Large Language Model Reasoning

Effectively training Large Language Models (LLMs) for complex, long-CoT reasoning is often bottlenecked by the need for massive high-quality reasoning data. Existing methods are either computationally expensive or fail to reliably distinguish high- from low-quality reasoning samples. To address this, we propose High-Entropy Sum (HES), a training-free metric that quantifies reasoning quality by summing only the entropy of the top (e.g., 0.5%) highest-entropy tokens in each reasoning sample. We validate HES across three mainstream training paradigms: Supervised Fine-tuning (SFT), Rejection Fine-tuning (RFT), and Reinforcement Learning (RL), with extensive results demonstrating its consistent effectiveness and significantly reduced computational overhead. In SFT, training on the top 20% HES-ranked data matches full-dataset performance, while using the lowest-HES data degrades it. In RFT, our HES-based training approach significantly outperforms baseline methods. In RL, HES-selected successful trajectories enable the model to learn strong reasoning patterns, significantly surpassing other compared methods. Our findings establish HES as a robust, training-free metric that enables a unified, effective, and efficient method for developing advanced reasoning in LLMs.

On-policy self-distillation (OPSD) is an emerging LLM post-training paradigm in which the model serves as its own teacher: conditioned on privileged information such as a reference trace or hint, the same policy provides dense token-level supervision on its own rollouts. However, recent studies show that OPSD degrades complex reasoning by suppressing predictive uncertainty, which supports exploration and hypothesis revision. Our token-level analysis shows that this failure arises from applying a uniform direction of teacher supervision across tokens with different uncertainty levels: conformity to the privileged self-teacher suppresses exploration at high entropy, while deviation from the teacher degrades step accuracy at low entropy. Accordingly, we propose \textbf{Direction-Adaptive Self-Distillation} (\textbf{DASD}), which reframes privileged self-distillation from uniform teacher imitation into entropy-routed directional supervision: high-entropy tokens are pushed away from the privileged teacher to preserve exploration, while low-entropy tokens are pulled toward the teacher to stabilize step-level execution. Across six mathematical reasoning benchmarks, DASD achieves the best macro Avg@16 over strong RLVR and self-distillation baselines. Pass@$k$, reasoning-health, and generalization analyses show that these average gains come from preserving exploration without sacrificing step-level execution.

Graph of Thoughts (GoT), a generalized form of recent prompting paradigms for large language models (LLMs), has been shown to be useful for elaborate problem solving. By executing a graph of operations, thoughts of the LLM are structured as an arbitrary graph, forming the actual graph of thoughts. Originally, the graph of operations is defined manually, which requires in-depth knowledge about the solution of the problem to solve. Such a static graph of operations is rigid and therefore lacks adaptability. We propose Reinforced Graph of Thoughts (RGoT), an automated approach to the GoT prompting paradigm that leverages reinforcement learning (RL) to adaptively generate a graph of operations from a human-defined set. Results indicate that, under certain constraints, it is possible to construct graphs of operations adaptively to the task's complexity in an automated way.

Recent progress in LLM reasoning has increasingly shifted from single-pass generation to explicit search over intermediate reasoning states. Tree-of-Thoughts (ToT) organizes inference to tree-structured search with branching and backtracking, but it substantially amplifies the Key--Value (KV) cache: retaining KV states for a frontier of partial trajectories quickly becomes a memory bottleneck that limits throughput and constrains search depth and width under fixed hardware budgets. We address this challenge by observing that KV reuse in ToT-style inference is governed by search dynamics: near-term decoding depends primarily on the active branch and its ancestors, whereas inactive subtrees have low short-term reuse probability yet must remain recoverable for backtracking. Motivated by this, we propose ArborKV, a structure-aware eviction framework that couples a lightweight value estimator with a tree-aware allocation policy, and performs purely token-extractive eviction with lazy rehydration to support revisits. Experiments on ToT-style reasoning benchmarks show that ArborKV achieves up to ~4x peak KV-memory reduction while preserving near-full-retention accuracy, enabling larger search configurations under fixed device budgets that would otherwise run out of memory.

Reinforcement learning from verifiable rewards (RLVR) has shown strong promise for LLM reasoning, but outcome-based RLVR remains inefficient on hard problems because correct final-answer rollouts are rare and sample-level credit assignment cannot use partial progress in failed attempts. We introduce SCRL (Subproblem Curriculum Reinforcement Learning), a curriculum RL framework that derives verifiable subproblems from reference reasoning chains and fixes the final subproblem as the original problem. This turns partial progress on hard problems into verifiable learning signals. Algorithmically, SCRL uses subproblem-level normalization, which normalizes rewards independently at each subproblem position and assigns the resulting advantages to the corresponding answer spans, enabling finer-grained credit assignment without external rubrics or reward models. Our analysis shows that subproblem curricula lift hard problems out of gradient dead zones, with larger relative gains as the original problem becomes harder. Across seven mathematical reasoning benchmarks, SCRL outperforms strong curriculum-learning baselines, improving average accuracy over GRPO by +4.1 points on Qwen3-4B-Base and +1.9 points on Qwen3-14B-Base. On AIME24, AIME25, and IMO-Bench, SCRL further improves pass@1 by +3.7 points and pass@64 by +4.6 points on Qwen3-4B-Base, indicating better exploration on hard reasoning problems.

Financial markets are characterized by extreme non-stationarity, low signal-to-noise ratios, and strong dependence on external information such as news, company fundamentals, and macroeconomic signals. Yet, existing approaches either abstract time-series into text or decouple forecasting from language-based reasoning, leading to a fundamental mismatch between qualitative reasoning and quantitative outcomes. To address this, we introduce StockR1, a time-series-enhanced LLM that unifies stock forecasting and financial reasoning through a verifiable forecast action. Based on a tool-call design, the model first emits a forecast action, which is a structured and interpretable representation of its qualitative market outlook. It then invokes a time-series decoder conditioned on this action to generate distributional future trajectories, leading to more informed question answering and financial reasoning. We optimize the full pipeline with reinforcement learning, where rewards jointly reflect answer validity, forecast accuracy, and consistency between generated actions and observed time-series dynamics. In addition, rewards are reweighted by a sample-level uncertainty scalar, encouraging the model to accommodate varying uncertainty in market dynamics. We evaluate StockR1 on financial question answering and stock forecasting over a large-scale 10-year benchmark. Our method consistently outperforms time-series baselines and general-purpose LLMs, improving reasoning accuracy by 17.7% (4B) and 25.9% (8B). These findings demonstrate that structuring the forecast actions establishes a powerful synergy between language reasoning and temporal prediction, enabling LLMs to reason through verifiable, interpretable, and numerically grounded decisions.

