterations. To support effective cross-trajectory collaboration, TMAS introduces hierarchical memories: the experience bank reuses low-level reliable intermediate conclusions and local feedback, while the guideline bank records previously explored high-level strategies to steer subsequent rollouts away from redundant reasoning patterns. Furthermore, we design a hybrid reward reinforcement learning scheme tailored to TMAS, which jointly preserves basic reasoning capability, enhances experience utilization, and encourages exploration beyond previously attempted solution strategies. Extensive experiments on challenging reasoning benchmarks show that TMAS achieves stronger iterative scaling than existing test-time scaling baselines, with hybrid reward training further improving scaling effectiveness and stability across iterations. Code and data are available at https://github.com/IQuestLab/tmas.

Tree-of-Thought (ToT) reasoning structures Large Language Model (LLM) inference as a tree-based search, demonstrating strong potential for solving complex mathematical and programming tasks. However, its efficiency is constrained by the reward dependency barrier -- a synchronization bottleneck caused by sequential reward-guided exploration that limits search parallelism and introduces substantial latency. Prior system optimizations, mainly designed for linear Chain-of-Thought (CoT) reasoning, cannot address these challenges, leaving the efficiency of ToT underexplored. To enhance ToT reasoning efficiency, we observe that the reasoning paths can be explored speculatively to break the reward synchronization barrier. Therefore, in this paper, we propose SPEX and introduce three key techniques: (i) intra-query speculative path selection to predict and expand high-potential branches of ToT, (ii) inter-query budget allocation to balance speculative resource allocation across queries dynamically, and (iii) adaptive early termination to prune deep and redundant branches for a skewed search tree. We implement SPEX on top of the SGLang framework and evaluate it across diverse ToT algorithms and LLMs. Extensive experiments show that SPEX achieves $1.2 \sim 3 \times$ speedup for different ToT reasoning algorithms. Moreover, SPEX synergizes with token-level speculative decoding, achieving cumulative speedups of up to $4.1\times$. Ablation studies further confirm the contributions of each technique. Overall, SPEX represents a significant step toward efficient and scalable ToT reasoning, unlocking the parallelism required for high-performance inference-time scaling for LLMs.

Process Reward Models (PRMs) are a powerful mechanism for steering large language model reasoning by providing fine-grained, step-level supervision. However, this effectiveness comes at a significant cost: PRMs require expert annotations for every reasoning step, making them costly and difficult to scale. Here, we propose a method for training unsupervised PRMs (uPRM) that requires no human supervision, neither at the level of step-by-step annotations nor through ground-truth verification of final answers. The key idea behind our approach is to define a scoring function, derived from LLM next-token probabilities, that jointly assesses candidate positions of first erroneous steps across a batch of reasoning trajectories. We demonstrate the effectiveness of uPRM across diverse scenarios: (i) uPRM achieves up to 15% absolute accuracy improvements over the LLM-as-a-Judge in identifying first erroneous steps on the ProcessBench dataset; (ii) as a verifier for test-time scaling, uPRM performs comparably to supervised PRMs and outperforms the majority voting baseline by up to 6.9%, and (iii) when used as a reward signal in reinforcement learning, uPRM enables more robust policy optimization throughout training compared to a supervised PRM trained using ground-truth labels. Overall, our results open a path toward scalable reward modeling for complex reasoning tasks.

Recent neural theorem provers use reinforcement learning with verifiable rewards (RLVR), where proof assistants provide binary correctness signals. While verifiable rewards are cheap and scalable without reward hacking issues, they suffer from sparse credit assignment: models receive no learning signal from difficult problems where partial progress goes unrewarded. This motivates learned reward models that can evaluate proof quality beyond binary verification. However, comparing reward models is challenging since it typically requires expensive RL training ablations. To address this, we introduce \textbf{FormalRewardBench}, the first benchmark for evaluating reward models in formal theorem proving with Lean 4. Our benchmark consists of 250 preference pairs where correct proofs are paired with incorrect variants generated through five expert curated error injection strategies: forced mistakes, minimal single-point variations, verbose incorrect proofs, natural language justification, and Python code injection. We evaluate frontier LLMs (e.g., Claude Opus 4.5), judge LLMs (e.g., CompassJudger-1-14B), general-purpose LLMs (e.g., Qwen2.5-72B-Instruct), and specialized theorem proving models (e.g., DeepSeek-Prover-V2-7B). Our results reveal that frontier LLMs achieve the highest performance (59.8%) while specialized theorem provers perform the worst (24.4%), suggesting that theorem proving ability does not transfer to proof evaluation. We provide further insights on various error injection mechanisms, highlighting the challenging nature of most injection mechanisms. We release \textbf{FormalRewardBench} publicly to encourage more research on developing reward models in formal mathematics.

Large language models (LLMs) exhibit strong reasoning capabilities when guided by high-quality demonstrations, yet such data is often distributed across organizations that cannot centralize it due to regulatory, proprietary, or institutional constraints. We study federated reasoning, where a server improves multi-step reasoning by coordinating with heterogeneous clients holding private demonstrations, without centralized training or raw data sharing. The key challenge is that client reliability is query-dependent, while the server cannot inspect client data to determine which contributions are trustworthy. To address this, we propose Uncertainty-Aware Federated Reasoning (FERA), a training-free framework based on iterative server-client co-refinement. Across communication rounds, clients generate reasoning traces with lightweight uncertainty estimates, and the server synthesizes them into improved reasoning that is redistributed as context for the next round, progressively improving both server outputs and client-side reasoning. Within each round, Uncertainty-Aware Self-Critique Aggregation (UA-SCA) resolves conflicts among heterogeneous client traces through query-dependent trust weighting and structured cross-client verification. Rather than simply discarding low-quality traces, UA-SCA revises flawed reasoning steps to recover useful information. We provide theoretical guarantees showing that the proposed iterative protocol converges and that uncertainty-aware weighting accelerates convergence. Experiments on multiple reasoning benchmarks show that FERA consistently outperforms both federated training and training-free baselines, achieving progressively higher accuracy across rounds while maintaining communication and computational efficiency.

