Large Language Model Reasoning

Reasoning in large language models (LLMs) is often grounded in human text, human demonstrations, and human-generated rationales. For equilibrium reasoning in complex games, however, relying on human data can be suboptimal. In fact, human play is often guided by intuition and heuristics and can deviate substantially from game equilibrium. This discrepancy is amplified in games with mixed-strategy equilibria, where human data is heavily biased toward pure strategies. Consequently, conditioning LLMs on this data yields weak game strategies. To grant LLMs the reasoning capacity in games, in this work, we study how to elicit equilibrium play using solver output. We propose Mixed-Strategy Decision Tree (MDT), which articulates the silent optimality of the equilibrium into sparse strategic rules that both humans and LLMs could understand. Using solver output rather than human annotation allows us to extend the input to arbitrarily new states and continuations. We instantiate this study on No-Limit Texas Hold'em by querying a solver oracle for over \textbf{250 million mixed-strategy decisions}; MDT together with other techniques \textbf{reduces the $\ell_1$ distance to the equilibrium by $52.6\%$} across $8$ different LLM configurations. A Route-only ablation tests the incremental contribution of the shadow-based contrast, while complete River-endgame and Liar's Dice experiments evaluate strategic fidelity and portability beyond the original NLH communication setting.

Vision-language-action (VLA) models provide strong semantic priors for robot navigation, but they often ignore embodiment-specific mobility constraints. A path that is semantically plausible for one robot may be physically infeasible for another. We propose CrossTracer, a hierarchical framework for cross-embodiment navigation through adaptive trace residuals. CrossTracer represents navigation plans as normalized image-plane waypoints, forming a unified pixel-space interface between semantic reasoning and physical grounding. First, Vision-Language Trace Proposer (VL-Tracer) adapts a pretrained VLA model to predict an initial navigation trace from egocentric observations and flexible goal specifications. Second, CE-Adapter refines this trace by predicting embodiment-conditioned residual corrections from visual traversability cues, robot identity, and the initial trace. To train the refinement module without costly manual annotation, Cross-Embodiment RRT* (CE-RRT*) converts panoptic segmentation into robot-conditioned traversability cost maps and generates cost-minimizing pixel-space traces. We evaluate CrossTracer on the NaviTrace benchmark, which tests whether a model can generate embodiment-consistent navigation traces from egocentric observations, language instructions, and robot embodiment types. CrossTracer achieves a total score of 45.68, outperforming the strongest evaluated general-purpose baseline, Gemini-2.5-Pro, by 10.01 points, corresponding to a 28.1% relative improvement. Real-world deployment on wheeled and legged robots further shows improved navigation success and execution efficiency.

Multilingual reasoning transfer is crucial for extending reasoning capabilities of large language models (LLMs) beyond high-resource languages. On-policy self-distillation (OPSD) and its variants have emerged as a promising paradigm, providing dense token-level supervision on student-generated rollouts, yet their objectives do not explicitly prioritize reasoning signals most critical to cross-lingual transfer. We characterize that target-language reasoning comprises the generation of both surface text and reasoning pivots, which are decisions that advance or redirect the reasoning process and shape subsequent inference. This motivates concentrating privileged distillation around such pivots. We therefore propose RP-OPSD, Reasoning-Pivot-guided On-Policy Self-Distillation, using the distributional shift between matched teacher views with and without an English reference solution as an operational proxy to guide privileged distillation and reference anchoring. Experiments on mathematical reasoning benchmarks covering 17 languages and multiple difficulty levels show that our method outperforms strong multilingual reasoning baselines and OPSD variants. Further analysis reveals that RP-OPSD concentrates privileged distillation on reasoning-control and problem-condistioned state-update tokens, while downweighting it for tokens that mainly support surface realization. Our code is available at https://github.com/NJUNLP/RP-OPSD.

On-Policy Distillation (OPD) is emerging as a promising alternative to reinforcement learning for LLM post-training, yet its effectiveness in multilingual settings remains underexplored. We study OPD and its advanced variant, On-Policy Delta Distillation (OPD$^2$), for mathematical reasoning in English, Korean, and Japanese. OPD$^2$ improves OPD by using the probability gap between a post-trained teacher and its base model as the learning signal. Experiments with Qwen3 show that OPD$^2$ consistently outperforms the original OPD, with particularly strong improvements in Korean and Japanese, and generally narrows the English-Korean performance gap. We further find that English-only OPD can also increase performance for Korean and Japanese, but often shifts the responses toward English, highlighting the importance of multilingual data to preserving target-language responses.

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

Test-time scaling improves LLM reasoning by using additional inference compute, but wider sampling alone can suffer from diminishing returns: new rollouts often repeat existing answer patterns instead of adding useful reasoning diversity. Verifier-based selection offers an alternative, but its performance depends on the calibration of an external reward model. We propose a verifier-free breadth--depth refinement framework that uses test-time compute to both explore and improve candidate solutions. The method samples multiple independent reasoning rollouts, refines each rollout through iterative self-critique and self-correction, and aggregates the refined answers by majority voting. Breadth preserves diverse initial attempts, while depth repairs local reasoning errors before aggregation. Across AIME24, AIME25, AMC, OlympiadBench, and MATH500, our method consistently improves over greedy decoding, majority voting, verifier-based best-of-$N$, beam search, and lookahead decoding across multiple open-weight models. For instance, with Qwen2.5-1.5B, accuracy increases from the strongest verifier-based baseline to $58.0\%$ on MATH500, and from $25.0\%$ to $32.5\%$ on AMC. These results show that test-time compute can be more effective when used to refine sampled trajectories rather than only to sample more candidates or rely on verifier-guided selection.

Video object removal must eliminate not only the target object but also its induced effects while maintaining high-fidelity and spatiotemporally coherent restoration. Existing methods mainly learn object-effect correspondences implicitly from predefined effect categories and fixed data distributions, limiting their generalization to complex real-world scenes involving compositional effects, spatially detached or weakly correlated effects, long-tail physical phenomena, and dynamically evolving interactions. We propose EffectLearner, a semantic-reasoning-enhanced framework that combines a VLM-based Object-Effect Reasoner with a DiT-based Video Eraser. Guided by a structured effect-analysis prompt, the Reasoner performs cross-modal reasoning over a target-highlighted video and extracts compact effect-aware context, which guides the Video Eraser toward comprehensive object-effect removal. Motion-aware mask guidance and motion-consistency supervision further improve removal coverage and spatiotemporal stability under object motion and evolving scene dynamics. To fully exploit the framework in challenging real-world scenarios, we further construct EffectWorld, a paired video dataset specifically designed for complex object-induced effects, and introduce a progressive training curriculum that combines common supervision with complex-effect data. On the standard ROSE-Bench, EffectLearner outperforms existing baselines on most metrics and achieves clear advantages on both EffectWorld-Eval and the challenging EffectWorld-Wild, demonstrating its ability to deliver high-quality video object removal in complex real-world scenes.