Reinforcement learning with verifiable rewards has become the standard recipe for improving LLM reasoning, but the dominant algorithm GRPO assigns a single trajectory-level advantage to every token, diluting the signal at pivotal reasoning steps and injecting noise at uninformative ones. Critic-free alternatives derived from on-policy distillation supply per-token signals through oracle-conditioned likelihood ratios, yet apply each signal in isolation from the trajectory-level evidence accumulated up to that position. We propose Oracle-Prompted Policy Optimization (OPPO), which rests on a single observation: the oracle signal used by prior distillation-style methods for local discrimination is also the natural Bayesian update of the model's belief about eventual success. Accumulating the signal along a trajectory yields, in closed form and at the cost of one extra forward pass, a running estimate of the success probability at every position, together with a token-level advantage that requires no learned value network and no additional rollouts. A first-order analysis factorizes the advantage into the per-token discrimination signal used by distillation methods modulated by a state weight that concentrates credit on genuinely pivotal tokens, with a directional variance-reduction guarantee. The framework admits two estimators differing only in which model scores the evidence: a \textit{self-oracle} that reuses the student and recovers the on-policy distillation reward as a strict special case, and a \textit{teacher-oracle} that delegates scoring to a stronger frozen model. On two base LLMs across seven mathematics, science, and code reasoning benchmarks, OPPO improves over GRPO, DAPO, and SDPO by up to $+6.0$ points on AMC'23 and $+5.2$ points on AIME'24, with gains that widen monotonically with response length.

Large language models frequently produce errors in reasoning tasks despite possessing the underlying knowledge required for correct reasoning. One possible approach to improve reasoning consistency is through activation steering. However, existing activation steering approaches apply fixed, pre-computed correction vectors, ignoring where the model currently sits along its generation trajectory; the result is indiscriminate perturbation that disrupts already-correct steps as freely as erroneous ones. We propose Manifold-Guided Attention Steering (MAGS), a trajectory-aware inference-time intervention grounded in a geometric observation: the output activations of specific attention heads diverge from a low-dimensional correctness manifold at the point of error, and this deviation compounds through subsequent steps. For each identified attention head, we learn a low-dimensional subspace from contrastive pairs of correct and incorrect traces that capture the directions along which error behavior deviates from correct behavior. During inference, we monitor each head's proximity to this manifold and apply a targeted projection correction when deviation exceeds a learned threshold, steering the attention output back toward the correct subspace before the error propagates. MAGS consistently outperforms both unsteered baselines and static steering approaches across benchmarks spanning mathematical reasoning (MATH-500, GSM8K), code generation (HumanEval, MBPP), and molecular generation (SMILES), suggesting that correctness manifolds are a general feature of LLM attention geometry.

Humans effortlessly go beyond literal meanings: If you mow the lawn, I will give you fifty dollars, is typically understood as implying that the speaker will pay only if the lawn is mowed, whereas If you are hungry, there is pizza in the oven implies that pizza is available regardless of the hearers hunger. Large Language Models - LLMs - show human-like performance on many tasks, yet it remains unclear whether they reason like humans. To address this, we conducted a population-matching experiment assessing how twentyfive LLMs compute conditional inferences across four languages, compared to an equal number of humans per language. We find that humans enrich logical reasoning through pragmatic inferences across languages. Model behavior is more variable. Some LLMs perfectly follow the truth-table of conditionals but they ignore pragmatic inferences, while others deviate from the truth-table, adhering to a single interpretation across the board, thus reflecting accurate rule-based processing but not human-like reasoning. Overall, LLMs are accurate semantic operators, but fail to capture the pragmatic enrichments characteristic of human reasoning. Crucially, LLM accuracy is neither predicted nor boosted by open vs. closed status, training orientation, or architecture type, suggesting that pragmatic reasoning is still an emerging ability in the cognitive toolkit of artificial systems.

Reinforcement learning with verifiable rewards (RLVR) has become an effective paradigm for improving reasoning language models on tasks such as mathematics, coding, and scientific question answering. However, widely used group-relative objectives, such as GRPO, summarize each sampled group with scalar statistics and therefore discard fine-grained relational information among candidate responses. This weakens credit assignment under sparse outcome rewards, especially when multiple generated solutions differ only subtly in reasoning quality. We propose \textbf{LamPO}, a \textbf{Lambda-Style Policy Optimization} method that replaces scalar group advantages with a \emph{Pairwise Decomposed Advantage}. LamPO aggregates pairwise reward gaps within each response group and modulates each comparison by a confidence-aware weight computed from sequence log-probability differences, while retaining the critic-free and clipped-update structure of PPO-style optimization. When reference solutions are available, we further add a lightweight ROUGE-L-based dense auxiliary reward to reduce reward sparsity. Experiments on AIME24, AIME25, MATH-500, and GPQA-Diamond with Qwen3-1.7B, Qwen3-4B, and Phi-4-mini show that LamPO consistently improves over GRPO and recent RLVR variants, with more stable training dynamics and better sample efficiency.

Reinforcement learning improves LLM reasoning, but PPO/GRPO typically use fixed clipping and decoding temperature, which makes training brittle and tuning-heavy. We propose Adaptive Group Policy Optimization (AGPO), a critic-free refinement of GRPO that uses group-level statistics to control both update magnitude and exploration. AGPO uses a shared probe-derived statistical state to drive two controllers: (i) adaptive clipping, which sets the trust-region size from reward dispersion and skewness, probe vote entropy, policy entropy, and step-wise KL drift; and (ii) bidirectional adaptive temperature sampling, which heats or cools decoding around a base temperature according to centered uncertainty relative to a running baseline. On nine English and Chinese math/STEM benchmarks, Qwen2.5-14B trained with AGPO outperforms PPO/GRPO under the same generated-token budget, reaching 67.3% on GSM8K and 40.5% on MATH. Gains transfer to Llama-3-8B and Gemma-2-9B, and ablations confirm both modules are complementary. Our implementation is publicly available at https://github.com/wandugu/paper_agpo.

Chain-of-thought (CoT) reasoning has become the default strategy for enhancing LLM capabilities, yet its application raises a fundamental question: when is explicit reasoning actually beneficial? Empirical evidence reveals a striking paradox: CoT often provides marginal or even negative gains on factual and open-ended tasks while multiplying token consumption. In this work, we show that LLM reasoning is not a static property of tasks or models, but a \emph{dynamic decoding state} that emerges during generation. Through systematic analysis, we find early-stage entropy dynamics provide a reliable signal of this state: tasks benefiting from CoT exhibit consistent entropy reduction, while others display unstable or increasing patterns. This behavior can be interpreted as a phase-transition-like shift from a high-entropy exploratory regime to a low-entropy structured reasoning regime. Based on these insights, we propose \textbf{EDRM} (Entropy Dynamics-based Reasoning Manifold), a lightweight and training-free routing framework that leverages early decoding entropy to adaptively select inference strategies. EDRM embeds entropy trajectories into a compact and interpretable manifold representation, enabling both zero-shot deployment and fine-grained instance-level adaptation. Across 15 benchmarks and 4 LLMs of varying scales and architectures, EDRM consistently outperforms static baselines. At the dataset level, EDRM achieves \textbf{41--55%} token reduction while improving accuracy with as few as 50 calibration samples. At the instance level, it further improves accuracy by up to \textbf{4.7%} while maintaining \textbf{27--45%} token savings. These results suggest that reasoning should be invoked selectively rather than by default, and demonstrate the effectiveness of entropy-driven decoding control for efficient and adaptive LLM inference.