Tool-using LLM agents must act on untrusted webpages, emails, files, and API outputs while issuing privileged tool calls. Existing defenses often mediate trust at the granularity of an entire tool invocation, forcing a brittle choice in mixed-trust workflows: allow external content to influence a call and risk hijacked destinations or commands, or quarantine the call and block benign retrieval-then-act behavior. The key observation behind this paper is that indirect prompt injection becomes dangerous not when untrusted content appears in context, but when it determines an authority-bearing argument. We present \textsc{PACT} (\emph{Provenance-Aware Capability Contracts}), a runtime monitor that assigns semantic roles to tool arguments, tracks value provenance across replanning steps, and checks whether each argument's origin satisfies its role-specific trust contract. Under oracle provenance, \textsc{PACT} achieves 100% utility and 100% security on mixed-trust diagnostic suites, while flat invocation-level monitors incur false positives or false negatives. In full AgentDojo deployments across five models, \textsc{PACT} reaches 100% security on the three strongest models while recovering 38.1--46.4% utility, 8--16 percentage points above CaMeL at the same security level. Ablations show that both semantic roles and cross-step provenance are necessary. \textsc{PACT} reframes agent security as authority binding, and isolates the remaining deployment bottleneck to provenance inference and contract synthesis.

Tool-integrated reasoning (TIR) enables large language models (LLMs) to enhance their capabilities by interacting with external tools, such as code interpreters (CI). Most recent studies focus on exploring various methods to equip LLMs with the ability to use tools. However, how to further boost the reasoning ability of already tool-capable LLMs at inference time remains underexplored. Improving reasoning at inference time requires no additional training and can help LLMs better leverage tools to solve problems. We observe that, during tool-capable LLM inference, both the number and the proportion of erroneous tool calls are negatively correlated with answer correctness. Moreover, erroneous tool calls are typically resolved successfully within a few subsequent turns. If not, LLMs often struggle to resolve such errors even with many additional turns. Building on the above observations, we propose PruneTIR, a rather effective yet efficient framework that enhances the tool-integrated reasoning at inference time. During LLM inference, PruneTIR prunes trajectories, resamples tool calls, and suspends tool usage through three components: Success-Triggered Pruning, Stuck-Triggered Pruning and Resampling, and Retry-Triggered Tool Suspension. These three components enable PruneTIR to mitigate the negative impact of erroneous tool calls and prevent LLMs from getting stuck in repeated failed resolution attempts, thereby improving overall LLM performance. Extensive experimental results demonstrate the effectiveness of PruneTIR, which significantly improves Pass@1 and efficiency while reducing the working context length for tool-capable LLMs.

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

Audio and vision provide complementary evidence for audio-visual question answering, yet current audio-visual large language models may suffer from cross-modal interference: information from one modality misguides the interpretation of another, thereby inducing hallucinations. We attribute this issue to uncontrolled cross-modal interactions during intermediate reasoning. To mitigate this, we propose Separate First, Fuse Later (SFFL), an audio-visual reasoning framework designed to reduce cross-modal interference. SFFL enforces modality-specific chain-of-thought reasoning, producing separate audio and visual reasoning traces and integrating evidence for answering. We construct modality-preference labels via a data pipeline under different modality input settings. We use these labels as an auxiliary reward in reinforcement learning to encourage a instance-dependent preference for modality cues when answering. We further introduce a modality-specific reasoning mechanism that preserves modality isolation during the separated reasoning stage while enabling full access to cross-modal information at the evidence fusion stage. Experiments demonstrate consistent improvements in both accuracy and robustness, yielding an average relative gain of 5.16% on general AVQA benchmarks and 11.17% on a cross-modal hallucination benchmark.

Next-generation visual assistants, such as smart glasses, embodied agents, and always-on life-logging systems, must reason over an entire day or more of continuous visual experience. In ultra-long video settings, relevant information is sparsely distributed across hours or days, making memory a fundamental challenge: models must accumulate information over time, recall prior states, track temporal order, and abstract recurring patterns. However, existing week-long video benchmarks are primarily designed for perception and recognition, such as moment localization or global summarization, rather than reasoning that requires integrating evidence across multiple days. To address this gap, we introduce EgoMemReason, a comprehensive benchmark that systematically evaluates week-long egocentric video understanding through memory-driven reasoning. EgoMemReason evaluates three complementary memory types: entity memory, tracking how object states evolve and change across days; event memory, recalling and ordering activities separated by hours or days; and behavior memory, abstracting recurring patterns from sparse, repeated observations over the whole week period. EgoMemReason comprises 500 questions across three memory types and six core challenges, with an average of 5.1 video segments of evidence per question and 25.9 hours of memory backtracking. We evaluate EgoMemReason on 17 methods across MLLMs and agentic frameworks, revealing that even the best model achieves only 39.6% overall accuracy. Further analysis shows that the three memory types fail for distinct reasons and that performance degrades as evidence spans longer temporal horizons, revealing that long-horizon memory remains far from solved. We believe EgoMemReason establishes a strong foundation for evaluating and advancing long-context, memory-aware multimodal systems.