Evolution Strategy (ES) is a promising alternative to gradient-based fine-tuning for resource-constrained Large Language Model (LLM) reasoning. However, directly applying ES to billion-parameter LLMs is highly ineffective. In such high-dimensional parameter spaces, most random perturbations are nearly orthogonal to useful update directions, leading to unstable optimization. We propose Hyper-ES, a subspace-based ES framework that avoids the weakness of ES in full-parameter search while exploiting its strength in low-dimensional optimization. Instead of asking ES to discover useful directions from random perturbations in the LLM parameter space, Hyper-ES first performs a small number of inexpensive gradient-based fine-tuning runs to obtain descent directions. Although each direction may provide only a limited improvement on its own, their span forms a compact adaptation subspace that captures useful reasoning updates. Hyper-ES then applies CMA-ES to optimize layer-wise DARE-TIES merging coefficients within this subspace, allowing ES to search over combinations of meaningful descent directions rather than over arbitrary full-model perturbations. We evaluate Hyper-ES on three Qwen2.5-Instruct and DeepSeek-R1-Distill backbones across six mathematical reasoning datasets. Results show that Hyper-ES consistently outperforms GRPO-LoRA by 1% while requiring 10% fewer space-consuming gradient updates. Code at https://github.com/kuangrepi/Hyper-ES.

Procedural generators produce useful verifiable reasoning problems at scale, but have received less attention as data for completion-supervised fine-tuning. We introduce Reasoning Core, a collection of 50 generators spanning mathematics, logic, planning, state tracking, formal languages, structured data, games, causality, and code, with semantic scorers, difficulty controls, and task evaluators. Under a matched completion-supervised protocol, we compare Reasoning Core with Procedural Warmup, Reasoning Gym, and SynLogic across four base-model settings and multiple training durations. In the primary 3B comparison, Reasoning Core achieves the highest mean scores on DROP, LogiQA, and ARC-Challenge, exceeding both the baseline without procedural data and all three alternative procedural collections. Task-level analyses show that semantic validity alone does not ensure training utility, highlighting compact targets and calibrated difficulty as important design factors. We ran audits combining model-assisted review, human adjudication, and regression testing. Applied throughout Reasoning Core development and to the other collections, they reveal subtle mismatches among generation, rendering, targets, and scoring, a reminder that procedural generation alone does not guarantee correctness. The library, generated datasets, and audit material are publicly available.

Long context reasoning in large language models (LLMs) is usually constrained by the fact that a single inference trajectory has to simultaneously explore the context, store intermediate state, verify evidence, and produce the final answer. This becomes particularly difficult in tasks that require extraction, counting, ordering, or multi-hop reasoning, where an early mistake can propagate until the final response. In this work, we propose Chained Recursive Language Models (Chained RLM), an inference-time architecture, in which the same underlying model is called repeatedly as a sequence of fresh reasoning roots. Each root receives the original problem and context, but does not inherit the full conversational history. Instead, it receives a compact plain-text summary, a plain-text blackboard, and some durable task-specific artifacts written by predecessor roots. The motivation is to manage the context by chopping into partial tasks rather than one large inference response; in each staged computation, intermediate artifacts can be inspected, corrected, and extended by a later fresh inference by the same model. We describe the system model, handoff mechanism, artifact workspace, and evaluation protocol for this system. We study when fresh-context artifact continuation gives a measurable gain in accuracy over direct LLM answering even with recursive tool-calling.

Chain-of-thought (CoT) reasoning offers a window into the decision-making of large language models (LLMs), which can be monitored for target behaviors by reading the reasoning trace, motivating work on CoT monitorability. Latent CoT approaches, however, replace the explicit tokens with a small number of continuous states, lowering inference costs but removing the readable trace this monitoring relies on. Monitoring then requires alternative access to the model, such as probing its activations or verbalizing the latent states back into text, but how much monitorability these alternatives preserve is unclear. We study this question with a hint-based intervention setup, a proxy for behaviors where models exploit biasing input cues, e.g., an inadvertently leaked answer or a belief stated by the user, without acknowledging them. Taking hint-reliance as the monitorability target, we compare monitors across reasoning modes, from explicit CoT to weakly- and strongly-supervised latent CoT, on math reasoning and question answering. We find that, in this setup, monitorability depends more on properties of the task (such as whether the correct answer constrains the supporting reasoning) and the level of access to model internals than on the reasoning mode.

Multi-agent LLM systems relay key--value caches instead of text and credit their gains to exchanged ``latent thoughts''. That credit is a claim about \emph{which} example's cache is relayed, not merely that one is. We audit it causally in released systems. The cache is replaced with deranged (mismatched-example), zeroed, and moment-matched random counterparts, under two regimes defined by whether the receiver needs the sender's private information. Where it does, the battery reads ceiling: 100% against 23--25% for answer-irrelevant relays on the primary backbone, a contrast replicated across three families, five checkpoints, and a prose document-QA surface. Where it does not, a pre-registered five-seed protocol establishes equivalence within 2.8 points, a margin anchored to the audited system's reported gain, under Holm-corrected TOST on GSM8K and ARC-Challenge across three Qwen3 scales and on MedQA at 8B (one cell shows a small detected advantage inside the margin); a second family shows no detected advantage. A large cache effect need not be a pairing effect. In one natural cell, zeroing the relay costs 14.7 points; a mismatched cache, 0.4. Nor is need sufficient: under the same test, delivered channels span ceiling (LatentMAS's native relay), partial (KVComm's layer subset), and no detected example-specific transfer (C2C's released projector). Benchmark deltas do not by themselves establish latent-thought transmission; establishing it takes a mismatched-cache audit, which we release.

Research agents increasingly conduct multi-round machine-learning experiments in industrial recommendation settings and retain the resulting trajectories to guide later decisions. Yet a completed trajectory is not automatically evidence: generated artifacts may be unsupported or incomplete, executed rounds may be invalid or confounded, and later modifications may obscure earlier findings. We study \textbf{trajectory-to-evidence conversion}, asking what a completed research process has actually established. We introduce an evidence-grounded framework that couples bounded verification of consequential artifacts with post-execution claim qualification. A context-isolated generate--verify--repair process checks artifacts for evidence violations and missing downstream requirements before release. After execution, validity and attribution checks consolidate evidence across rounds, qualify intervention-level claims as actionable repairs, diagnostic guards, or withheld findings, and preserve admitted claims as auditable records with explicit provenance and applicability boundaries. A hybrid LLM-assisted controller subsequently applies, defers, or rejects records based on available target evidence. Record audits characterize which claims survive qualification, while downstream diagnostics identify affirmative applicability judgment as a bottleneck for the tested controller. Across paper-to-target adaptations, later rounds often improve on the first, while final rounds frequently underperform an earlier best, exposing non-monotonic trajectory evolution. Candidates produced through the complete workflow also yielded positive online lifts relative to deployed baselines.

Large Reasoning Models (LRMs) excel on complex tasks through long chain-of-thought (CoT) reasoning, but their lengthy intermediate steps cause severe overthinking that inflates inference cost. KV-cache compression is a common solution, yet existing reasoning-oriented methods apply a uniform policy across the trajectory and judge compression only by what it removes from the cache. Two observations point the other way. First, a reasoning state's tolerance to context loss varies along the trajectory, and process reward tracks it: deleting tokens at high-reward steps preserves accuracy far better than deleting the same budget at random. Second, compression is not free on the generation side, since a smaller cache leads the model to generate more tokens, partly canceling the saving. Together these motivate coordinating both sides under a single process reward. We propose ReCo (Reward-Coordinated Compression), a step-wise framework in which a lightweight process-reward estimator scores each completed step and drives three components: (1) reward-adaptive KV-cache compression that shrinks the retained cache harder at high-reward steps and less at low-reward ones, (2) a reward-banded penalty on reflection tokens that curbs redundant generation, and (3) confidence-based early stopping that triggers when the reasoning is reliable. Across three reasoning models and six benchmarks, ReCo reduces generated tokens by 37%-65% and end-to-end latency by 2.08x-2.35x over Full CoT, all while largely preserving accuracy.