Large Language Models (LLMs) perform well on many language tasks, but their Theory of Mind (ToM) reasoning is still uneven in complex social settings. Existing benchmarks, including ExploreToM, do not always test the recursive beliefs and information asymmetries that make these settings difficult. This paper presents OSCToM (Observer-Self Conflict Theory of Mind), an approach for modeling nested belief conflicts in LLM-based ToM tasks. The key case is one in which an observer's view of another agent conflicts with the observer's own belief state. Such cases go beyond simple perspective-taking and require recursive, multi-layered reasoning. OSCToM combines reinforcement learning (RL), an extended domain-specific language, and compositional surrogate models to generate observer-self conflicts. In our experiments, OSCToM-8B gives the best overall result among the systems tested. It improves on the reported ExploreToM results on FANToM and remains competitive on Hi-ToM and BigToM. On the information-asymmetric FANToM benchmark, OSCToM reaches 76% accuracy, compared with the 0.2% reported by ExploreToM. The data-synthesis procedure is also 6x more efficient, indicating that targeted training data can help smaller models handle advanced cognitive reasoning. The project code is available at https://github.com/sharminsrishty/osct.

Large language models (LLMs) are increasingly integrated into high-stakes decision-making. Inspired by the theory of \emph{inattentional blindness} in human cognition, we investigate whether LLMs, trained on human-preferred corpora that embed attentional biases, exhibit a similar limitation: \emph{failing to attend to subtle yet important contextual cues under explicit task instructions}. To evaluate this, we introduce the task of \textbf{explicit-implicit reasoning} and present \textbf{MixRea}, a benchmark of 2,246 multiple-choice questions across 9 reasoning types with varying distributions of explicit and implicit information. Evaluation of 21 advanced LLMs shows that even the best-performing reasoning model (Gemini 2.5 Pro) achieves only 42.8% consistency, revealing widespread inattentional blindness. To mitigate this, we propose \textbf{Potential Relation Completion Prompting (PRCP)}, a prompting method that improves reasoning by recovering overlooked causal relations. Further analysis shows that this limitation persists across diverse multi-source reasoning tasks, highlighting the need for more cognitively aligned models.

Chain-of-thought (CoT) is a standard approach for eliciting reasoning capabilities from large language models (LLMs). However, the common CoT paradigm treats thinking as a prerequisite for answering, which can delay access to plausible answers and incur unnecessary token costs even when the model is able to identify an answer before extended thinking, a behavior known as performative reasoning. In this paper, we introduce CopT, a reformulated reasoning pipeline that reverses the usual order of thinking and answering. Instead of thinking before answering, CopT first elicits a draft answer and then invokes subsequent on-policy thinking conditioned on its own draft answer for reflection and correction. To assess whether the draft answer should be trusted, CopT recasts continuous embeddings as inference-time contrastive verifiers. Specifically, it contrasts the model's support for the same generated tokens under discrete-token inputs and continuous-embedding inputs, yielding a sequence-level reverse KL estimator for answer reliability. Our analysis shows that under certain assumptions, the expected estimate equals the mutual information between the unresolved latent state and the emitted answer token, explaining why it captures answer-relevant uncertainty rather than arbitrary uncertainty in the latent state. When the answer is deemed insufficiently reliable, CopT performs further on-policy thinking, where a second KL estimator dynamically controls draft-answer visibility, preserving useful partial information while reducing the risk of being misled by unreliable content. Across mathematics, coding, and agentic reasoning tasks, CopT improves peak accuracy by up to 23% and reduces token usage by up to 57% at comparable or higher accuracy, without any additional training. The code is available at https://github.com/sdc17/CopT.

While empirical scaling laws for LLM reasoning are well-documented, the theoretical mechanisms governing out-of-distribution (OOD) generalization remain elusive. We formalize reasoning via optimal transport, projecting discrete trajectories into a continuous metric space to quantify domain shifts using the Wasserstein-1 distance. Invoking Kantorovich duality, we bound OOD generalization via architectural Lipschitz continuity and functional approximation limits. This exposes two primary constraints. First, position-dependent attention (e.g., Absolute Positional Encoding) fails to preserve shift invariance, yielding an $Ω(1)$ Lipschitz constant and expected risk, whereas shift-invariant mechanisms (e.g., Rotary Embeddings) preserve equivariance and bound the error. Second, by mapping sequential backtracking to a Dyck-$k$ language, we establish a strict circuit depth lower bound for $\text{TC}^0$ Transformers. Scaling physical layer depth is necessary to avert representation collapse -- a constraint that scaling representation width cannot bypass due to irreducible approximation bounds in Barron spaces. Evaluations across 54 Transformer configurations on combinatorial search corroborate these bounds, demonstrating that generalization risk degrades monotonically with the Wasserstein domain shift.

Tool-augmented reasoning has emerged as a promising direction for enhancing the reasoning capabilities of multimodal large language models (MLLMs). However, existing studies mainly focus on enabling models to perform tool invocation, while neglecting the necessity of invoking tools. We argue that tool usage is not always beneficial, as redundant or inappropriate invocations largely increase reasoning overhead and even mislead model predictions. To address this issue, we introduce AutoTool, a model that adaptively decides whether to invoke tools according to the characteristics of each query. Within a reinforcement learning framework, we design an explicit dual-mode reasoning strategy with mode-specific reward functions to guide the model toward producing accurate responses. Moreover, to prevent premature bias toward a single reasoning mode, AutoTool jointly explores and balances tool-assisted and text-centric reasoning throughout training, and promotes free exploration in later stages. Extensive experiments demonstrate that AutoTool exhibits outstanding performance and high efficiency, yielding a 21.8% accuracy gain on V* benchmark compared to the base model, and a 44.9% improvement in efficiency over existing tool-augmented methods on POPE benchmark. Code is available at https://github.com/MQinghe/AutoTool.

Mathematical reasoning is essential for problem-solving in education, science, and industry, serving as a crucial benchmark for evaluating artificial intelligence systems. As Large Language Models (LLMs) improve their reasoning capabilities, understanding how well they perform mathematical reasoning has become increasingly important. This survey synthesizes recent advancements in mathematical reasoning with LLMs through a structured analysis of datasets, architectures, training strategies, and evaluation protocols. Our systematic review encompasses approximately 120 peer-reviewed studies and preprints, examining the evolution of this research area and providing a unified analytical framework to understand current progress and limitations. Our study particularly introduces a unified taxonomy of mathematical datasets, distinguishing between pretraining corpora, supervised fine-tuning resources, and evaluation benchmarks across varying levels of reasoning complexity. A systematic analysis of reasoning architectures and training strategies, including tool integration, verifier-guided reasoning, and parameter-efficient adaptation, is presented to assess their effects on reasoning robustness and generalization. Moreover, a comparative evaluation of existing metrics highlights the gap between final-answer accuracy and process-level reasoning verification. By synthesizing insights across these areas, our analysis identifies recurring failure modes, such as reasoning faithfulness issues, benchmark biases, and generalization limitations, and outlines key research directions toward improving symbolic grounding, evaluation reliability, and the development of more robust and trustworthy LLM-based reasoning systems.