Post-training techniques combined with inference-time scaling significantly enhance the reasoning and alignment capabilities of large language models (LLMs). However, a fundamental tension arises: inference-time methods benefit from diverse sampling from a relatively flattened probability distribution, whereas reinforcement learning (RL)-based post-training inherently sharpens these distributions. To address this, we propose Exploration-Driven Optimization (EDO), which extends reward-biasing style exploration objectives to iterative post-training and integrates them into standard RL objectives, encouraging greater diversity in sampled solutions while facilitating more effective inference-time computation. We incorporate EDO into iterative Direct Preference Optimization (iDPO) and Group Relative Policy Optimization (GRPO), resulting in two variants: ED-iDPO and ED-GRPO. Extensive experiments demonstrate that both ED-iDPO and ED-GRPO exhibit greater solution diversity and improved reasoning abilities, particularly when combined with test-time computation techniques like self-consistency. Across three in-distribution reasoning benchmarks, EDO achieves a 1.0-1.3% improvement over the strongest baselines, and delivers an additional 1.5% average gain on five out-of-distribution tasks. Beyond accuracy, EDO preserves model entropy and stabilizes RL training dynamics, highlighting its effectiveness in preventing over-optimization collapse. Taken together, these results establish EDO as a practical framework for balancing exploration and exploitation in LLM reasoning, especially in settings that rely on test-time scaling.

Large reasoning models, such as OpenAI o1 and DeepSeek-R1, tend to become increasingly verbose as their reasoning capabilities improve. These inflated Chain-of-Thought (CoT) trajectories often exceed what the underlying problems require, wasting compute, latency, and context budgets. While introducing length-based efficiency rewards during reinforcement learning offers a natural remedy, existing methods struggle with two fundamental challenges: the optimal balance between correctness and efficiency is non-stationary throughout training, and intrinsic reasoning budgets vary drastically across problems. Relying on static reward weights and global length constraints inevitably forces a compromise between degraded accuracy and unrealized compression. To overcome these limitations, we propose LEAD (Length-Efficient Adaptive and Dynamic reasoning), a method that replaces static heuristics with online, self-adaptive mechanisms. LEAD dynamically calibrates the correctness-efficiency trade-off at each step using a Potential-Scaled Instability, directing optimization capacity to the most informative learning signal. Furthermore, it estimates an adaptive per-problem target length online based on the model's own correct rollouts, applying a symmetric efficiency reward that penalizes both overthinking and over-compression. Evaluated on five mathematical reasoning benchmarks, LEAD achieves the highest accuracy and Accuracy-Efficiency Score among RL-trained efficient-reasoning methods while producing substantially shorter outputs than the base model.

Large-scale 3D vision-language models (VLMs) like LLaVA-3D offer strong spatial reasoning but are difficult to deploy due to high computational costs. We propose a knowledge distillation framework that transfers spatial reasoning from a 7B teacher to a 2.29B student model. Our approach achieves 8.7x lower inference latency and a 3x reduction in model size while retaining 54-72% of the teacher's performance. The framework utilizes VGGT as the vision encoder and a multi-task distillation pipeline with uncertainty-aware loss weighting. To improve reasoning without chain-of-thought (CoT) data, we introduce "Hidden CoT": learnable latent tokens that serve as an internal scratchpad before answer generation. This is the first use of latent scratchpad reasoning in distilled 3D VLMs. The student model jointly performs spatial description, depth estimation, and object detection. Experiments on ScanNet and 3D-FRONT show strong spatial understanding, reaching 68-72% accuracy on proximity and contact tasks. Our framework enables efficient 3D scene QA on resource-constrained platforms.

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

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

Although large language models rely on chain-of-thought for complex reasoning, the overthinking phenomenon severely degrades inference efficiency. Existing reinforcement learning methods compress reasoning chains by designing elaborate reward functions, which renders high-quality samples extremely sparse in the exploration space and creates a sampling bottleneck for the prior policy. Inspired by cognitive science, we theoretically prove that a posterior distribution guided by reference answers achieves higher expected utility than the prior distribution, thus capable of breaking through the sampling bottleneck of high-quality samples. However, the posterior distribution is unavailable during inference. To this end, we formalize efficient reasoning as a variational inference problem and introduce an efficiency-aware evidence lower bound as the theoretical foundation. Based on this, we propose the VPG-EA framework. It adopts a parameter-shared dual-stream architecture to instantiate both the posterior distribution and the prior policy; after filtering out pseudo-efficient paths via cross-view evaluation, it unidirectionally transfers the posterior's efficient patterns to the prior policy through variational distillation. Experiments on DeepSeek-R1-Distill-Qwen-1.5B and 7B scales demonstrate that VPG-EA improves the comprehensive efficiency metric epsilon cubed by 8.73% and 12.37% over the strongest baselines on each model size, respectively.

Large language models (LLMs) have achieved remarkable progress in mathematical reasoning, but this ability is not equally accessible across languages. Especially low-resource languages exhibit much lower reasoning performance. To address this, we propose Crosslingual On-Policy Self-Distillation (COPSD), which transfers a model's own high-resource reasoning behavior to low-resource languages. COPSD uses the same model as student and teacher: the student sees only the low-resource problem, while the teacher receives privileged crosslingual context, including the problem translation and reference solution in English. Training minimizes full-distribution token-level divergence on the student's own rollouts, providing dense supervision while avoiding the sparsity and instability of outcome-only reinforcement learning (RL). Experiments on 17 low-resource African languages show that COPSD consistently improves low-resource mathematical reasoning across model sizes and substantially outperforms Group Relative Policy Optimization (GRPO). Further analyses show that COPSD improves answer-format adherence, strengthens test-time scaling, and generalizes to harder multilingual reasoning benchmarks, with especially large gains for lower-resource languages. We make our code and data available at: https://github.com/cisnlp/COPSD.