Large Language Models (LLMs) have transformed computational linguistics and achieved remarkable performance across numerous natural language processing tasks, yet significant gaps persist in understanding how these systems process culturally embedded linguistic expressions. This paper introduces ProverbIT, a novel Italian benchmark comprising 100 multiple-choice questions designed to evaluate LLMs' ability to complete Italian proverbs. We assess 13 frontier models, including Large Reasoning Models (LRMs) and traditional LLMs, across three tasks: proverb completion, multiple-choice selection with correct answers, and multiple-choice selection without correct answers. Our evaluation reveals surprising results: while nearly all models demonstrate knowledge of the proverbs through successful completion tasks, performance drops dramatically when transitioning to multiple-choice formats without correct answers, with even state-of-the-art reasoning models showing substantial degradation. Through detailed Chain-of-Thought analysis of two LRMs, we uncover that models exhibit a strong bias toward selecting literal synonyms and frequently mention correct proverb endings during reasoning without successfully identifying their absence from the given options. These findings suggest that current LLMs rely heavily on memorized patterns rather than deeper semantic understanding of culturally grounded expressions, highlighting important limitations in their reasoning capabilities for figurative language comprehension.

Large language models (LLMs) have recently shown strong performance on Theory of Mind (ToM) tests, prompting debate about the nature and validity of the underlying capabilities. At the same time, reasoning-oriented LLMs trained via reinforcement learning with verifiable rewards have demonstrated notable improvements across a range of benchmarks. In this work, we examine the behavior of such reasoning models in ToM tasks using novel adaptations of machine psychological experiments together with results from established benchmarks. We observe that reasoning models consistently exhibit increased robustness to prompt variations and task perturbations. Our analysis suggests these gains come at least partly from models being more robust at reaching the correct answer under prompt and task variation. We read this as evidence for a robustness-based account rather than for a new ToM-specific ability.

LLM-based multi-agent systems (MAS) typically optimize a single topology, restricting reasoning to a narrow trajectory and limiting comprehensive analytical capacity. Naively merging multiple topologies into a composite graph introduces redundant noise propagation across irrelevant connections, degrading solution quality. To address this dilemma, we propose \textbf{Hierarchical Sparse Coordination over a Union of Complementary Topologies for MAS (HELENA)}, a multi-agent framework that balances diverse reasoning paths with sparse task-dependent execution. \helena{} constructs a union MAS graph from complementary candidate topologies selected via Monte Carlo Tree Search and Determinantal Point Process, broadening the reasoning trajectory for comprehensive analysis of complex problems. A Hierarchical Sparse Coordination module then activates only a sparse subgraph at each step while agents exchange compressed latent briefs to suppress redundant noise propagation. Finally, a Local Self-Refinement stage identifies decision units with discrepancy evidence and rewrites them only when contrastive evidence simultaneously confirms a reliable solution-side failure and a challenger-side improvement. Experiments across eight benchmarks show that \helena{} achieves state-of-the-art results on all benchmarks, with an average gain of \pctup{3.47} over the strongest baseline and up to \pctup{10.34} on MMLU-Pro, achieving larger improvements on harder benchmarks at a reasonable additional cost.

Synthetic generation of Cognitive Behavioral Therapy (CBT) sessions is challenged by two competing demands: adhering to strict therapeutic structure while modeling the resistant, unpredictable behavior of real patients. Existing script-based methods fail to capture dynamic therapeutic interactions, while multi-agent approaches struggle to adhere to CBT's sequential structure; both suffer from sycophancy, producing overly compliant patients that misrepresent real clinical settings. In this work we introduce ODRA, a novel framework for synthesizing therapy dialogues through a Chain-of-Thought (CoT) strategy grounded in foundational CBT guidelines (Beck, 2020). ODRA further incorporates a resistance orchestrator to solve patient sycophancy, which employs steering techniques to elicit behaviors aligned with their resistance level. Automated and expert evaluations show that ODRA significantly outperforms existing methods across therapeutic skills, CBT alignment, and patient behavioral fidelity, with licensed psychologists preferring ODRA sessions across 12 of 13 clinical metrics. Furthermore, models fine-tuned on our dataset demonstrate superior therapeutic performance against both cooperative and resistant patients, validating that explicit resistance modeling in synthetic training data directly translates to downstream clinical robustness.

Existing document understanding benchmarks have largely focused on locating page elements, yet real-world document intelligence requires models to reason jointly about region semantics, spatial relations, and visual structure. We present ADOPD 2026, a reasoning-oriented extension of ADOPD that turns page decomposition into spatially grounded document understanding. ADOPD 2026 enriches page anchors inherited from ADOPD 2024 dataset with human-cleaned captions, semantic tags, and generated chain-of-thought (CoT) traces grounded to document regions. Instead of treating boxes, masks, and tags as independent supervision signals, we cast text blocks, visual entities, semantic labels, bounding boxes, and polygon masks as a shared vocabulary of visual anchors. This representation supports three connected capabilities. First, region-level semantic tagging asks models to identify document element types from both page context and local appearance, revealing long-tail semantic failures that standard layout benchmarks often hide. Second, unified vision-language grounding generates text regions and visual entities together with coordinates or polygonal outlines, transforming detection and segmentation outputs into structured anchors that can be reused by downstream reasoning systems. Third, current state-of-the-art models still struggle with dense counting tasks evaluated on DocCount, a benchmark derived from ADOPD 2026, highlighting the need for the Thinking-with-Anchors pipeline in document semantic understanding. By connecting page decomposition to verifiable visual-anchor reasoning, ADOPD 2026 provides a task framework that moves document understanding beyond localization toward anchor-grounded document intelligence.

We present a structured large-language-model (LLM) architecture for zero-shot human--robot coordination in a cooperative construction task with private goal views. Guided by a Dec-POMDP formulation, the architecture decomposes decision-making into (i) action-conditioned Theory-of-Mind (ToM) inference, (ii) hierarchical planning, (iii) conversation interpretation, (iv) action verification, and (v) feedback-based replanning. We compare the proposed method with an ablation without ToM inference and a multi-agent reinforcement-learning policy trained offline over many goal pairs. In human-participant experiments, the proposed method required fewer interaction steps and yielded higher post-interaction trust ratings than both baselines. These results suggest that systematically decomposing the team decision problem, using LLMs as tractable surrogates for otherwise intractable inference and planning computations, and retaining conventional verification for physical feasibility can improve both task coordination and the human experience.

Video question answering requires models to ground language queries in visual evidence and, when necessary, reason over that evidence across time. Existing methods typically rely on long textual chain-of-thought rationales, even though many questions can be answered as soon as the relevant object, action, or frame is localized. We propose Dynamic Latent Reasoning (DyLaR), which first grounds a question in a short block of perception latents (continuous hidden states that encode query-relevant visual evidence), and then adaptively decides whether to append reasoning latents (continuous thoughts that reason over this evidence in latent space) before answering. DyLaR learns this behavior by grounding perception latents in verified visual evidence and distilling verified rationales into reasoning latents, followed by reinforcement learning that further refines when to reason. Across nine video benchmarks and four multimodal language model backbones, DyLaR improves average accuracy over same-backbone baselines while generating fewer than 20 tokens per query. On Qwen3-VL-4B, for example, DyLaR improves average accuracy over Qwen3-VL-4B-Thinking from 54.0 to 58.2 while reducing response length from 1,220.7 to 18.5 tokens per query. Ablations further show that grounded perception latents, rationale-supervised reasoning latents, and adaptive routing each improve accuracy.