Large language models (LLMs) have achieved remarkable success in complex reasoning tasks via long chain-of-thought (CoT), yet their immense computational overhead hinders real-world deployment. LLM reasoning distillation addresses this by transferring reasoning capabilities from formidable teacher models to compact student models. However, existing distillation paradigms face a fundamental dilemma. Typical off-policy distillation strictly utilizes teacher-generated golden trajectories, suffering from an exposure bias due to the mismatch between training distributions and student-generated inference contexts, which leads to error cascades in long CoT reasoning. To address this, on-policy distillation allows students to explore their own trajectories, but we demonstrate that it inherently introduces a reciprocal reversed exposure bias: the teacher model also struggles to provide positive guidance when conditioned on student-generated sub-optimal contexts. To resolve this dual exposure biases problem, we propose Monitoring Trajectories and Backtracking when it strays (MOTAB), a new LLM reasoning distillation pipeline. Specifically, MOTAB dynamically monitors the student's on-policy generation against an adaptive safety boundary. When the generation strays and exceeds this threshold, MOTAB backtracks to the last safe state and leverages teacher intervention to correct the course. This approach inherently tolerates minor student errors to mitigate exposure bias, while preventing sub-optimal contexts to circumvent reversed exposure bias. Extensive experiments on the LIMO-v2 and AceReason datasets demonstrate that MOTAB effectively alleviates the dual exposure biases, yielding a roughly 3% average performance improvement in reasoning tasks.

Entropy-based deep reasoning has emerged as a promising direction for improving the reasoning capabilities of Large Language Models (LLMs), but existing methods often either increase response length indiscriminately or shorten responses at the cost of accuracy. To better balance this trade-off, we introduce Conditional Entropy Shaping (CES), a framework that dynamically controls token-level response entropy, enabling LLMs to produce concise solutions on simple problems while encouraging deeper exploration on hard ones. Built on DAPO, CES uses token-level entropy as an uncertainty signal and applies a conditional bidirectional policy: it penalizes high-entropy "forking point" tokens on correct reasoning paths to improve conciseness, and rewards them on incorrect paths to encourage exploration and error correction. We implement CES on DeepSeek-R1-Distill-7B and evaluate it on 12 mathematical benchmarks. CES consistently improves average accuracy while reducing response length relative to DAPO, and supplementary experiments show similar trends on a smaller 1.5B backbone and on out-of-domain benchmarks.

Large Language Models have achieved strong performance on reasoning tasks with objective answers by generating step-by-step solutions, but diagnosing where a multi-step reasoning trace might fail remains difficult. Confidence estimation offers a diagnostic signal, yet existing methods are restricted to final answers or require internal model access. In this paper, we introduce Stepwise Confidence Attribution (SCA), a framework for closed-source LLMs that assigns step-level confidence based only on generated reasoning traces. SCA applies the Information Bottleneck principle: steps aligning with consensus structures across correct solutions receive high confidence, while deviations are flagged as potentially erroneous. We propose two complementary methods: (1) NIBS, a non-parametric IB approach measuring consistency without graph structures, and (2) GIBS, a graph-based IB model that learns subgraphs through a differentiable mask to capture logical variability. Extensive experiments on mathematical reasoning and multi-hop question answering show that SCA reliably identifies low-confidence steps strongly correlated with reasoning errors. Moreover, using step-level confidence to guide self-correction improves the correction success rate by up to 13.5% over answer-level feedback.

Large language models (LLMs) have increasingly leveraged tool invocation to enhance their reasoning capabilities. However, existing approaches typically tightly couple tool invocation with immediate execution. Such immediate tool interaction may disrupt the reasoning coherence of LLMs and constrain their expressivity, ultimately degrading reasoning performance. To this end, for the first time, we propose and formalize the problem of decoupling tool invocation from execution during reasoning, and introduce delayed execution with explicit control to enhance tool-integrated reasoning (TIR). Furthermore, we propose a hierarchical control framework and theoretically derive a surrogate loss that enables an implicitly hierarchical policy to learn behavior equivalent to that of an explicit hierarchical policy, leading to the proposed IH-GRPO algorithm. Extensive experiments on IH-GRPO achieve absolute improvements of 1.87%, 2.16%, and 2.53% on Qwen3-1.7B, Qwen3-4B, and Qwen3-8B across six out-of-domain mathematical reasoning benchmarks over the strongest baseline method, while also yielding consistent performance gains in other domains. Our code is available at https://github.com/Lumina04/IH-GRPO-01.

Reinforcement learning with verifiable rewards (RLVR) has demonstrated promising potential to enhance the reasoning capabilities of large language models (LLMs) in domains such as mathematics and coding. However, its applications on knowledge-intensive domains have not been effectively explored due to the scarcity of high-quality verifiable data. Furthermore, current RLVR focuses solely on the correctness of final answers, leading to the limitations of flawed reasoning and sparse reward signals. In this work, we propose Knowledge-to-Verification (K2V), a framework that extends RLVR to knowledge-intensive domains through automated verifiable data synthesis, while enabling verification of the LLM's reasoning process. Extensive experiments demonstrate that K2V enhances the reasoning of LLM in knowledge-intensive domains without significantly compromising the model's general capabilities. This study also suggests that integrating automated data synthesis with reasoning verification is a promising direction to enhance model capabilities in these broader domains. Code is available at https://github.com/SeedScientist/K2V.

Code reasoning refers to the task of predicting the output of a program given its source code and specific inputs. It can measure the reasoning capability of large language models (LLMs) and also benefit downstream tasks such as code generation and mathematical reasoning. Existing work has verified the effectiveness of reinforcement learning on the task. However, these methods design rewards solely based on final outputs or coarse-grained signals, and neglect the inherent consistency of the stepwise reasoning process in the task. Therefore, these methods often result in sparse reward or reward hacking, which limits the full play of enhanced learning capabilities. To alleviate these issues, we propose CodeThinker, a consistency-driven reinforcement learning framework for code reasoning. Specifically, CodeThinker has three key components: (1) a stepwise reasoning-aware model training module, which utilizes a consistency tracing paradigm as a template to synthesize training data that captures the stepwise reasoning process; (2) a dynamic beam sampling strategy, which aims to improve the quality of sampled outputs under a fixed sampling budget; and (3) a consistency reward mechanism that can effectively alleviate reward hacking. Experiments on three popular benchmarks show that CodeThinker achieves state-of-the-art performance across multiple LLMs. For instance, it outperforms the strongest baseline by 4.3% in accuracy when deployed on Qwen2.5-Coder-7B-Instruct. We also validate the effectiveness of CodeThinker on downstream tasks. Results show that, without additional training, CodeThinker obtains average accuracy gains of 5.33 and 3.11 percentage points on mathematical reasoning and code reasoning tasks covering 17 programming languages, respectively.