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

Chain-of-thought (CoT) prompting assumes that generated reasoning reflects a model's internal computation. We show this assumption is wrong in a specific, measurable way: models internally detect their own reasoning errors but outwardly express confidence in them. A linear probe on hidden states predicts trace correctness with 0.95 AUROC -- from the very first reasoning step (0.79) -- while verbalized confidence for wrong traces is 4.55/5, nearly identical to correct ones (4.87/5). A text-surface classifier achieves only 0.59 on the same data, confirming a 0.20-point gap invisible in the generated text. This hidden error awareness holds across three model families (Qwen, Llama, Phi), 1.5B-72B parameters, and RL-trained reasoning models (DeepSeek-R1, 0.852 AUROC). The natural question is whether this signal can fix the errors it detects. It cannot. Four interventions -- activation steering, probe-guided best-of-N, self-correction, and activation patching -- all fail; patching destroys output coherence entirely. The signal is diagnostic, not causal: a readout of computation quality, not a lever to redirect it. This delineates a boundary for mechanistic interpretability: error representations during reasoning are fundamentally different from the factual knowledge representations that prior work has successfully edited.

Reasoning LLMs produce thousands of chain-of-thought tokens whose KV cache must reside in scarce GPU HBM. The dominant response -- permanently evicting low-importance tokens -- is catastrophic for reasoning: accuracy collapses to 0-2.5% when half the cache is removed. We ask a different question: must every token live in HBM, or can some live elsewhere? We introduce a semantics-aware memory hierarchy that sorts tokens into four tiers -- HBM, DDR, compressed, and evicted -- using cumulative attention scoring. Low-importance tokens are moved to CPU memory rather than destroyed; before each attention step they are prefetched back at full precision, contributing exactly the same terms as if they had never left the GPU. We formalize this as zero-approximation-error offloading and derive our central finding: accuracy depends solely on how many tokens are permanently discarded (the eviction ratio), not on how many remain in HBM. A controlled 3x3 grid over HBM and eviction ratios confirms this across three model scales (7B-32B) and four benchmarks. With only 3% eviction, the hierarchy retains 91% of full-cache accuracy on GSM8K and 71% on MATH-500 (n=200); at 14B scale it matches the uncompressed baseline (90% vs. 86%) while halving HBM occupancy. A head-to-head reproduction of R-KV -- the current SOTA eviction method -- on our setup achieves only 0-32% at comparable budgets. A system prototype with real GPU-CPU data movement shows that the price of this preservation is modest -- 5-7% transfer overhead -- and scaling analysis projects 2-48 GB HBM savings at production batch sizes.

Knowledge Tracing (KT) models students' knowledge states based on learning interactions to predict performance. While deep learning-based KT models have boosted predictive accuracy, most models rely on deterministic vector embeddings and opaque latent state transitions, limiting interpretability regarding how specific past behaviors influence predictions. To address this limitation, we propose Probabilistic Logical Knowledge Tracing (PLKT), an interpretable KT framework that formulates prediction as a goal-conditioned evidence reasoning process over historical learning behaviors. Instead of representing knowledge states as deterministic vector embeddings, PLKT employs robust Beta-distributed probabilistic embeddings to represent student knowledge states. This probabilistic foundation allows us to model the uncertainty of historical behaviors and perform explicit logical operations (e.g., conjunction), constructing transparent reasoning paths that reveal how specific past interactions contribute to the prediction. Extensive experiments show that PLKT outperforms state-of-the-art KT methods while achieving superior interpretability. Our code is available at https://anonymous.4open.science/r/PLKT-D3CE/.

The Chain-of-Thought (CoT) paradigm, while enhancing the interpretability of Large Language Models (LLMs), is constrained by the inefficiencies and expressive limits of natural language. Latent Chain-of-Thought (latent CoT) reasoning, which operates in a continuous latent space, offers a promising alternative but faces challenges from structural complexities in existing multi-step or multi-model paradigms, such as error propagation and coordination overhead. In this paper, we introduce One-Model One-Step, a novel compression framework for Latent Reasoning with Rule-Based Priors(RuPLaR) to address this challenge. Our method trains an LLM to autonomously generate latent reasoning tokens in a single training stage, guided by rule-based prior probability distributions, thereby eliminating cascaded processes and inter-model dependencies. To ensure reasoning quality, we design a joint training objective that enforces answer consistency via cross-entropy, aligns soft tokens with rule-based priors via KL divergence (the Soft Thinking constraint), and adds a problem-thought semantic alignment constraint in the representation space. Extensive experiments show that our compression framework not only improves accuracy by 11.1% over existing latent CoT methods but also achieves this with minimal token usage, underscoring its effectiveness and extensibility. Code: https://github.com/xiaocen-luo/RuPLaR.

Large language models fail at counting repeated tokens despite strong performance on broader reasoning benchmarks. These failures are commonly attributed to limitations in internal count tracking. We show this attribution is wrong. Linear probes on the residual stream decode the correct count with near-perfect accuracy at every post-embedding layer, across all model depths. This holds even at the exact layers where the wrong answer crystallizes while the model simultaneously outputs an incorrect count. Attention patterns show no evidence of collapse over repeated tokens and tokenization artifacts account for none of the failure. Instead, a format-triggered multi-layer perceptron (MLP) block overwrites the correctly-encoded count with a fixed wrong answer at roughly 88--93,% network depth. This prior fires for repeated word-tokens in space-separated list format and is absent for repeated digit-tokens. It is suppressed by comma-separated delimiters in larger models but persists in smaller ones. The finding holds across Llama-3.2 (1B and 3B) and Qwen2.5 (1.5B, 3B and 7B) at consistent relative depth. Counting failure is a failure of routing not of representation and the two require different interventions.

