Overview

Large reasoning models study how language models can solve difficult problems by allocating computation across intermediate steps, tools, and alternative solution paths. Common approaches include reinforcement learning with verifiable rewards, supervised reasoning-trace distillation, tool use, search over sequential or parallel paths, and latent recurrent computation. Evaluation spans mathematics, coding, science question answering, logic, and multimodal tasks, measuring not only accuracy but also compute use, robustness, and the quality of intermediate reasoning. A central tension is that extra thinking can help on hard problems but also produce loops, errors, and safety risks, so researchers are developing adaptive stopping and compression while testing whether reasoning traces reliably reflect how answers are formed.

Latest papers

349 papers8 papers in the last four weeks, down 74% on the four weeks before. 0.1% of all new papers.

Nov 10, 2026

Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout

During extended wakefulness, mammalian neocortical networks experience localized periods of transient neuronal silence termed off-states. Here, we examine how analogous spatiotemporally concentrated disruptions affect sequential reasoning in autoregressive Large Language Models (LLMs). Evaluating 250 paired derivations on the GSM8K benchmark across two 8B models (Qwen3-8B and LLaMA-3.1-8B-Instruct), we contrast clustered bursts of attention head masking against budget-matched diffuse i.i.d. dropout. Concentrated bursts induce substantial reasoning degradation: at onset parameter B=0.10\mathcal{B}=0.10 (scheduling ∼5.7%\sim 5.7\% of vulnerable head-token volume), Qwen3-8B accuracy drops from 91.2%91.2\% under diffuse noise to 78.8%78.8\% under burst dropout (Δ=−12.4Δ= -12.4 percentage points [pp], exact McNemar p=5.5×10−6p = 5.5 \times 10^{-6}), while LLaMA-3.1-8B-Instruct drops from 84.0%84.0\% to 76.8%76.8\% (Δ=−7.2Δ= -7.2 pp, p=0.0198p = 0.0198). Conversely, dispersing the identical cumulative budget uniformly across generation steps via stratified quintile matching yields substantially milder degradation (retaining 91.2%91.2\% in Qwen and 84.0%84.0\% in LLaMA). In an exploratory factorial dissociation (N=100N=100), single-step dropout density is associated with an acute drop in a single schedule draw (−10.0-10.0 pp on Qwen, pholm=0.0425p_{\text{holm}} = 0.0425), while multi-token continuity exhibits an inconclusive compounding trend (−7.0-7.0 pp, pholm=0.1892p_{\text{holm}} = 0.1892). Error analysis reveals that Qwen failures reflect derivation errors within formatted responses (conditional accuracy drop of −11.4-11.4 pp), whereas in LLaMA 22.2%22.2\% of discordant losses lack a final answer format line (conditional gap narrows to −5.0-5.0 pp). These findings demonstrate that autoregressive reasoning chains are particularly vulnerable to concentrated structural disruptions, providing empirical benchmarks for fault tolerance in transformer architectures.
Nov 9, 2026

Adaptive Entropy-Sparing for Efficient Reasoning

Large Reasoning Models (LRMs) produce explicit chains of thought to improve transparency, but they frequently overthink, generating long yet inefficient reasoning chains that inflate computational costs and risk introducing errors. While numerous reinforcement learning methods struggle to penalize verbosity to achieve conciseness, they typically apply the penalty uniformly across lengthy sequences, which leads to credit confounding, that is, the failure to distinguish essential reasoning steps from superfluous ones. In this paper, we study a measurable proxy for this confounding: high-entropy tokens that often coincide with surface transition points where reasoning re-enters verification. Based on this insight, we propose Adaptive Entropy-Sparing (AES), which selectively spares high-entropy tokens that occur before an efficient in-group reference, while penalizing overlength high-entropy suffix tokens associated with superfluous branching. Our AES thereby navigates the suppression-exploration trade-off, reducing overthinking while preserving average reasoning accuracy. We also present an entropy-length model that motivates why local uncertainty can serve as a proxy for future reasoning cost. Across eight reasoning benchmarks, AES reduces reasoning length by up to 52.9% while increasing mean accuracy by up to 3.8% over the base models, consistently outperforming existing efficiency-oriented methods. Our code is available at https://github.com/MobiusDai/AES.
Nov 4, 2026

ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models

Large reasoning models (LRMs) perform well on challenging benchmarks, yet it remains unclear whether they form generalizable abstractions across conceptually related reasoning tasks. The Abstraction and Reasoning Corpus (ARC) provides a controlled testbed for studying abstraction, requiring solvers to infer transformation rules over two-dimensional grids from a small set of examples. While ARC evaluates rule abstraction within individual tasks, different tasks may share higher-level structure despite requiring different transformations\unicodex2014\unicode{x2014}for example, several tasks may involve reasoning about objects contained within other objects. Benchmarks such as ConceptARC capture this structure by grouping ARC-style tasks into recurring spatial and semantic concepts, but they often contain only a few manually constructed tasks per concept. We introduce ConceptARC at Scale, a generator for creating thousands of ARC-style tasks targeting a specific concept. Using this generator, we construct ConceptARC-100, a dataset of 1,600 new, manually reviewed tasks, with 100 tasks for each of the 16 concepts defined by ConceptARC\unicodex2014\unicode{x2014}ten times the size of the original benchmark. We evaluate a range of recent LRMs and models trained from scratch on ConceptARC-100 to characterize which concepts they solve reliably and where they struggle. We further analyze the models' internal representations and find clear semantic structure: task representations cluster according to the operations applied to objects, such as recoloring, removal, or translation, even when realized through substantially different rules and scenarios. Together, our results suggest that models form stable abstractions of transformation operations that generalize across diverse tasks.
Oct 31, 2026

Compile as Hypothesis: A Headroom Typology for Prompt Optimization on Real Reasoning Models

Automated prompt compilers search for or rewrite prompts under the implicit assumption that optimization consistently improves downstream task performance. We evaluate this assumption on frontier reasoning models through rigorous paired, same-day measurements. Across four optimizer families (genome search, profile-guided mutation, GEPA, and ESPO) and eight tasks, structured search yields no statistically significant accuracy gains on strong reasoners. However, a discrete structural genome provides essential architectural stability, preventing catastrophic bimodal seed-collapse (where unconstrained search collapses toward the judge floor in 3/5 runs). Through 190-problem evaluations on knowledge-intensive benchmarks, we establish a three-dimensional headroom typology: prompt optimization cannot compensate for intrinsic knowledge deficits or physical token truncation, but successfully resolves protocol and formatting errors. On gpt-6, while free-text reflective search (GEPA) yields a +.14 gain on a task where structured search remains flat, qualitative ablation reveals that ~.035 stems from judge-rubric reconstruction while ~.105 reflects genuine instruction refinement. Our pre-registered AutoAPC-Select deployment gate matches the oracle policy in 5/7 groups and rescues seed-collapse arms with zero regret. Finally, we expose silent backend-mixing in commercial APIs during outage windows and provide automated per-call audit verification tools to ensure empirical evaluation fidelity.
Oct 31, 2026

Steer, Don't Tell: Breaking Reasoning Plateaus with Reachability-Aligned Self-Distillation

Post training for large reasoning models remains insufficient on challenging problems, whose correct trajectories are hard to sample. Though prior works address this difficulty by integrating supervised fine tuning on expert trajectories with reinforcement learning, such external experts are not always available. Existing self improvement methods instead use reference solutions to create a privileged self-teacher. Yet privileged teacher correctness does not imply student reachability. Solution conditioning shifts the teacher distribution toward actions unlikely under the unprivileged student, producing seemingly "correct" trajectories that are unreachable or even mal-formed learning targets. Accordingly, we identify reachability alignment as the key to self constructed supervision and propose Self-Steered Reasoner (SSR). SSR realizes this principle through student support projection, which transforms the privileged teacher distribution into a directional signal only over reasoning actions locally reachable under the unguided student. Verification then retains correct trajectories sampled from this reachable space. To further boost the sampling efficiency, an iterative refinement framework is adopted, using diagnostic guidance based on current trajectories. On problems with pass@64 = 0 even after extensive RL, SSR recovers correct and learnable trajectories for 74.6% without parameter updates. Fine tuning on these trajectories followed by resumed reinforcement learning improves aggregate accuracy by 4.9 percentage points on AIME25/26 and HMMT26. These results show that privileged guidance becomes effective self-improvement data when aligned with the reasoning actions the student can reach.
Oct 30, 2026