Existing deception datasets label completed outputs as honest or deceptive, treating deception as a property of the final response rather than a function of the model's reasoning trace. This obscures a more fundamental question: when does a language model become committed to deception? We introduce counterfactual localization: for each sentence prefix in a reasoning trace, we fix the prefix, resample continuations, and estimate the probability of a deceptive outcome. To scale this, we construct five environments (spanning strategic bluffing, maze guidance, financial advice, used-car sales, and offer negotiation) in which deception is never prompted but emerges from strategic incentives and labels follow mechanically from environment state rather than subjective human judgment. The resulting corpus localizes $\sim$1.46M sentences across four reasoning models, drawn from over 94.1M sampled continuations, 91.5B generated tokens, and over 100K scenarios. Sentence-level human evaluation confirms that detected commitment points correspond to interpretable shifts in decision state. Using this resource, we show that lexical cues for commitment prediction transfer poorly across environments, whereas attention-based transition features generalize out of distribution, suggesting that deceptive commitment is reflected in reusable changes in reasoning dynamics rather than surface form. We further identify compact attention-head sets (under 10% of heads) that, selected on one environment, causally suppress deceptive commitment across held-out environments. We release the corpus as a substrate for studying deception, and more broadly commitment, in language-model reasoning.

With the continuous advancement of reasoning abilities in Large Language Models (LLMs), their application to scientific reasoning tasks has gained significant research attention. Current research primarily emphasizes boosting LLMs' performance on scientific QA benchmarks by training on larger, more comprehensive datasets with extended reasoning chains. However, these approaches neglect the essence of the scientific reasoning process -- logicality, which is the rational foundation to ensure the validity of reasoning steps leading to reliable conclusions. In this work, we make the first systematic investigation into the internal logicality underlying LLM scientific reasoning, and develop a scientific logicality-enriched methodology, including a set of assessment criteria and data sampling methods for logicality-guided training, to improve the logical faithfulness as well as task performance. Further, we take physics, characterized by its diverse logical structures and formalisms, as an exemplar discipline to practise the above methodology. For data construction, we extract scientific problems from academic literature and sample a high-quality dataset exhibiting strong logicality. Experiments based on three different backbone LLMs reveal that: 1) the training data we constructed can effectively improve the scientific logicality in LLM reasoning; and 2) the enriched scientific logicality plays a critical role in solving scientific problems. Code is available at \href{https://github.com/ScienceOne-AI/PhysLogic}{https://github.com/ScienceOne-AI/PhysLogic}.

Recent advances in large language models (LLMs) have shown that Chain-of-Thought (CoT) reasoning can substantially improve performance on complex reasoning tasks. At the same time, In-Context Learning (ICL) has become an important mechanism for adapting LLMs to new tasks without updating model parameters, using only examples provided in the prompt. However, standard ICL often struggles on tasks that require multi-step reasoning, because the demonstrations usually contain only input-output pairs and lack explicit intermediate reasoning steps. This paper introduces an Automatic Chain-of-Thought (Auto-CoT) framework to improve ICL by automatically constructing reasoning-enhanced demonstrations. Auto-CoT generates reasoning chains for input-output examples, augments the prompt context with structured intermediate explanations, and removes irrelevant or low-quality demonstrations through a systematic selection process. By incorporating high-quality reasoning examples into the ICL prompt, Auto-CoT guides the model toward more reliable reasoning and improves prediction accuracy. Experiments across multiple reasoning tasks demonstrate that the proposed framework improves ICL performance by providing explicit intermediate reasoning guidance.

Large language models (LLMs) generate not only reasoning text, but also token-level confidence trajectories that record how uncertainty evolves during inference. Whether these trajectories are relevant to reasoning correctness remains unclear. Here we show that confidence trajectories encode a content-agnostic confidence geometry associated with trace-level final-answer correctness. Using only token-level confidence values, without access to the input question, reasoning text, hidden states, or external verifiers, we find that low-dimensional representations of confidence trajectories separate correct from incorrect reasoning traces. Across GSM8K, MATH, and MMLU, this geometric separation is quantitatively linked to downstream predictability: stronger clustering of correct and incorrect traces, measured by the Davies--Bouldin index, consistently corresponds to higher correctness-discrimination AUC. We further show that correctness-related information is enriched in the tail of reasoning, suggesting that late-stage confidence dynamics carry key correctness signals. We propose NeuralConf, a lightweight estimator that learns from confidence trajectories for correctness evaluation. Under a fixed trace budget, NeuralConf-derived scores improve confidence-weighted answer aggregation over majority voting, tail confidence, and other static baselines. These results reveal that LLMs expose trace-intrinsic statistical signals of correctness through their own confidence dynamics, offering a route to improve inference using information already present within generation.

Reinforcement Learning with Verifiable Reward (RLVR) has proven effective in improving Large Language Model's (LLM) reasoning ability. However, the learning dynamics of RLVR remain underexplored. In this paper, we reveal a counterintuitive phenomenon: among hard examples that the model initially struggles with, a substantial subset remains unlearnable even when correct rollouts are present. To understand the phenomenon, we first demonstrate that existing optimization and sampling techniques fail to resolve unlearnability. With cross-example gradient analysis, we show that unlearnable examples have fundamental representation issue, characterized by low gradient similarity with the rest of the examples and ungeneralizable reasoning patterns. We further show that representation flaws are difficult to mitigate in RL, as data augmentation does not improve gradient similarity. Our study provides the first systematic characterization of unlearnable data in RLVR training and reveals fundamental limitations in current RL approaches for reasoning tasks. Code and data are available at \url{https://github.com/yulinchen99/unlearnability-rlvr}.

When Large Language Models produce structured outputs such as travel plans, code solutions, or multi-step proofs, individual reasoning steps may appear correct while the output as a whole violates budgets, fails test cases, or contradicts earlier deductions. We propose a decomposed energy function that combines a learned quality scorer with deterministic analytical constraint penalties for verifying structured LLM outputs. The quality scorer is a heterogeneous ensemble of low-rank adapters on a single frozen encoder (3% trainable parameters); the ensemble mean ranks candidates while the standard deviation quantifies epistemic uncertainty, driving a two-pass inference loop that triggers targeted regeneration or abstention. Across five benchmarks (GSM8K, MuSR, TravelPlanner, TACO, Knights & Knaves), our 149M-parameter verifier orchestrating a pool of 7-26B open generators outperforms single-shot Qwen-72B on every benchmark, matches Claude Sonnet 4.6 on MuSR (67.7% vs. 68.0%), and reduces constraint violations by 53% relative to Opus 4.6 on TravelPlanner (oracle 0.028, random 0.231). The two routes are complementary: structural verification wins when constraints are checkable (the verifier captures signal frontier models cannot self-detect), while pretraining-scale priors win where they are not (narrative inference, code semantics). A cross-dataset confounding analysis confirms genuine quality discrimination on four reasoning tasks and identifies a model-identity shortcut on code, mitigated via last-layer retraining. Scorers trained on difficult data transfer zero-shot: a MuSR-trained scorer achieves 93.9% on GSM8K without seeing a math problem.