Recent work shows that large language models (LLMs) encode behavioural traits ("personas") as linear directions in activation space, often called "persona vectors". Prior work has used such directions as static handles for behavioural steering. Building on this, we treat them as dynamic signals instead: probes we can monitor and intervene on as reasoning unfolds. We use the term polylogue to denote the time series of alignments between persona vectors and hidden activations over the course of generation. Experiments across four open-weight models show that polylogue features predict correctness on MMLU-Pro competitively with low-dimensional activation baselines, while remaining interpretable through their associated persona directions. They also suggest concrete steering targets, namely which latent directions to modulate at different stages of a response. We instantiate this as a simple paragraph-conditioned intervention that improves accuracy on three of four models, pointing to stage-aware latent steering as a promising direction for reasoning-time control. Together, this positions the polylogue as an interpretable tool for reasoning-time monitoring and intervention.

Humanoid Visual Search (HVS) requires agents to actively explore immersive 360$^\circ$ environments. While prior methods treat this as a monolithic task relying on cumulative, multi-turn Chain-of-Thought (CoT) reasoning, they impose heavy cognitive burdens and require expensive trajectory-level annotations. In this paper, we propose Imagining in 360$^\circ$, a novel framework that decouples the exploration process into a specialized Imaginator and an Actor. The Imaginator functions as a probabilistic predictor of spatial priors; instead of maintaining a cumulative reasoning chain, it infers the semantic layout of both observed and unobserved regions in a single step. By sampling multiple hypotheses within this semantic space, we provide the Actor with a distribution of effective spatial information, offering robust guidance that hedges against uncertainty during active search. This decoupled architecture significantly lowers data engineering costs by eliminating the need for full-trajectory CoT annotations, enabling the generation of over 1.96 million curated training samples. Extensive experiments demonstrate that explicitly modeling semantic spatial priors drastically improves search efficiency and success rates in complex, in-the-wild environments.

Despite surpassing human performance across mathematics, coding, and other knowledge-intensive tasks, large language models (LLMs) continue to struggle with causal reasoning. A core obstacle is the target data itself: causal systems are complex and often expressed in non-executable forms, while ground-truth answers to causal queries are inherently scarce. We introduce CauSim, a framework that turns causal reasoning from a scarce-label problem into a scalable supervised one. CauSim constructs increasingly complex causal simulators: executable structural causal models (SCMs), incrementally built by LLMs, that scale to globally complex systems while maintaining verifiable answers to causal queries. CauSim operates across representations by formalizing non-executable causal knowledge into code, enabling data augmentation, and translating executable SCMs into natural language, enabling supervision in previously difficult-to-supervise representations. We structure our research into two parts: (1) how to construct increasingly complex causal simulators, and (2) a systematic study of what CauSim enables, demonstrating generalization across representations, consistent gains from curriculum scaling and data volume, LLM self-improvement through self-generated simulators, and data augmentation via formalization of existing domain knowledge.

Following the recent achievement of gold-medal performance on the IMO by frontier LLMs, the community is searching for the next meaningful and challenging target for measuring LLM reasoning. Whereas olympiad-style problems measure step-by-step reasoning alone, research-level problems use such reasoning to advance the frontier of mathematical knowledge itself, emerging as a compelling alternative. Yet research-level math benchmarks remain scarce because such problems are difficult to source (e.g., Riemann Bench and FrontierMath-Tier 4 contain 25 and 50 problems, respectively). To support reliable evaluation of next-generation frontier models, we introduce Soohak, a 439-problem benchmark newly authored from scratch by 64 mathematicians. Soohak comprises two subsets. On the Challenge subset, frontier models including Gemini-3-Pro, GPT-5, and Claude-Opus-4.5 reach 30.4%, 26.4%, and 10.4% respectively, leaving substantial headroom, while leading open-weight models such as Qwen3-235B, GPT-OSS-120B, and Kimi-2.5 remain below 15%. Notably, beyond standard problem solving, Soohak introduces a refusal subset that probes a capability intrinsic to research mathematics: recognizing ill-posed problems and pausing rather than producing confident but unjustified answers. On this subset, no model exceeds 50%, identifying refusal as a new optimization target that current models do not directly address. To prevent contamination, the dataset will be publicly released in late 2026, with model evaluations available upon request in the interim.

Chain-of-Thought (CoT) prompting symbolized a huge improvement of reasoning capabilities of Large Language Models (LLMs). However, scaling up test-time computation yields extensive CoT sequences, introducing severe inference latency and key-value (KV) cache memory bottlenecks. While structural pruning offers a fundamental, hardware-aware solution to alleviate static parameter burdens, existing magnitude-based methods may cut off the neurons of CoT: by over-indexing on discrete cross-entropy objectives, these heuristics fall into a \textit{magnitude trap}: they prioritize high-frequency, low-information syntactic tokens and trigger a disappointing reasoning collapse at high sparsities (e.g., 40%). To overcome this topological phase transition, we propose \textsc{Relative Kinetic Utility} (RKU), a novel theoretical framework that elevates discrete pruning to a continuous kinetic integral over the depth manifold of the model based on Alternating Gradient Flow(AGF). By modifying it with Fisher trace normalization, RKU acts as a lightweight curvature-aware normalization to isolate \textit{kinetic spikes} -- the fundamental structural pathways responsible for high-curvature logical routing. Extensive experiments on Qwen-2.5-7B and LLaMA-3-8B improves performance in the high-sparsity regime around 40%. RKU attains 13.34% accuracy on GSM8K at 40% sparsity, outperforming the strongest baseline, and appears to better preserve reasoning-relevant representations under out-of-distribution evaluation.