Early Stopping in Large Reasoning Models: A Survey and Empirical Audit of Efficiency and Reliability

Large reasoning models can improve accuracy by generating longer reasoning traces, but some of this computation may be unnecessary. Early-stopping methods aim to stop once enough reasoning is available to answer. Evaluating these methods requires accounting for the cost of deciding when to stop and the reliability of the resulting answer. We evaluate different methods across five reasoning models on mathematical and scientific reasoning, code generation, and factuality tasks by using a shared rollout replay protocol. We introduce Target-Integrated Token Efficiency (TITE) to measure efficiency relative to the oracle across accuracy targets. We also examine correctness at later exit positions, transfer across models and tasks, post-stop reasoning continuation, and retained justification quality. In our experiments, correctness can decrease at later positions beyond what the fitted decoding-noise model predicts. Reliability checks also reveal continued reasoning after forced termination and weaker supporting reasoning in some correct early-stopping outputs. The implementation can be found in https://github.com/hung20gg/audit-early-stopping.
Oct 26, 2026

Reasoning Patterns for Efficient AI Coding: An Empirical Study

Large Reasoning Models (LRMs) generate long reasoning traces that expose intermediate inference behavior, offering opportunities to study and control reasoning in AI coding. However, the underlying patterns in these traces and their potential for analyzing and guiding coding models remain poorly understood, limiting systematic characterization and inference-time control of LRM reasoning. We conduct an empirical study of reasoning patterns across four representative coding tasks. We transform reasoning traces into structured action sequences and mine recurring fragments using within-task support, cross-task contrasts, and valid--invalid support contrasts. We characterize four general reasoning patterns by coding agents: design-to-implementation for code generation, state tracing for execution reasoning, evidence--fix--recheck for debugging, and semantic mapping for translation. We further identify positive and negative patterns associated with valid and invalid reasoning traces, respectively. Positive patterns reflect verification-driven convergence, whereas negative patterns reflect repeated design revision, reflection, and knowledge recall without closure. To demonstrate their practical utility, we evaluate three pattern-based applications: task identification, pattern-guided prompting, and pattern-aware early stopping. Features capturing general reasoning patterns enable accurate task identification from reasoning traces without accessing task descriptions or final answers. Prompting guided by positive and negative patterns improves the overall accuracy of downstream tasks, while negative pattern-aware early stopping substantially reduces API-reported token usage with only a modest decrease in pass@1. All three applications also generalize to previously unseen models, showing that reasoning patterns can support both the analysis and control of LRM behavior in AI coding.
Oct 19, 2026

CALM: Consistency-Aware Logit Masking in Decoding Space for Mitigating Overthinking of Large Reasoning Models

Large reasoning models (LRMs) have achieved strong performance on challenging mathematical and scientific tasks through long chain-of-thought reasoning, but they often suffer from overthinking with unnecessary reasoning steps. Existing truncation-based methods typically detect consecutive trial-answer agreement and force early exit, yet such local stopping signals do not necessarily indicate final-answer stability and can disrupt useful subsequent reflection or verification, potentially leading to performance degradation. We propose consistency-aware logit masking (CALM), a training-free decoding intervention for mitigating overthinking in LRMs. CALM first monitors trial answers at step boundaries and then applies logit masking to action transition point (ATP) tokens such as "Wait", "Alternatively", and "Hmm", suppressing redundant transitional reasoning while allowing the model to continue its thought process more naturally. Experiments across four advanced LRMs and five benchmarks show that CALM reduces token consumption by 18.6%-25.0% compared with full CoT reasoning while maintaining performance, and can even improve accuracy by 0.07%-1.48% by reducing redundant reasoning. LLM-as-a-Judge results also demonstrate that CALM provides more coherent and concise results.
Oct 15, 2026

ARBOR: Locating Reasoning Failures in Long Chains of Thought through Reasoning State Tracking