Large Language Models (LLMs) have demonstrated strong capabilities across diverse NLP applications, such as translation, text generation, and question answering. Nevertheless, they remain limited in complex settings that demand deep reasoning and logical inference. Since these models are trained on large-scale text corpora, their generation process may still introduce irrelevant, noisy, or factually inconsistent content. To mitigate this problem, we introduce SGR, a stepwise framework that enhances LLM reasoning through external subgraph generation. SGR builds query-specific subgraphs from external knowledge bases and uses their semantic structure to support multi-step inference. By grounding intermediate reasoning steps in structured external knowledge, the framework helps the model concentrate on relevant entities, relations, and supporting evidence. In particular, SGR first constructs a subgraph tailored to the input question. It then guides the model to reason progressively over the generated structure and combines multiple reasoning trajectories to obtain the final prediction. Experimental results across several benchmark datasets show that SGR achieves consistent improvements over competitive baselines, highlighting its value for improving both reasoning accuracy and factual reliability.

Recent studies observe that reinforcement learning with verifiable rewards (RLVR) reliably improves pass@1 on reasoning tasks, yet often fails to yield comparable gains in pass@k, raising the question of whether RLVR genuinely enables large language models to acquire novel reasoning abilities or merely enhances the efficiency of sampling reasoning modes already present in the base model. Prior analyses largely support the latter view, attributing this limitation to structural properties of standard RLVR objectives that result in insufficient exploration pressure. In this work, we argue that a central structural constraint arises from reverse-KL regularization, which stabilizes training but inherently anchors the policy to the reference distribution, thereby suppressing the emergence of alternative reasoning modes. However, we show that neither removing the KL term nor replacing it with forward-KL provides a satisfactory solution, as both disrupt the efficiency-coverage trade-off by either inducing reward hacking or allocating probability mass to off-target regions. To resolve this tension, we propose SAGE, a principled framework that enables controllable empirical support expansion by reshaping the reverse-KL anchor distribution itself through a guide function q(x,y), achieving consistent improvements in both pass@1 and pass@k across challenging mathematical reasoning benchmarks. Our code is available at https://github.com/tally0818/SAGE.

Test-time compute scaling is a primary axis for improving LLM reasoning. Existing methods primarily scale depth by extending a single reasoning trace. Scaling breadth by sampling multiple candidates in parallel is straightforward, but introduces a selection bottleneck: choosing the best candidate without a ground-truth verifier, since pointwise LLM judging is noisy and biased. To address this, we introduce OpenDeepThink, a population-based test-time compute framework that selects via pairwise Bradley-Terry comparison. Each generation, the LLM judges random pairs of candidates and aggregates votes via Bradley-Terry into a global ranking; top-ranked candidates are preserved and the top three quarters are mutated using the natural-language critiques produced during comparison; the bottom quarter is discarded. OpenDeepThink raises Gemini 3.1 Pro's effective Codeforces Elo by +405 points in eight sequential LLM-call rounds (~27 minutes wall-clock). The pipeline transfers across weaker and stronger models without retuning, and on the multi-domain HLE benchmark, gains appear concentrated in objectively verifiable domains and reverse in subjective ones. We release CF-73, a curated set of 73 expert-rated Codeforces problems with International Grandmaster annotation and 99% local-evaluation agreement against the official verdict.

Large Language Models (LLMs) have demonstrated remarkable abilities in reasoning. However, maximizing their potential through inference-time scaling faces challenges in trade-off between sampling budget and reasoning quality. Current strategies remain inefficient as they typically treat sampling width and depth as orthogonal objectives, where width consensus methods risk reinforcing hallucinations, while depth pruning mechanisms prematurely truncate complex yet valid reasoning chains. Therefore, we propose Dual-Dimensional Consistency (DDC), a unified framework that bridges path quality with adaptive termination. By coupling Confidence-Weighted Bayesian protocol with a Trend-Aware Stratified Pruning, our method ensures that computational resources are concentrated on high quality reasoning paths, filtering hallucinations while accelerating consensus. Evaluations across five benchmarks demonstrate that this approach reduces token consumption by over 10 times while maintaining or exceeding the accuracy of strong baselines across various LLMs.

Tool use extends large language models beyond parametric knowledge, but reliable execution requires balancing appropriate reasoning depth with strict structural validity. We approach this problem from a case-based perspective to present CAST, a case-driven framework that treats historical execution trajectories as structured cases. Instead of reusing raw exemplar outputs, CAST extracts case-derived signals to identify complexity profiles for estimating optimal reasoning strategies, alongside failure profiles to map likely structural breakdowns. The framework translates this knowledge into a fine-grained reward design and adaptive reasoning, enabling the model to autonomously internalize case-based strategies during reinforcement learning. Experiments on BFCLv2 and ToolBench demonstrate that CAST improves both schema-faithful execution and task-level tool-use success while reducing unnecessary deliberation. The approach achieves up to 5.85 percentage points gain in overall execution accuracy and reduces average reasoning length by 26%, significantly mitigating high-impact structural errors. Ultimately, this demonstrates how historical execution cases can provide reusable adaptation knowledge for calibrated tool use.

Multi-step LLM reasoning over structured tables fails because planning and execution share no explicit cell-grounding contract. Existing methods constrain the planner to a left-to-right factorization at odds with table permutation invariance, and score intermediate states by generated content alone, overlooking cell grounding. We conduct a pilot study showing that diffusion language models (DLMs) produce more human-aligned and permutation-stable cell attention on tables than autoregressive models, with a 40.2% median reduction in attention-AUROC variability under row reordering. Motivated by this, we propose TABALIGN, a planned table reasoning framework that operationalizes the contract. TABALIGN pairs a masked DLM planner, whose bidirectional denoising emits plan steps as binary cell masks, with TABATTN, a lightweight verifier trained on 1,600 human-verified attention standards to score each step by its attention overlap with the plan-designated mask. Across eight benchmarks covering table question answering and fact verification, TABALIGN improves average accuracy by 15.76 percentage points over the strongest open-source baseline at comparable 8B-class scale, with a matched-backbone ablation attributing 2.87 percentage points of this gain to the DLM planner over an AR planner on a fixed reasoner. Cleaner DLM plans also accelerate downstream reasoning execution by 44.64%.

We present Darwin Family, a framework for training-free evolutionary merging of large language models via gradient-free weight-space recombination. We ask whether frontier-level reasoning performance can be improved without additional training, by reorganizing latent capabilities already encoded in existing checkpoints. Darwin introduces three key ideas: (i) a 14-dimensional adaptive merge genome enabling fine-grained component- and block-level recombination; (ii) MRI-Trust Fusion, which adaptively balances diagnostic layer-importance signals with evolutionary search through a learnable trust parameter; and (iii) an Architecture Mapper that enables cross-architecture breeding between heterogeneous model families. Empirically, the flagship Darwin-27B-Opus achieves 86.9% on GPQA Diamond, ranking #6 among 1,252 evaluated models, and outperforming its fully trained foundation model without any gradient-based training. Across scales from 4B to 35B parameters, Darwin models consistently improve over their parents, support recursive multi-generation evolution, and enable a training-free evolutionary merge that combines Transformer- and Mamba-based components. Together, the Darwin Family demonstrates that diagnostic-guided evolutionary merging is a practical and reproducible alternative to costly post-training pipelines for reasoning-centric language models.