Reasoning-based end-to-end (E2E) autonomous driving has recently emerged as a promising approach to improving the interpretability of driving decisions as it can generate human-readable reasoning together with predicted trajectories. Such approaches commonly generate multiple trajectories to capture diverse future behaviors, and they fall into two categories: (1) multi-reasoning, where one reasoning sequence is generated per trajectory, and (2) single-reasoning, where a single reasoning is shared across all trajectories. The former offers richer diversity at the cost of redundant computation, while the latter is more efficient but is often assumed to sacrifice diversity. Alpamayo 1, a representative system, adopts the multi-reasoning approach and achieves competitive trajectory prediction performance. However, the efficiency of this design remains largely unexplored, making it a well-motivated subject for investigation. In this paper, we systematically analyze and improve Alpamayo 1 in two ways. First, we reduce inference latency while preserving trajectory diversity by redesigning Alpamayo 1 into a single-reasoning system. Through extensive experiments, we find that replacing multi-reasoning with single-reasoning does not meaningfully degrade trajectory diversity. Second, we accelerate diffusion-based action generation by eliminating inter-block overhead arising from unnecessary copy operations and inefficient kernel execution. Through closed-loop and open-loop experiments, we validate both optimizations, demonstrating a 69.23% reduction in inference latency while maintaining trajectory diversity and prediction quality. These results highlight the importance of jointly analyzing system architecture and runtime execution to improve the efficiency of reasoning-based E2E AD systems.

Large Reasoning Models possess remarkable capabilities for self-correction in general domain; however, they frequently struggle to recover from unsafe reasoning trajectories under adversarial attacks. Existing alignment methods attempt to mitigate this vulnerability by fine-tuning the model on expert data including reflection traces or adversarial prefixes. Crucially, these approaches are often hindered by static training data which inevitably deviate from model's dynamic, on-policy reasoning traces, resulting in model hardly covering its vast generation space and learning to recover from its own failures. To bridge this gap, we propose Self-ReSET, a pure reinforcement learning framework designed to equip LRMs with the intrinsic capacity to recover from their own safety error trajectories, which are subsequently reused as an initial state for reinforcement learning. Extensive experiments across various LRMs and benchmarks demonstrate that Self-ReSET significantly enhances robustness against adversarial attacks especially out-of-distribution (OOD) jailbreak prompts while maintaining general utility, along with efficient data utilization. Further analysis reveals that our method effectively fosters self-recovery patterns, enabling models to better identify and recover from unsafe intermediate error states back to benign paths. Our codes and data are available at https://github.com/Ing1024/Self-ReSET.

While explicit Chain-of-Thought (CoT) empowers large reasoning models (LRMs), it enables the generation of riskier final answers. Current alignment paradigms primarily rely on externally enforced compliance, optimizing models to detect malicious prompts rather than evaluating the safety of their own outputs. We argue that this approach remains largely behavioral: our empirical analysis reveals that ostensibly aligned models lack intrinsic safety understanding, often failing to verify their own response safety and remaining vulnerable to adversarial jailbreaks. To address this fundamental limitation, we propose Safety Internal (SInternal), a framework that internalizes safety specifications by training LRMs exclusively on safety verification tasks to critique their own generated answers using expert reasoning trajectories. We demonstrate that learning to verify induces a strong generalization for response safety, significantly enhancing robustness against out-of-domain jailbreaks. Furthermore, when combined with reinforcement learning, SInternal serves as a superior initialization compared to standard supervised fine-tuning, suggesting that internalizing safety understanding creates a more robust foundation for alignment than merely mimicking safe behaviors. Our codes are available at https://github.com/AlphaLab-USTC/SInternal

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

Large Language Models (LLMs) are increasingly deployed as autonomous agents that execute tool-augmented, multi-step tasks, where latency is a critical factor for real-world applications. Yet an overlooked threat is Reasoning-Level Denial-of-Service (R-DoS), in which an attacker preserves task correctness but degrades availability by inflating an agent's reasoning depth or tool-use budget. We introduce OTora, the first unified, two-stage red-teaming framework for instantiating R-DoS attacks. Stage I optimizes an adversarial trigger that induces targeted tool invocations using insertion-aware scoring and dynamic target co-evolution, supporting both black-box and white-box settings. Stage II generates agent-aware reasoning payloads via an ICL-guided genetic search that amplifies overthinking while maintaining correct task outcomes. Across WebShop, Email, and OS agents built on multiple backbone models such as LLaMA-70B and GPT-OSS-120B, OTora achieves up to 10 times increases in reasoning tokens and order-of-magnitude latency slowdowns, all while preserving near-baseline task accuracy. Finally, we discuss mitigation strategies for detecting and constraining abnormal reasoning and latency spikes. The code is available at https://github.com/llm2409/OTora.

Larger language models consistently outperform smaller ones on reasoning benchmarks, yet the reasoning differences underlying this gap remain underexplored. Across benchmarks in mathematics, physics, chemistry, and programming, we observe stable performance gaps: averaged over datasets, Qwen3-32B outperforms Qwen3-8B by 6.43%, while GPT-OSS-120B exceeds GPT-OSS-20B by 7.38%. To study the reasoning differences behind these gains, we develop AdvCluster, an automated framework that identifies questions where the larger model shows a stable advantage, extracts fine-grained advantage descriptions from paired reasoning traces produced by larger and smaller models, and organizes them through semantic clustering with quantitative evaluation and selection guided by a reviewer model. Our analysis yields a systematic taxonomy of larger model reasoning advantages, spanning both common advantages that recur across domains and specialized advantages associated with particular domains. Across these patterns, a recurring theme is Constraint-Guided Reasoning: larger models are better at identifying explicit and implicit constraints, organizing them into structured reasoning, and using them to rule out infeasible paths and verify intermediate steps.