Large reasoning models solve complex problems through long chains of thought, in which prior work has documented failure modes such as overthinking, underthinking, and unsupported deductions. Existing methods detect these failures from local signals: textual similarity to nearby steps, switch tokens, or an LLM's judgment of a step's correctness. However, what a step contributes to the solution depends on the state that earlier steps have established, so identifying failures requires a global view that tracks that state across the trajectory. Therefore, we propose ARBOR, a reasoning-failure attribution framework that represents each trajectory as a tree of reasoning states, built as each step creates, verifies, prunes, or revisits a state, and attributes each failure to a relation between a step and the state preceding it. Using ARBOR, we refine the three failure modes into distinct subtypes that harm reasoning through different mechanisms and at different costs. We find that the tree state ARBOR maintains before a step separates failure steps from the at-risk steps at which the failure did not occur. We discuss how ARBOR integrates with existing mitigations at inference time and in training data. Code, data and interactive examples are available at https://inklab.usc.edu/ARBOR.
Oct 15, 2026

Are Large Reasoning Models Good Everyday Reasoners? A Comparison with Non-Reasoning Models on Logical Reasoning Tasks

Large reasoning models (LRMs) that leverage test-time computation achieve strong performance on mathematical and scientific benchmarks and are increasingly deployed in real-world applications. However, their ability to perform everyday logical reasoning, which often departs from textbook formal logic, remains underexplored. In this work, we systematically evaluate how well reasoning models align with everyday reasoning patterns by introducing a benchmark based on linguistic and pragmatic reasoning phenomena. We also compare reasoning models with non-reasoning models under controlled prompting conditions. We find that, in many cases, (1) reasoning models show degraded performance in everyday reasoning when such reasoning deviates from formal logical rules, (2) some non-reasoning models with chain-of-thought prompting outperform comparable reasoning models in everyday reasoning, and (3) prompting with additional instructions provides limited control over reasoning behavior. These findings suggest that current reasoning models do not reliably capture everyday reasoning and point to limitations in post-training, motivating improved training objectives and evaluation across diverse forms of reasoning.
Oct 15, 2026

Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally

When a speaker says ``some students passed'', listeners infer that not all of them did. Reasoning models now let users decide how long a model thinks before it answers, but we do not know what extra thinking does to such everyday inferences. We study this directly: we keep a model's weights fixed, change only how much it thinks, and record which reading it chooses. A second tool cuts the model's own reasoning at successive points and asks for an answer, showing where the reading changes. Thinking makes models read logical words as logic: it cancels the implicatures of some (not all) and or (not both) in most tested open reasoning models and in GPT-5, while context-driven inferences such as indirect answers improve. The trace shows why: models first entertain the enriched reading and cancel it later. Answering from that early reading, an exploratory inference-time remedy, recovers these implicatures without losing literal accuracy. Evaluations of reasoning models should measure how they interpret, not only how accurate they are.
Oct 14, 2026

Beyond Correctness: Compute-Responsive Data Selection for Reasoning Model Post-Training

Reasoning language models increasingly rely on test-time compute scaling, yet most existing work studies how to allocate inference computation while treating the model itself as fixed. We ask a complementary question: can reasoning models be post-trained to make better use of a given amount of test-time compute, and which examples are most useful for this goal? Our key observation is that reasoning problems differ substantially in their response to additional computation. Some are already solved under limited compute, others remain unresolved even after further reasoning, while a distinct subset changes from complete initial failure to success when additional reasoning is provided. We call these compute-responsive (CR) examples and introduce Compute-Responsive Data Selection (CRDS), a post-training data-selection principle that prioritizes examples with empirically demonstrated marginal utility from additional reasoning. On Qwen3-8B, SFT on CR examples recovers 45.3% of held-out low-compute failures, compared with 40.6% when training on already-solved examples. The advantage extends beyond the selected CR split: SFT-CR reaches 51.8% overall accuracy on an independently sampled test distribution, with the gain concentrated on compute-responsive problems. CR training also shifts the compute--accuracy frontier across inference budgets from 512 to 4,096 tokens, outperforming matched non-CR training throughout and reaching 48.1% recovery at the largest budget. The same pattern is supported across multiple data sources, remains visible under on-policy distillation, and extends to Llama-3.1-8B-Instruct when CR sets are reconstructed separately for that model. Better test-time compute utilization can therefore be learned through post-training, with recoverable responses to additional reasoning providing a particularly informative signal for selecting training examples.
Oct 14, 2026