Arguments are a fundamental aspect of human reasoning, in which claims are supported, challenged, and weighed against one another. We present an end-to-end large language model (LLM)-based system for reconstructing arguments from natural language text into abstract argument graphs. The system follows a multi-stage pipeline that progressively identifies argumentative components, selects relevant elements, and uncovers their logical relations. These elements are represented as directed acyclic graphs consisting of two component types (premises and conclusions) and three relation types (support, attack, and undercut). We conduct two complementary experiments to evaluate the system. First, we perform a manual evaluation on arguments drawn from an argumentation theory textbook to assess the system's ability to recover argumentative structure. Second, we conduct a quantitative evaluation on benchmark datasets, allowing comparison with prior work by mapping our outputs to established annotation schemes. Results show that the system can adequately recover argumentative structures and, when adapted to different annotation schemes, achieve reasonable performance across benchmark datasets. These findings highlight the potential of LLM-based pipelines for scalable argument mining.

While many-shot ICL achieves remarkable performance, prior studies of its scaling behavior have mainly focused on non-reasoning tasks. In this work, we study many-shot ICL on reasoning tasks, with a particular focus on many-shot chain-of-thought in-context learning (CoT-ICL). Analyzing across non-reasoning and reasoning tasks and across non-reasoning and reasoning-oriented LLMs, we identify several distinctive properties of many-shot CoT-ICL. We further interpret these findings by viewing many-shot CoT-ICL as in-context test-time learning rather than scaled pattern matching, and suggest two principles: (i) demonstrations should be easy for the target model to understand, and (ii) they should be ordered to support a smooth conceptual progression. Guided by the principle, we propose Curvilinear Demonstration Selection (CDS), a simple ordering method that yields up to a 5.42 percentage-point gain on a math task with 64 demonstrations. Overall, our results reframe the long context window from a retrieval buffer into a structured curriculum for in-context test-time learning.

Recent advances in Reinforcement Learning (RL) have underscored its potential for incentivizing reasoning capabilities of Large Language Models (LLMs). However, existing step-level efforts suffer from costly annotations that limit domain coverage, while scalar scores further impose an information bottleneck, offering insufficient semantic bandwidth to improve intermediate decisions. Alternative language-critique approaches, which rely on frozen or external critics, provide richer textual feedback but lack the scalability needed for sustained policy improvement. In this work, we propose language-driven stepwise trajectory redirection, termed as STRIDE, a novel training framework that shifts process supervision from scalar rewards to learnable stepwise language feedback. Specifically, we co-train a generator and a generative verifier using only outcome-based rewards, eliminating external annotations, while delivering sustained policy improvement through jointly aligned verifier training. The verifier's stepwise language critiques explicitly localize and explain failures, enabling the generator to redirect reasoning trajectories at intermediate steps toward alternative decisions. The trajectory redirection design guarantees harmless policy improvement, even under noisy or suboptimal verifier feedback. Experiments on diverse reasoning benchmarks show that STRIDE significantly outperforms state-of-the-art baselines, as well as achieving breakthroughs on zero-pass-rate problems where scalar methods yield no learning signal in our ablation studies, demonstrating the effectiveness of learnable stepwise language feedback for enhancing LLM reasoning.

Large Reasoning Models (LRMs) are increasingly integrated into systems requiring reliable multi-step inference, yet this growing dependence exposes new vulnerabilities related to computational availability. In particular, LRMs exhibit a tendency to "overthink", producing excessively long and redundant reasoning traces, when confronted with incomplete or logically inconsistent inputs. This behavior significantly increases inference latency and energy consumption, forming a potential vector for denial-of-service (DoS) style resource exhaustion. In this work, we investigate this attack surface and propose an automated black-box framework that induces overthinking in LRMs by systematically perturbing the logical structure of input problems. Our method employs a hierarchical genetic algorithm (HGA) operating on structured problem decompositions, and optimizes a composite fitness function designed to maximize both response length and reflective overthinking markers. Across four state-of-the-art reasoning models, the proposed method substantially amplifies output length, achieving up to a 26.1x increase on the MATH benchmark and consistently outperforming benign and manually crafted missing-premise baselines. We further demonstrate strong transferability, showing that adversarial inputs evolved using a small proxy model retain high effectiveness against large commercial LRMs. These findings highlight overthinking as a shared and exploitable vulnerability in modern reasoning systems, underscoring the need for more robust defenses.

On-policy self-distillation trains a reasoning model on its own rollouts while a teacher, often the same model conditioned on privileged context, provides dense token-level supervision. Existing objectives typically weight the teacher's token-level signal uniformly across a chain-of-thought sequence, despite substantial variation in the entropy of the teacher's predictive distribution. We propose EGRSD (Entropy-Guided Reinforced Self-Distillation), which unifies token-level updates through three signals: a reward-grounded direction, a teacher-student likelihood-ratio magnitude, and the proposed teacher-entropy confidence gate that down-weights high-entropy token positions while maintaining a nonzero lower bound on every token weight. We further introduce CL-EGRSD, a causal-lookahead variant that distinguishes sustained high-entropy spans from transient high-entropy positions whose following context rapidly becomes low entropy. Experiments with Qwen3-4B and Qwen3-8B in thinking mode show that EGRSD and CL-EGRSD advance the accuracy-length frontier among the compared trainable methods.

Multi-agent debate has emerged as a promising approach for improving LLM reasoning on ground-truth tasks, yet current methodologies face certain structural limitations: debate tends to induce a martingale over belief trajectories, majority voting accounts for most observed gains, and LLMs exhibit confidence escalation rather than calibration across rounds. We argue that the genuine value of debate, and dialectic systems as a whole, lies not in ground-truth tasks but in defeasible domains, where every position can in principle be defeated by better reasoning. We present the Council of Hierarchical Agentic Language (CHAL), a multi-agent dialectic framework that treats defeasible argumentation as an engine for belief optimization. Each agent maintains a CHAL Belief Schema (CBS), a graph-structured belief representation with a Bayesian-inspired architecture, that facilitates belief revision through a gradient-informed dynamic mechanism by leveraging the strength of the belief's thesis as a differentiable objective. Meta-cognitive value systems spanning epistemology, logic, and ethics are elevated to configurable hyperparameters governing agent reasoning and adjudication outcomes. We provide a series of ablation experiments that demonstrate systematic and interpretable effects: the adjudicator's value system determines the debate's overall trajectories in latent belief space, council diversity refines beliefs for all participants, and the framework generalizes across broad fields. CHAL is, to our knowledge, the first framework to treat multi-agent debate as structured belief optimization over defeasible domains. Further, the auditable belief artifacts it produces establish the foundation for dedicated evaluation suites for defeasible argumentation, with broader implications for building AI systems whose reasoning and value commitments are transparent, aligned, and subject to human oversight.