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

On-policy training has emerged as a powerful post-training paradigm for improving the reasoning capabilities of large language models, and is often enhanced by golden trajectories from stronger expert models. However, when the expert fails on harder problems, existing trajectory-guided methods lose their main source of supervision, and these failed trajectories are typically discarded as negative samples. We argue that such failures, which we call Golden Negative Trajectories, can still provide valuable reasoning signals when treated not as demonstrations to imitate, but as flawed trajectories to reflect upon. We identify a Reflection Advantage: for hard problems, reflecting on a flawed trajectory can be easier and more effective than solving the problem directly from scratch. Motivated by this, we propose ReflectRL, a lightweight plug-and-play framework that learns from Golden Negative Trajectories during on-policy training. ReflectRL first uses these trajectories to elicit Reflective Reasoning, then applies Reflective-to-Direct Policy Transition to transfer the acquired reasoning behavior back to Direct Reasoning. Experiments across 9 benchmarks, 4 LLM backbones, and 4 on-policy training methods show that ReflectRL consistently improves reasoning performance with minimal overhead.

When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate. Across a five-model panel we identify two propagation regimes - spike-and-suppress (Phi-3.5, Gemma-2-9B) and late-accumulation (Llama-3, Mistral, Qwen2.5-7B) - and on the two models meeting an 80% identity-patch gate, sensitivity and causality are anti-correlated (rho = -0.72 to -0.88). Within-family scaling on Qwen2.5 (1.5B to 14B) shows the late-accumulation signature strengthening monotonically with scale, corroborated on a second family. We propose cascade disruption as the mechanism behind the dissociation: adapters placed at causally implicated early layers break intact downstream computation, making diagnostic-flagged sites the worst adapter placements. A fixed-harness layer sweep across four models (3.8-8B) confirms the core prediction on chain-of-thought GSM8K - the flagged sites are the most damaging adapter windows on every adjudicable model - and is sign-consistent but strongly attenuated on a multiple-choice control, consistent with damage that compounds with generation length. The sweep yields practical guidance: a training-free LRD pre-screen and a default-deepest placement rule, though absolute gains over no-adapter baselines remain small. Finally, apparent gains from a representation-stability loss reverse under an adequate generation budget - truncated chain-of-thought had been scored as empty - a methodological warning for any intervention evaluated on chain-of-thought tasks.

Reasoning-based guard models improve LLM safeguards, but decoding explicit rationales for every interaction makes them costly to deploy. Although latent-reasoning methods reduce token generation by moving reasoning into continuous states, they remain underexplored for safety moderation and lack an inspection interface for deployment. In this paper, we propose LatentGuard, an efficient and inspectable safeguard framework that brings continuous latent reasoning to guard models. LatentGuard uses a staged curriculum to progressively compress task-aligned textual rationales into compact latent states, enabling safety verdicts to be predicted directly from continuous representations. To preserve inspectability, an isolated auxiliary decoder generates compact audit artifacts on demand, keeping rationale generation off the standard inference path. Experiments show that LatentGuard-8B improves mean weighted F1 from 83.95 to 84.91 over GuardReasoner-8B, while reducing critical-path reasoning cost from 268.56 generated rationale tokens to 1.60 latent reasoning tokens. Its audit decoder achieves an audit utility score of 85.75, demonstrating an efficient and inspectable path toward deployable LLM safeguards.

Chain-of-Thought (CoT) reasoning offers a promising window into model monitoring. However, monitoring relies on faithfulness, i.e., the model output strictly derives from its reasoning trace. We identify an alignment tension where a model must be faithful enough to be monitored, yet robust enough to reject unsafe reasoning. We demonstrate that this counterbalance exists in current Large Reasoning Models (LRMs), and show ways in which it can be addressed. We introduce HazMart, a human-written dataset set in an autonomous AI shopkeeper scenario. Unlike prior work that relies on providing hints in prompts to test faithfulness (e.g., "A Stanford professor said it should be Answer A"), we propose a novel replacement-based technique, which we call Targeted Reasoning Replacement (TRR), that directly intervenes in the reasoning chain to substitute in unsafe or illogical thoughts (e.g., "Wait, the answer must be Option B [was Option A] because it is the most fitting"). DeepSeek-R1-Llama-70B exhibits high faithfulness (97.5%) but fails to reject Unsafe Reasoning (12.3%), while QwQ-32B is more robust (73.9% safety) at the cost of lower faithfulness (74.7%). Mechanistic analyses of QwQ-32B reveal that these properties are represented by anti-correlated internal directions peaking at the action-commit token. Finally, we demonstrate that representation steering can independently amplify the safety direction, increasing safe behavior by 9 percentage points while maintaining base capabilities.

World-action modeling has emerged as a promising paradigm for robotic control, as it empowers models to go beyond reacting to observations and anticipate how a scene will evolve. However, existing WAMs often incur substantial computational overhead. Pixel-space methods often allocate substantial capacity to visual details that may not be directly relevant to control, while some latent-space methods require multi-stage training to construct the reasoning space. The resulting training cost can make such methods difficult to train under modest computational budgets. In this work, we propose LiLa-WAM, a lightweight world-action model that reasons about the future in a compact latent space and can be trained end-to-end on a single 24GB GPU. Its core design is a compact latent reasoning space jointly shaped by future-state prediction and action generation, which keeps the model lightweight while remaining well aligned with control. For task specification, we further propose the Visual Transition Token(VTT), a language-free task representation that encodes each task as a direction in visual feature space. Experiments on RoboTwin~2.0, LIBERO, and real-robot tasks demonstrate LiLa-WAM's effectiveness, achieving 90.48% success across 50 RoboTwin tasks with single-GPU training.

Out-of-distribution (OOD) recommendation is vulnerable to preference shifts induced by a latent environment. Existing methods can infer latent states from logged interactions, yet the statistical meaning of the latent environment and its effect on preference remain underdetermined. We formulate this task as conditionally identifiable risk-aware recommendation (CI-RR) and propose Conditionally Identifiable Latent-Environment Recommendation (CILER). CILER uses a user-conditioned exponential family to model the latent environment and a feature-indexed polynomial to specify how it changes preference. It predicts by marginalizing item probabilities over the inferred environment distribution. Under sufficient variation, correct specification, and decoder regularity, CILER identifies the environment-sensitive representation up to the stated equivalence class. We further bound excess deployment log-risk by environment-inference error. Controlled studies test the observable consequences of sufficient variation and model specification. Experiments on three datasets show that CILER improves all twelve OOD ranking metrics under feature, temporal, and geographical shifts within shared support.

Chain-of-Thought (CoT) prompting remains the standard baseline for evaluating models' reasoning abilities. Originally, this technique was introduced to elicit step-by-step reasoning from large language models (LLMs), which would otherwise tend to directly output the final answer. However, many modern LLMs produce CoT-style responses \textit{natively} when presented with reasoning tasks, which made us revisit the effectiveness of standard CoT prompting. We evaluate several modern mid-sized language models on a math problem-solving task and find that models specialized for reasoning achieve better performance in a simple zero-shot setting than when using few-shot CoT examples - significantly surpassing officially reported results at no additional cost (e.g., from $\sim$77% to $\sim$84% for Mathstral on GSM8K). For the tested general-purpose model, a zero-shot CoT prompt is also sufficient to outperform a few-shot CoT baseline. We attribute this to a `guidance-distraction' tradeoff: standard CoT prompting also demands style adaptation, formatting compliance, and potentially undesired contextualization, which can distract models from the core reasoning task. Our findings suggest that using standard CoT prompting increasingly acts as a source of distraction as models grow stronger.