Large language model agents increasingly operate through environment-facing scaffolds that expose files, web pages, APIs, and logs. These observations influence tool use, state tracking, and action sequencing, yet their reliability and authority are often uncertain. Environmental grounding is therefore a systems-level problem involving context admission, evidence provenance, freshness checking, verification policy, action gating, and model reasoning. Existing agent benchmarks mainly evaluate task capability or specific attacks such as prompt injection and memory poisoning, but they under-specify a fundamental reliability question: whether agents remain grounded in the true environment state when observations are stale, incorrect, or malicious. We introduce EnvTrustBench, an agentic framework for benchmarking this failure mode. We define an evidence-grounding defect (EGD) as a behavioral failure in which an agent treats an environment-facing claim as sufficient evidence for action without resolving it against available current evidence, leading to a task-incorrect false path under the true environment state. Given a task scenario, EnvTrustBench generates the workspace, environment, agent-facing objective, and validation oracle, executes the evaluated agent, records its action-observation trajectory and final state, and applies the oracle to produce a verdict. Using 6 LLM backbones and 5 widely used scaffolds, we evaluate 55 generated cases across 11 task scenarios, with each scenario expanded through five feedback-guided generation iterations. Results show that EGDs consistently emerge across operational workflows, highlighting environmental grounding as a core agent reliability problem with important security implications.

Reasoning-centric large language models (LLMs) achieve strong performance by generating intermediate reasoning trajectories, but often incur excessive token usage and high inference-time decoding cost. We observe that, when solving the same problems, larger reasoning models can often produce more concise traces, whereas smaller reasoning models tend to generate longer and more redundant trajectories. This is especially problematic in real-world deployment, where memory, latency, and serving-cost constraints often favor smaller models. Our observations suggest that reasoning compression can be transferred from large models to small ones rather than enforced through explicit length constraints. Based on this insight, we propose Mixed-Policy Distillation (MPD), a reasoning compression framework that transfers concise reasoning behavior from a larger-sized teacher to a smaller student by distilling teacher-compressed student trajectories. Unlike on-policy distillation, which aligns the student with teacher distributions over verbose student trajectories, or off-policy distillation, which relies on teacher-generated trajectories and may suffer from distribution mismatch, MPD combines the strengths of both. Given a student-sampled trajectory, the teacher rewrites it into a more concise reasoning trace, and the student is trained via KL-based alignment on the compressed trajectory. This preserves student-policy exploration while injecting teacher-guided compression. Experiments on Qwen3-1.7B show that MPD reduces token usage by up to 27.1% while improving performance across multiple reasoning benchmarks, demonstrating an effective approach to efficient small-model reasoning.

Multi-agent reasoning has shown promise for improving the problem-solving ability of large language models by allowing multiple agents to explore diverse reasoning paths. However, most existing multi-agent methods rely on inference-time debate or aggregation, which can be vulnerable to incorrect peer influence and biased consensus. Moreover, the agents themselves remain static, as their underlying reasoning skills do not evolve across tasks. In this paper, we introduce \textbf{AgentPSO}, a particle-swarm-inspired framework for evolving multi-agent reasoning skills. AgentPSO treats each agent as a particle-like reasoner whose state is a natural-language skill and whose velocity is a semantic update direction, iteratively guiding agents toward higher-performing skill configurations. Across training iterations, each agent updates its skill by combining its previous velocity, personal-best skill, global-best skill, and a self-reflective direction derived from peer reasoning trajectories. This enables agents to learn reusable reasoning behaviors by drawing on their own experience and on the strongest skills found by the population, without updating the parameters of the backbone language model. Experiments on mathematical and general reasoning benchmarks show that AgentPSO improves over static single-agent skills and test-time-only multi-agent reasoning baselines. The evolved skills further transfer across benchmarks and to another backbone model, suggesting that AgentPSO captures reusable reasoning procedures rather than merely optimizing benchmark-specific prompts. Code is publicly available at https://github.com/HYUNMIN-HWANG/AgentPSO/.

Large reasoning models achieve high accuracy through extended chain-of-thought but generate 5--8 more tokens than necessary, applying verbose reasoning uniformly regardless of problem difficulty. We propose Hint Tuning, a data-efficient approach that teaches models to calibrate reasoning depth. Our key insight: the corresponding instruct model serves as an ideal difficulty probe. By testing what the instruct model can solve with varying guidance, we automatically construct training data across three states: No-Hint (direct answer), Sparse-Hint (minimal prefix), and Full-Hint (complete reasoning). This converts the abstract challenge of difficulty labeling into a measurable consistency check between the instruct and reasoning models. With only 1K self-annotated samples, Hint Tuning achieves 24--66% token reduction (31.5% average) across mainstream reasoning models (Qwen3-Thinking, DeepSeek-R1-Distill) at multiple scales (4B--32B) while maintaining competitive accuracy on five benchmarks. Unlike methods requiring massive distillation datasets or expensive RL, we achieve superior efficiency through simple alignment with the instruct model's capabilities. Code and data are available at https://github.com/redai-infra/hint-tuning.