We study on-policy self-distillation (OPSD), where a language model improves its reasoning ability by distilling privileged teacher distributions along its own on-policy trajectories. Despite its promise, OPSD can suffer from training instability due to a pattern mismatch between teacher and student responses. Self-reflected teacher responses may introduce reflection-induced biases and response templates that miscalibrate token-level supervision, ultimately harming the student's reasoning ability. To mitigate this issue, we propose OGLS-SD, an outcome-guided logit-steering framework that leverages verifiable outcome rewards to calibrate privileged teacher logits. Specifically, OGLS-SD contrasts teacher logits induced by successful and failed on-policy trajectories, constructing an outcome-discriminative steering direction for token-level guidance. Experiments on mathematical reasoning benchmarks show that OGLS-SD stabilizes self-distillation and improves performance over standard OPSD and other variants.

While Large Language Models (LLMs) are widely used in open-domain Question Answering (QA), their ability to handle inferential questions-where answers must be derived rather than directly retrieved-remains still underexplored. This study investigates how the structure and quality of passages influence LLM performance on such questions. We focus on convergence, a measure of how effectively sentences (hints) eliminate incorrect answers, as a criterion for constructing passages. Using subsets of the TriviaHG dataset, we form passages by combining sentences with varying convergence levels and evaluate six LLMs of different sizes and architectures. Our results show that passages built from higher convergence sentences lead to substantially better answer accuracy than those selected by cosine similarity, indicating that convergence captures meaningful relevance for inferential reasoning. Additionally, ordering sentences by descending convergence slightly improves performance, suggesting that LLMs tend to prioritize earlier, information-rich cues. These findings highlight convergence as a practical signal for guiding passage construction and analyzing inferential reasoning behavior in LLMs.

Recent soft prompt research has tried to improve reasoning by inserting trained vectors into LLM inputs, yet whether the gain comes from the learned content or from the act of injection itself has not been carefully separated. We study Random Soft Prompts (RSPs), which drop the training step entirely and append a freshly drawn sequence of random embedding vectors to the input. Each RSP vector is sampled from an isotropic Gaussian fitted to the entrywise mean and variance of the pretrained embedding table; the sequence carries no learned content, and yet reaches accuracy comparable to optimized soft prompts on math reasoning benchmarks in several settings. The mechanism unfolds in two stages: because attention has to absorb a never-seen-before random position, the distribution over the first few generated tokens flattens and reasoning trajectories branch, and as generation continues this influence dilutes naturally so the response commits to a single completion. We show that during inference RSPs lift early-stage token diversity and, combined with temperature sampling, widen Pass@N, the probability that at least one out of N attempts is correct. Beyond inference, we carry the same effect into DAPO training and demonstrate practical gains. Our contributions are: (i) RSP isolates the simplest form of soft prompt -- training-free, freshly resampled -- providing a unified lens for the structural effect of injection that variants otherwise differing in training and form all share; (ii) a theoretical and empirical validation of the underlying mechanism; and (iii) an extension from inference to training.

The reasoning frontier of Large Language Models (LLMs) has advanced significantly through modern post-training paradigms (e.g., Reinforcement Learning from Verifiable Rewards (RLVR)). However, the efficacy of these methods remains fundamentally constrained by the diversity and complexity of the training data. One practical solution is data synthesis; yet, prevalent methods relying on unstructured mutation or exploration suffer from homogeneity collapse, failing to systematically expand the reasoning frontier. To overcome this, we propose Evoutionary Task Discovery (EvoTD), a framework that treats data synthesis as a directed search over a dual-axis manifold of Algorithmic Skills and Complexity Attributes. We introduce structured evolutionary operators to navigate this space: a Crossover operator that synthesizes novel skill compositions to enhance diversity, and a Parametric Mutation operator that scales structural constraints (e.g., input size, tree depth) to drive robust generalization. Crucially, we integrate a dynamic Zone of Proximal Development filter, ensuring tasks lie within the learnable region of the model. Empirically, EvoTD delivers substantial reasoning gains that generalize consistently across model architectures, pretraining regimes, and scales, demonstrating that structured evolutionary curricula can effectively support reasoning improvement. We release our code on https://github.com/liqinye/EvoTD.

We present Seirênes, a self-play RL framework that transforms contextual interference from a failure mode of LLM reasoning into an internal training signal for co-evolving more resilient reasoners. While RL with verifiable rewards has significantly advanced reasoning capabilities, models can still exhibit fragility when encountering non-idealized contexts: scenarios characterized by superfluous information, tangential instructions, or incidental correlations that differ from the clean distributions typical of standard benchmarks. Seirênes harnesses this vulnerability through a parameter-shared and adversarial self-play loop. Within this framework, a single model is trained to both construct plausible yet distracting contexts that expose its own reasoning blind spots, and solve problems by discerning the essential task from these perturbations to recover the core underlying logic. By pitting these competing objectives against each other, Seirênes compels the model to move beyond superficial pattern matching and anchors its capabilities in robust underlying reasoning. This continuous interaction sustains an informative co-evolutionary curriculum as the model improves. Across seven mathematical reasoning benchmarks and model scales from 4B to 30B, Seirênes achieves average gains of +10.2, +9.1, and +7.2 points. Besides, distracting contexts produced by the 4B Seirênes model reduce the accuracy of top-tier closed-source models (GPT and Gemini) by roughly 4--5 points, revealing Seirênes' general ability to uncover reasoning models' blind spots.

Recent multimodal large language models (MLLMs) have shown strong chain-of-thought (CoT) reasoning ability on vision-language tasks, but their direct deployment in real-world systems is often limited by latency and resource constraints. In practice, smaller MLLMs are preferred for online serving, yet their reasoning performance is bottlenecked by the lack of large-scale, high-quality multimodal CoT supervision. In this paper, we present OmniThoughtVis, a scalable data curation and distillation pipeline for transferring multimodal reasoning capabilities from high-capacity teacher models to smaller, deployment-oriented MLLMs. Starting from a diverse open-source seed pool, our pipeline generates structured CoT traces and performs joint annotation of reasoning difficulty, answer quality, and semantic task tags. To maintain data quality at scale, we combine rule-based filtering, difficulty-aware selection, and tag-based diversity sampling, resulting in a curated corpus of 1.8M samples that supports controllable subset construction for downstream training. We use OmniThoughtVis to distill Qwen3-VL models from 2B to 8B parameters and evaluate them on nine multimodal reasoning benchmarks. The resulting distilled models show consistent gains across model scales, including improvements of up to +16.8 points on MathVerse and +5.6 points on MMMU-Pro for the 4B model. Notably, the distilled 4B model matches or surpasses the undistilled 8B baseline on several tasks, highlighting the practical value of scalable reasoning distillation for deployment-oriented MLLMs.

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.

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.

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.

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.

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.