Self-consistency assumes the most frequent answer among sampled reasoning traces is the most reliable, but this can fail in causal reasoning: samples often repeat the same confounding error, and votes fragment across multiple valid answers, letting an invalid answer win despite a valid minority trace. We introduce CALVER (Causal Axiom-Level VERification), a training-free symbolic verifier that scores structured traces against Pearl's causal criteria, including -separation, backdoor adjustment, and intervention, and selects the highest-scoring candidate without consulting a reference answer. On CLEAR find-one-valid queries that admit multiple graph-valid answers, CALVER reaches 42.1% where plurality, a reward model, an LLM judge, and model confidence remain near 30% on identical frozen pools. Scaling the judge to 72B does not close the gap. In an audited clean-core subset, 11 of 21 graph-valid CALVER selections differ from the benchmark's listed answer while still satisfying the requested predicate. The advantage widens with the sampling budget and reproduces across ten published Bayesian networks, a second model family, and settings where the model must build the graph from text. CALVER also improves thresholded average-treatment-effect decisions against exact ground truth, generalizes to logic under a truth-table checker, and scores each candidate in milliseconds on CPU. CALVER needs only a causal structure, supplied outright or built from the text; wherever that holds, selection can aggregate via causal validity.

Reasoning in Multimodal Large Language Models (MLLMs) requires both fine-grained visual perception and rigorous logical deduction. Explicit text-based Chain-of-Thought (CoT) is computationally expensive and prone to visual hallucinations, while existing latent reasoning methods typically require costly training. Furthermore, directly adapting training-free LLM reasoning mechanisms to the multimodal setting yields unstable performance. We identify that this failure stems from their reliance on token-level entropy, which fundamentally conflates perceptual ambiguity (e.g., unclear visual details) with logical uncertainty (e.g., complex reasoning steps). To overcome this bottleneck, we present a novel training-free inference strategy for MLLMs that explicitly decouples perception and reasoning. We propose a novel metric, the vision-to-text attention ratio, to dynamically gauge the model's cognitive focus. Guided by this metric, our proposed framework, Attention-Guided Switching (AGS), adaptively triggers latent reasoning for perceptual tokens to preserve high-fidelity visual information in the continuous space, while enforcing explicit text generation for logical tokens to maintain structural anchoring. Extensive experiments demonstrate that our method achieves state-of-the-art performance, significantly improving both accuracy and inference efficiency by reducing autoregressive steps and latency. Code is released at https://github.com/swordAndSnow/MM26-AGS.

Large language models often reason at length before answering, increasing cost and latency. Prompts and trained settings can shorten this reasoning, but a shorter trace may only show that the model stopped sooner. Here, we evaluate paired runs of the same question at matched reasoning horizons across 198 GPQA Diamond and 500 MMLU-Pro questions. We test a numeric/concision prompt that announces a token limit for Qwen3-14B and the trained effort settings of gpt-oss-20b and -120b. The Qwen prompt shortens reasoning traces by 12-17%, while accuracy changes at matched token limits are small and mixed. A concise/early-answer instruction raises MMLU-Pro accuracy by 3.8 percentage points at 512 tokens, including +2.7 points when both runs are unfinished. Its gain at 2,048 tokens is uncertain. For gpt-oss, candidate-logit answers from completed low- and medium-effort reasoning are 14.5-26.3 points more accurate than matched-horizon high-effort answers. Most of the 512-token advantage comes from lower effort finishing earlier, while differences among unfinished runs are smaller and mixed. Wrong early answers often concentrate probability on the chosen option, so earlier stopping does not uniformly improve probability quality. In these tests, a tight deadline can favor lower effort or a concise instruction, whereas allowing high effort to finish can recover higher final accuracy. Evaluations should report correct completion before a deadline, the answer obtained when a run is stopped, differences among unfinished runs, and probability assigned to the correct answer separately.

Chain-of-thought (CoT) reasoning improves large language model (LLM) performance while also providing an observable interface to the model's reasoning process. Existing approaches that leverage verbalized CoTs to monitor reasoning correctness, however, largely evaluate the semantic correctness or consistency of individual intermediate steps, rather than how the reasoning process evolves across the trace. As a result, failures distributed across the reasoning trajectory, rather than those localized to a single incorrect step, remain comparatively underexplored. Furthermore, verbalized CoTs need not faithfully reflect the model's internal reasoning, motivating analyses that do not treat individual statements as literal accounts of internal computation. In this work, we therefore ask whether the dynamics of visible CoT can be leveraged to systematically distinguish successful from failed reasoning without assuming such semantic faithfulness. We study a range of LLMs on verifiable Boolean satisfiability tasks with variable complexity, enabling controlled comparisons near each model's capability frontier. Tagging CoT sentences by reasoning function reveals premature verification collapse on SAT problems: incorrect traces enter clause checking earlier, repeat similar operations, and finalize sooner. On UNSAT problems, models presumptuously move towards incorrect SAT conclusions, checking candidate assignments rather than deriving contradictions across constructed cases. Subsequently, a targeted proof-search prompt intervention raises Llama3-70B accuracy from 13.3% to 85%, correcting 84.6% of these errors. These results show that capability failures can manifest as distributed, task-dependent changes in the structure of visible reasoning, and that CoT dynamics agnostic to whether the verbalized trace reflects the model's internal computations can help diagnose and correct failures.

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

We test whether the "compositional ignition" reported in latent-reasoning models is real computation, an instrument artifact, or inherited from verbal training data. We grow an independent realization of a published 30M-parameter recurrent-depth reasoner from scratch (same recipe and seed), film its development, certify fidelity through a pre-registered whole-signature gate, and measure resolution in two channels at once: the vocabulary readout and the hidden state. The ignition is real and lives at the readout: arrival time rises lawfully with problem depth, resolution is sharp and holds, and the signature reproduces across two same-seed realizations with divergent training trajectories. At commitment the decision margin jumps 5.8-8.0 logits in one iteration, exceeding the 90th percentile of near-threshold non-event steps in 96% of cases; the signed margin's zero-crossing there is definitional and carries no evidential weight, so the evidence is that conditioned magnitude. The hidden-state direction snaps in raw geometry, meeting its pre-registered criterion (in the decoder's LayerNorm coordinates it attenuates just below our bar, so the composite decoder-coordinate claim is not confirmed), and then freezes in both (descriptively so in decoder coordinates; angular steps 52.9 to 1.2 degrees over eight iterations), while subsequent displacement is predominantly radial (0.961 of squared-norm) and readout-null to a measured bound (radial logit effect <=5.7e-6). An earlier velocity-trough claim is withdrawn: pre-registered normalization controls showed it coordinate-dependent. Intermediates were never recoverable through the tied readout (relay 0.00). All criteria were frozen before their data; the predictions ledger, including this paper's own withdrawn headline, ships in the companion repository.

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

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

Multi-domain machine translation (MDMT) requires more than fluent generation: it demands domain-sensitive translation decisions such as domain disambiguation, terminology control, and stylistic adaptation. Large reasoning models (LRMs) make such decisions explicit through intermediate translation steps, but our analysis across 15 domains and four translation directions shows that this explicit reasoning is double-edged: it improves long-form and high-difficulty translation, yet often drifts in terminology-intensive and stylistically constrained settings. We trace this failure to a credit-assignment bottleneck: existing methods optimize final outputs or coarse trajectories, but cannot identify which translation steps actually help the final translation. To address this, we propose PAMT, a process-aligned training framework that combines cold-start domain-aware Long-CoT supervision with reinforcement learning. PAMT uses sequence-level format and outcome rewards for the final translation, together with a step-level process reward that measures how much each explicit translation step increases the likelihood of the reference translation. Across two backbones, PAMT improves over base models, outperforms MT-specialized baselines on average, and remains competitive with strong LLMs/LRMs across in-domain, OOD, and multilingual settings.