Traditional simulation methods reproduce occurred emergency instances through presetting to assist people in risk assessment and emergency decision-making. However, due to the lack of randomness and diversity, existing simulation systems struggle to fully explore the potential risk as emergency instances are scarce. In contrast, Large Models (LMs) can dynamically adjust generation strategies to introduce controllable randomness, while also possessing extensive prior knowledge and cross-domain knowledge transfer capabilities. Inspired by it, we propose the LMs-driven World Line Divergence System (WLDS), which enables diversified visualization and deduction of emergency instances in different domains. WLDS leverages LMs to deduce emergency instances in various development directions, and introduces the factual calibration and logical calibration mechanism to ensure factual accuracy and logical rigor during the deduction process. The interactive module can independently select deduction directions to avoid potential hallucinations that are difficult for the system to identify. Furthermore, by introducing the visualization module, WLDS forms simulation and deduction that combine text and images, which enhances interpretability. Extensive experiments conducted on the proposed Emergency Instances Deduction (EID) benchmark dataset demonstrate that WLDS achieves high-precision and high-fidelity simulation and deduction of emergency instances in multiple specific domains. Relevant experiments further demonstrate that WLDS can generate more emergency instances deduction data for users and provide support for better decision-making in similar emergency instances in the future.

We introduce MathConstraint, a hard, adaptive benchmark for evaluating the combinatorial reasoning capabilities of LLMs. We combine constraint satisfaction problems with rigorous solver-based verification and design an adaptive generator to create instances that remain challenging as the LLMs improve in their reasoning capabilities. Unlike existing benchmarks that quickly saturate on fixed datasets or use LLM-as-a-judge for checking solutions,MathConstraint uses parameterized problem types that enable scalable generation of arbitrarily difficult and automatically verifiable instances. We release MathConstraint-Easy ($266$ instances), on which frontier models achieve between $72.6\%$ (gemini-3.1-flash-lite) and $87.6\%$ (gpt-5.5) accuracy, and MathConstraint ($329$ instances) on which the same models drop to between $18.5\%$ (claude-4.6-sonnet) and $66.9\%$ (gpt-5.5) accuracy, demonstrating the resilience of our benchmark generator against rapid progress in LLM reasoning capabilities. We evaluate 12 frontier and open-weight models with and without access to a sandboxed Python environment that includes generic SAT/SMT solvers. Tool access roughly doubles frontier accuracy on MathConstraint (mean $+28$pp; up to $+52$pp for claude-4.6-sonnet). Further, halving the tool-call budget from $8$ to $4$ rounds erases up to $37$ points -- a sensitivity that most single-budget benchmarks miss. We release the generator, dataset, and evaluation harness as a robust environment for studying combinatorial reasoning and tool-use behavior under adversarially-tunable difficulty.

The effectiveness of Reinforcement Learning (RL) in Large Language Models (LLMs) depends on the nature and diversity of the data used before and during RL. In particular, reasoning problems can often be approached in multiple ways that rely on different forms of reasoning, and exposure to only a limited range of such approaches in the training data may limit the effectiveness of RL. Motivated by this, we investigate using diverse self-generated data during mid-training as an intermediate step before RL training. Specifically, we adopt a bootstrapped data-generation framework guided by George Polya's problem-solving approaches for generating multiple variants of correct answers for each question in the training data, and then perform fine-tuning. We first provide a theoretical perspective on how mid-training on such data improves RL and explain how policy-gradient updates can incentivize combining multiple approaches. We then empirically demonstrate that RL-trained models initialized with our mid-training data achieve consistent improvements across various mathematical reasoning benchmarks and other OOD tasks like code generation and narrative reasoning. Overall, our investigative study shows that a language model learning multiple problem-solving approaches, through self-generated data helps subsequent RL.

Recent advances in large language models (LLMs) have demonstrated remarkable reasoning capabilities, largely stimulated by Reinforcement Learning with Verifiable Rewards (RLVR). However, existing RL algorithms face a fundamental limitation: their exploration remains largely constrained by the inherent capability boundary of the policy model. Although recent methods introduce external expert demonstrations to extend this boundary, they typically rely on complete trajectory-level guidance, which is sample-inefficient, information-sparse, and may confine exploration to a static guidance space. Inspired by the potential of multi-agent systems, we propose $\textbf{AIPO}$, an enhanced reinforcement learning framework that improves LLM reasoning through active multi-agent interaction during exploration. Specifically, AIPO enables the policy model to proactively consult three functional collaborative agents, $\textit{Verify Agent}$, $\textit{Knowledge Agent}$, and $\textit{Reasoning Agent}$, when encountering reasoning bottlenecks, thereby receiving fine-grained and targeted guidance to actively expand its capability boundary during training. We further introduce a tailored importance sampling coefficient together with a clipping strategy to mitigate the off-policy bias and gradient vanishing issues that arise when learning from agent-provided feedback. After training, the policy model performs reasoning independently without relying on collaborative agents. Extensive experiments on diverse reasoning benchmarks, including AIME, MATH500, GPQA-Diamond, and LiveCodeBench, show that AIPO consistently improves reasoning performance, generalizes robustly across different policy models and RLVR algorithms, and effectively expands the reasoning capability boundary of the policy model.

Hallucination detection methods for large language models increasingly operate on chain-of-thought reasoning traces, yet it remains unclear whether they evaluate the reasoning itself or merely exploit surface correlates of the final answer. We introduce a controlled-invariance methodology that exposes this distinction through two oracle tests: \textsc{Force}, which replaces each response's final answer with the ground truth while preserving the reasoning trace, and \textsc{Remove}, which strips answer-announcement steps while leaving the trajectory intact. This reveals if their predictive power derives from answer-level artifacts rather than from the structure or validity of intermediate reasoning. We further show that once these artifacts are controlled for, effective detection does not necessarily require complex learned representations: TRACT, a lightweight scorer built on lexical trajectory features (hedging trends, step-length dynamics, and cross-response vocabulary convergence), achieves strong robustness while remaining competitive with or outperforming existing baselines on unperturbed traces. These findings suggest that the current central challenge in reasoning-aware hallucination detection is not the absence of signal in the trace, but the failure to isolate it from endpoint cues.