Reinforcement learning (RL) has emerged as an effective method for enhancing the reasoning capabilities of large language models (LLMs). However, existing methods suffer from insufficient precision in feedback on generated answer trajectories and exhibit the phenomenon of problem difficulty drift. To address these challenges, we propose CVPO - Curriculum-guided Value-Variance Policy Optimization. At the response trajectory level, we find that token-level value-variance correlates with exploration intensity. Our theoretical analysis shows this variance bounds policy update magnitude. We then use the estimated trajectory value-variance to quantify the intrinsic randomness in generation. Based on this, we design a variance-aware advantage adjustment mechanism for different reward types. At the question level, we introduce a dynamic curriculum weighting method that adapts to question difficulty. This helps the model focus on tasks matched to its current ability during each training stage. Experimental results show our method outperforms strong value-based baselines like VAPO. It achieves better performance and stronger exploration, enabling more accurate and robust reasoning in language models across various math tasks.

LLM-based CAD agents produce executable parametric programs, but their correction loops may lose evidence about satisfied requirements, faulty operations, and prior repairs. We introduce TraceCAD, a recovery layer that links requested features, modeling steps, failure evidence, and candidate outcomes as persistent state. TraceCAD diagnoses likely faulty operations, searches bounded edits in their dependency regions, validates candidates through execution and preservation checks, and retains successful and failed repair outcomes in reusable skill memory. On DeepCAD-derived benchmarks with 200-model ablations and a 1K-model comparison, TraceCAD achieves competitive geometric quality in terms of IoU, Chamfer distance, and Hausdorff distance. Removing persistent state nearly halves recovery score; removing localized search more than doubles geometric regression and doubles code-agent invocations. Initializing the skill store on disjoint training models further reduces retries, token cost, and latency. These results demonstrate that persistent, localized, and reusable recovery improves final CAD quality and repair reliability.

Long-context reasoning remains a critical bottleneck for large language models, as recent recurrent-memory approaches face two inherent challenges: sequential chunk-wise updates can overwrite early critical evidence with later irrelevant content, and serial inter-chunk dependencies limit parallelism and cause latency to increase with context length. To address these issues, we propose PI-Mem (Parallel-Iterative Memory), a mechanism that processes all chunks in parallel and iteratively refines a shared memory over a bounded number of turns. In each turn, PI-Mem reads all chunks in parallel conditioned on the current memory, selects new or complementary evidence from each chunk, and merges the selected evidence into a compact shared memory for the next turn. To discourage redundant turns, we optimize the workflow through reinforcement learning with an auxiliary turn-efficiency reward, enabling the model to adaptively exit once sufficient evidence has been accumulated. We evaluate PI-Mem with Qwen3.5-35B-A3B and Qwen2.5-7B on the HotpotQA benchmark across context lengths up to 3.6 million tokens and find that it outperforms the recurrent-memory baseline by +6.25 and +7.81 absolute points while achieving 6.1$\times$ and 2.1$\times$ inference speedups, respectively. These results demonstrate that PI-Mem breaks the accuracy--efficiency trade-off in long-context reasoning and provides a scalable approach to complex multi-hop question answering over extremely long documents.

Large language models (LLMs) have demonstrated impressive performance in MQM-based translation quality (TQ) evaluation, and recent advances in large reasoning models (LRMs) promise even greater improvements. However, both LLMs and LRMs are computationally expensive to deploy at scale, while small language models (SLMs)---though much more efficient---struggle with the complex reasoning required for evaluation tasks. In this work, we present an extensive empirical study benchmarking SLMs, LLMs, and LRMs across a wide range of TQ evaluation setups, providing a comprehensive view of the current landscape and establishing best practices. To address the scalability challenge, we introduce TQLite, a novel distillation framework that enables SLMs to approach the MQM evaluation performance of the best LRM-based evaluators. Our approach leverages a multi-LRM jury to generate high-quality synthetic training data via practical data curation techniques and aggregation of evaluation responses across a diverse panel of models. Our results demonstrate that SLMs trained via TQLite achieve strong MQM evaluation performance that far exceeds off-the-shelf evaluation capabilities of standard SLMs, offering a scalable and cost-effective alternative to LLM- and LRM-based evaluators.

Although reinforcement learning with verifiable rewards (RLVR) has improved the performance of large language models (LLMs) across a variety of reasoning tasks, there is significant debate as to whether RLVR expands the reasoning capability boundary, or just improves sampling efficiency. In this paper, we investigate the nature of test-time exploration in RLVR-trained LLMs by employing controlled maze-solving experiments and extracting a tree structure from mathematical reasoning traces (BODHI-Trees) based on semantic equivalence. This helps us delineate between entropy arising from stylistic variations and genuine inferential branching. Our findings demonstrate that the policy entropy collapse observed in RLVR models is not merely syntactic, and is accompanied by a significant reduction in semantic branching entropy. While RLVR improves adherence to environmental constraints and backtracking capabilities, it constricts the space of continuations; we provide evidence suggesting that this might be responsible for the sample efficiency gains of RLVR, albeit at the cost of genuine rollout diversity.

Chain-of-thought (CoT) monitoring is an increasingly important component of AI safety stacks but relies on the assumption that a model's reasoning trace is informative about its actions. This work studies the limits of CoT monitoring through the lens of model poisoning. We demonstrate that backdoors can be implanted into reasoning models to elicit an attacker-chosen behavior while their CoT traces appear entirely benign. We find that these CoT-Hidden backdoors can be induced through simple fine-tuning recipes across reasoning-model architectures and sizes. When direct poisoning is ineffective, we introduce a curriculum training approach that progressively teaches the model to produce an attacker-chosen output while concealing the behavior from its reasoning traces. These findings suggest that CoT monitoring may be better framed as a question about the consistency between a model's reasoning trace and its final response than as anomaly detection within a trace. We further examine the mechanisms that allow models to suppress evidence of the target behavior from their reasoning traces. Causal interventions locate a trigger-conditioned activation pathway that does not depend on the visible reasoning, and residual stream verbalizations provide an anomaly warning near answer generation, but do not identify the trigger, target, or backdoor mechanism.

Optimization-based latent reasoning improves large language model outputs by optimizing instance-specific continuous states at test time while keeping model parameters frozen. Existing methods, however, typically connect these states to the reasoning trajectory through decoded tokens, making sequence-level credit assignment indirect and obscuring how latent updates shape subsequent reasoning. We introduce GradCuit (gradient through circuit), which inserts optimizable latent states at a selected Transformer layer between the hidden representations of the prompt and the generated continuation. Causal self-attention provides every continuation-token log-probability with a differentiable path to every preceding latent state through the remaining Transformer blocks, enabling reward-weighted gradients from the entire continuation to be assigned directly to the latents. Across five instruction-tuned backbones, three reasoning benchmarks, and two answer formats, GradCuit achieves an average accuracy of 64.5%, outperforming chain-of-thought prompting by 6.6 percentage points and the strongest competing method by 2.4 points. GradCuit also demonstrates greater robustness: across seven learning-rate settings, it consistently outperforms LatentSeek while reducing the standard deviation of accuracy from 1.53 to 0.82, and even its random-walk variant remains competitive with LatentSeek. For interpretability, token-level gradient attribution reveals that latent influence concentrates on reasoning-connector tokens, while layer analysis identifies early-to-middle Transformer layers as the most effective optimization space. By directly optimizing internal reasoning from outcome feedback, GradCuit opens a new axis of robust and interpretable test-time scaling, where LLMs adapt how they reason rather than merely regenerate, sample, or rerank outputs.

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

We assess 18 chapter-specific reviews of the ten mathematical results announced by OpenAI on 1 August 2026, alongside review standards, Lean formalizations, subsequent research, and mathematical references. The article audits this review record without claiming a complete reconstruction of all ten proofs. No confirmed substantive mathematical error in a principal result remains in the examined assessments, although review depth varies and some dependencies remain partly checked. Chapter 8 presents the strongest reservation: a specialist review requests major revision of compressed analytic arguments. In Chapter 6, an apparent polarity error was withdrawn after an overbar lost during PDF extraction was recovered from the typeset source. Subsequent research independently reuses the Chapter 3 proof mechanism and confirms that Connes's rigidity conjecture is false, without independently reproducing Chapter 4's stronger infinite-family result. Among the cited follow-ups, Chapter 7 receives the strongest direct theorem-level corroboration through a stronger hardness theorem. Related equality results in Chapter 8 do not verify the analytic inequality proof. Some follow-ups disclose material AI assistance. We argue that confidence should combine formal checking, human reconstruction, independent mathematical use, and a public record supporting correction of both proofs and reviews.

Reinforcement learning has become a central paradigm for improving the reasoning capabilities of large language models. Existing methods generally aim to reduce the failure probabilities induced across problems. In this paper, we introduce a moment-based perspective on policy optimization for LLM reasoning by treating the failure probability of a randomly sampled problem as a random variable and characterizing optimization objectives through its moments. Under this perspective, many existing methods optimize only a single moment of the failure-probability distribution, leaving its broader distributional structure largely uncharacterized. We propose \textbf{M}ulti-\textbf{M}oment \textbf{P}olicy \textbf{O}ptimization (MMPO), a novel policy optimization framework that jointly minimizes multiple moments of the failure-probability distribution. MMPO admits a direct operational interpretation as minimizing the expected truncated time required to obtain the first successful response. Beyond MMPO, we further develop a general moment-transformation framework that systematically induces different moment profiles and provides a unified view of a broader family of policy optimization objectives. Experiments across five mathematical reasoning benchmarks and models of different scales demonstrate that MMPO consistently outperforms strong baselines. We hope this moment-based perspective offers new insights into the design of policy optimization objectives for LLM reasoning.

Multimodal representation learning is a cornerstone of modern AI. By encoding multimodal queries and targets into vectors, it powers industrial search and recommendation and underpins modern agents. Real-world platforms with complex modalities and massive-scale content, such as Douyin, Xiaohongshu, and YouTube, demand both efficiency under billion-scale indexing and fine-grained discrimination for hard matching. Existing MLLM embedding models rarely satisfy both. Contrastive models are efficient but rely on pair-level supervision too coarse for fine-grained distinctions, while CoT-based models improve discrimination through explicit generation impractical to serve online. We present Douyin Multimodal Embedding (DME), a model trained in two stages to combine both strengths. Stage 1 performs large-scale contrastive pre-training that establishes a unified multimodal embedding space with broad modality and task coverage. Stage 2 supplements semantic sufficiency, the property that an embedding is grounded in retrieval-relevant evidence and preserves fine-grained counterpart-side semantics, via two mechanisms. Evidence-Grounded Typed Latent Reasoning organizes retrieval evidence through hidden-space latent reasoning, and Cross-Conditional Reconstruction enforces counterpart-side semantics through cross-directional autoregressive reconstruction. Both act only during training and add only marginal query-side overhead, so DME serves as efficiently as a standard contrastive encoder. On MMEB-v2, DME reaches state-of-the-art results at comparable scales for its 2B and 9B variants (74.8 and 78.4), with especially strong video and visual-document tasks. In production, DME delivers a 2.92% relative gain on Douyin's in-house offline evaluation set, is deployed across Douyin scenarios such as generative, image, and AI search, and yields a 0.1% Lifetime (LT) gain in online A/B testing on Douyin search.

Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers' trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.

Reflection is a powerful mechanism for LLM reasoning, yet its effectiveness hinges on accurately attributing failures to specific reasoning steps, a capability that current models notably lack. Existing failure attribution methods either require expensive step-by-step counterfactual testing that scales poorly with trajectory length, or treat reasoning traces as flat sequences that ignore the inherent non-linear logical dependencies. We propose a hypergraph-based paired failure attribution (HPFA) framework that attributes the failure root cause by comparing the hyperedges of the targeted failure reasoning path against a reference successful path. By reducing the search space, our method efficiently localizes root causes and enables scalable synthesis of attribution data for training a lightweight attributor model via supervised fine-tuning and reinforcement learning. Experiments on mathematical reasoning and agentic coding tasks demonstrate that HPFA can dramatically increase attribution accuracy and efficiency, and the trained attributor consistently improves reasoning accuracy at test time, outperforming baselines that lack graph structure or paired analysis.

Agentic retrieval-augmented generation (RAG) systems can fail before evidence-conditioned reasoning is tested: an agent may retrieve candidate snippets but finalize without inspecting them. We study this failure mode as a procedural property of the agent trajectory, decomposing wrong answers into pre-evidence discipline failures and post-gold-read failures using saved tool-call traces, retrieved evidence, read passages, and final answers. Across 12,000 paired trajectories on HotpotQA, 2WikiMultiHopQA, and MuSiQue, the two failure types are largely non-redundant: the both-trigger rate is in [11.2%, 13.1%] across regex and spaCy entity extractors. We then evaluate Read-Gate, a minimal runtime invariant requiring an agent to read after search and before finalization. Forced reading improves LLM-Acc by 14.9-19.9 points on trajectories that would otherwise skip reading and by 3.2-9.4 points on full minimal-reasoning cells. Additional diagnostics show that larger hidden thinking budgets do not necessarily increase evidence inspection. Together, these results indicate that evidence-gathering should be evaluated as a trajectory-level control problem, separately from answer-side reasoning.

Chain-of-thought (CoT) reasoning can improve performance on difficult video questions but often wastes decoding tokens on simple ones. We study whether a video multimodal large language model can adapt its reasoning effort to each question. We propose AdaThinkV, an adaptive framework for video reasoning that learns whether to reason explicitly without offline difficulty labels, manually tuned confidence thresholds, or an external router. During reinforcement learning, AdaThinkV samples matched rollouts in explicit reasoning and direct answering modes for each prompt. ThinkGain estimates the prompt-level utility of explicit reasoning by balancing its accuracy gain against additional response length, providing supervision for both conditional response generation and autonomous mode selection. For difficult prompts, limited rollout exploration can yield groups in which every response is unsuccessful and accuracy rewards show little variation, providing insufficient signal for learning. We therefore introduce Variance Recovery Policy Optimization (VRPO), which retains and progressively expands these groups to recover informative signals from prompts that are difficult yet solvable. At inference, AdaThinkV selects a response mode and generates the response in a single autoregressive sequence. Across a unified suite of video reasoning evaluations, AdaThinkV achieves a mean accuracy of 40.79 with an average of 257.20 output tokens, outperforming the strongest evaluated adaptive baseline by 2.98 points while using 22.7% fewer tokens. Project page: https://trilarflagz.github.io/AdaThinkV/

Multimodal Large Reasoning Models (MLRMs) have achieved strong performance on tasks requiring visual understanding and multi-step inference. However, as reasoning trajectories grow, models may become less effective at using information established earlier in the context, increasing the risk of reasoning errors. Existing approaches primarily address this problem by sustaining visual grounding throughout reasoning. However, reasoning also transforms visual observations into task-specific relations, constraints, and intermediate conclusions whose influence may weaken over long trajectories. Our attribution analysis suggests that correctness is not consistently separated by image attribution alone, but is more closely associated with whether trajectories retain and integrate such reasoning-derived information across stages. Motivated by this, we introduce TRAM (TRajectory-derived Auxiliary Memory), a training-free method that augments standard decoding with an auxiliary memory pathway derived from the model's own reasoning trajectory. TRAM consolidates completed reasoning into a compact latent memory, updates it online through fast and slow recurrent streams, and feeds it back into selected decoder layers through a lightweight residual pathway. Experiments across four MLRM variants on eight benchmarks show that TRAM improves performance over vanilla decoding on mathematical, scientific, and general visual reasoning tasks without additional training.

Multimodal large language models (MLLMs) have advanced visual understanding and reasoning, yet their static parametric knowledge limits their ability to address knowledge-intensive and dynamically evolving open-world problems. To move beyond this limitation, multimodal deep search has emerged as a key direction for open-world information access, evolving from single-turn factual retrieval toward long-horizon, multi-turn search guided by visual evidence. However, existing methods typically confine vision to the input or answer stage, overlooking its role in intermediate reasoning, and lack designs tailored to long-horizon interaction. Consequently, visual evidence rarely drives continued retrieval, constraining both interaction depth and reasoning span. To address these limitations, we propose DeepVoyager-VL, a long-horizon multimodal deep-search framework for vision-in-the-loop search. Specifically, we construct a multimodal event graph to drive data synthesis, yielding problems with intermediate visual dependencies and long reasoning chains. We then design an agent framework for active visual acquisition and on-demand image loading. Finally, we fine-tune models on the synthesized data without reinforcement learning. Extensive experiments across ten multimodal search benchmarks demonstrate the effectiveness of our method.

Verifier-guided reinforcement learning has become a powerful paradigm for improving LLM reasoning. In multi-turn settings, models receive a verifier score after each turn and iteratively refine their outputs. Although such scores provide dense feedback, they do not directly provide dense credit: a score measures the quality of the current output, while credit should measure how the current turn changes the refinement trajectory. We propose TCPO, a turn-level credit assignment method for verifier-guided multi-turn RL. TCPO casts credit assignment as score-to-credit conversion and constructs turn-level advantages through reference-based comparisons: retrospective credit captures immediate progress and regression relative to the best prior state; hindsight delayed credit identifies non-improving turns with later payoff; and selective fixed-history counterfactual estimation refines high-surprisal turns under the same history. Experiments on math reasoning, code generation, and AppWorld agent tasks show that TCPO improves or matches the strongest baselines across model scales, task domains, and verifier types. TCPO achieves the best or tied-best best-turn Pass@8 on Qwen3-4B and DeepSeek-R1-Distill-Llama-8B, reduces turns to success, and improves multi-turn agent performance. These results highlight score-to-credit conversion as a central ingredient for verifier-guided multi-turn policy optimization.

KV cache compression is commonly evaluated by final-answer accuracy, implicitly assuming that preserving the answer also preserves the reasoning that supports it. We test this assumption for large reasoning models and show that it can fail: under compression, correct answers and the validity of their visible supporting rationales can be preserved at different rates. We study this failure with a controlled fixed-trace replay protocol, which holds reasoning content fixed and isolates whether compression preserves usable information from an already available trace. We evaluate ten token-eviction KV compression methods and one quantization method on three models across mathematical reasoning, scientific QA, clinical calculation, and long-context retrieval. We measure final accuracy, answer-chain consistency, and perturbation faithfulness. Across tasks, token-eviction methods can preserve competitive final-answer accuracy while substantially degrading chain support or perturbation faithfulness. We call this the answer-evidence gap. A coverage-preserving quantization control is substantially less affected, suggesting that the failure is tied less to KV memory reduction itself than to losing access to parts of the reasoning trace. Code is available at https://github.com/famous-blue-raincoat/Safe_KV_Compress.

Simulating human-like Theory of Mind (ToM) has been a longstanding problem in natural language processing (NLP). To address this, existing works introduce a reasoning step of event hiding (a.k.a. perspective-taking), where events unknown to a character are removed before question answering. However, resorting to event hiding for ToM reasoning presents a performance degradation issue due to the strict output format constraints involved in event hiding. To mitigate this issue, we propose generating perspective-taking outputs as free-form explanations without event hiding, but this poses a notable yet underexplored challenge: LLMs need to inhibit responses to events unknown to characters, because the absence of event hiding exposes LLMs to these events throughout reasoning. To address this challenge, we hypothesize and empirically verify that LLMs can achieve such inhibition if a character's lack of knowledge about events is made explicit during reasoning. Based on this finding, we introduce PICTURE, a new prompting method that enables LLMs to generate a character's lack of knowledge within free-form Chain-of-Thought (CoT). Experimental results show that PICTURE outperforms existing prompting methods by an average of 7.3% on false-belief tasks.

Latent reasoning allows language models to carry out intermediate reasoning in continuous latent representations rather than fully externalizing it as discrete chains of thought. However, assigning credit to such latent thoughts from answer-only rewards is difficult: a single final answer mixes thought quality with answer-sampling noise. We propose \textbf{Latent Thought Credit (LTC)}, a hierarchical credit-assignment framework for latent reasoning. For each prompt, LTC samples multiple latent thoughts, fixes the context after each thought, and estimates thought-level expected reward by averaging rewards over multiple answers generated from that fixed context. LTC uses thought-level advantages to optimize the latent-thought phase, answer-level advantages to optimize the answer phase, and an advantage-weighted thought-matching objective that helps the policy reproduce high-credit latent thoughts. We instantiate LTC in a GRPO-style on-policy training framework and evaluate it across mathematical reasoning and STEM multiple-choice tasks. LTC achieves the best average accuracy among the compared methods, while ablations and fixed-context diagnostics show that multi-answer estimation reduces reward-estimation error and mitigates ambiguous or incorrect thought-level credit.

Tool-using language-model agents repeatedly rediscover procedures they have already executed, producing traces that mix reusable structure with retries, exploration, accidental ordering, and repeated lookups. We present TraceCompiler, a skill-guided system that mines clusters of noisy agent traces and compiles them into executable, mostly deterministic workflows. It admits an inter-tool dependency only when a consumer argument contains a value attributable uniquely to an earlier producer; every hard edge carries an auditable evidence tuple, and ambiguous relations are marked suspected and impose no ordering constraint. Bindings are classified as constants, user inputs, copied outputs, transforms, or residual LLM decisions. On T1, a mechanized form of the rule recovers producer-consumer dependencies at 0.928 precision and 0.943 recall over 15,775 def-use edges of its training split, against 0.711 F1 for adjacency and 0.712 for a frequency-thresholded directly-follows measure on identical data; the compiler skill run blind reaches 0.992 on 250 of those edges. On AppWorld we replay released trajectories in the deterministic simulator to recover masked return values and measure the rule against 563 token edges at 0.993 precision - a self-consistency check, since replay injects tokens by a related heuristic. We compile two recurring intents: a Venmo money-request intent reduces 34 observed API calls to 11 runtime calls and, under leave-one-out execution against the benchmark's own state tests, passes 15 of 21, the failing fold escalating rather than acting because its required branch was never observed; and a Spotify/Todoist intent the compiler correctly refuses to compile, because an irreversible side effect is under-determined. We measure call reduction but not offline compilation cost, so we claim no net efficiency result.