To Reason but Not to Stop: Understanding and Recovering Pruned Large Reasoning Models

Weight pruning can degrade the accuracy of large reasoning models (LRMs), yet it remains unclear whether it damages reasoning itself. We find that much of this degradation stems from a failure to stop rather than to reason. As sparsity increases, pruned LRMs increasingly fail to terminate within the token budget, often after the correct answer has appeared in their reasoning, instead falling into repetitive loops, even though they still recognize where the dense model would stop. We further find that the hidden-state difference between terminating and looping continuations is consistent across pruned checkpoints, suggesting a shared direction associated with committing to an answer. Building on this, we propose Termination-Guided Activation Steering (TGAS), a training-free method that adds this direction to the residual stream at inference time. Across six LRMs, three pruning pipelines, and math and code benchmarks, TGAS improves termination and overall accuracy while reducing repetitive reasoning and generation length, showing that this failure is actionable.
Oct 13, 2026

Efficiently Learning To Reason or Not to Reason: Root-token Policy Optimization for Adaptive Thinking

Large reasoning models (LRMs) achieve strong performance by externalizing explicit reasoning traces before producing the answer, yet suffer from overthinking challenge that allocates uniformly heavy computation to queries of varying difficulty. While proprietary models mitigate this via opaque routing, open-source LRMs still lack an efficient mechanism to internalize adaptive reasoning due to both expensive training cost and limited disclosure of training recipes. In response, we introduce RPO (Root-token Policy Optimization), a framework that enables LRMs to self-determine when to reason by training only the initial root token (e.g., whether to invoke the <think> tag) via group relative reward and group-wise advantages. By focusing on this pivotal branching point, RPO drastically reduces training overhead and VRAM usage. Across multiple model families and scales, RPO learns difficulty-aware adaptive thinking at just ~2% of the training compute of prior adaptive-reasoning methods.
Oct 11, 2026

From Candidate Tokens to Final Outcomes: Decoding Relative Long-Horizon Candidate Value from Reasoning Model Representations

Large reasoning models generate reasoning trajectories token by token. At a given prefix, multiple next-token candidates may be plausible, yet shape subsequent reasoning and differ in their chances of leading to the correct answer. Whether these differences are reflected in internal representations remains unclear. To study this question, we define relative long-horizon candidate value to quantify within-prefix differences and construct the Reasoning Candidate-Token Value (RCTV) dataset, pairing mathematical reasoning prefixes with plausible next-token candidates and their empirical long-horizon candidate values. Using RCTV, we examine hidden states, multilayer perceptron outputs, and attention outputs at candidate positions with sparse linear probes. Experimental results provide evidence that relative long-horizon candidate value is linearly decodable from these representations at multiple layers. Candidate-value probes can also outperform probability- and embedding-based baselines. Furthermore, a controlled online intervention experiment shows that using the decoded signals to guide token selection during generation can improve the chances of eventual success. Our findings provide evidence that internal representations encode information about relative long-horizon candidate value, offering insights into how local candidate-level representations relate to long-horizon reasoning outcomes.
Oct 11, 2026

RECAP: Calibrating Low-Bit Reasoning Models on the History They Read

Low-bit quantization reduces the cost of serving reasoning models but can substantially degrade long chains of thought. Each step reads cached keys and values spanning thousands of tokens, allowing quantization errors to accumulate along the trajectory. Standard calibration operates on short segments that omit the earlier context needed by later reasoning steps. Even when this context is retained, cache reconstruction error alone does not capture how quantization errors affect subsequent attention. We present RECAP, a post-training quantization method that calibrates reasoning models on the history they read. Dynamic-prefix calibration separates long historical context from short-block reconstruction. It regenerates prefix keys and values with the current quantizer and rotates the active block to supervise every position along the trajectory. Prefix read geometry further supervises historical cache errors according to their effects on attention outputs, weighting key and value errors asymmetrically. With both components applied only during calibration, RECAP outperforms the baseline across multiple reasoning models and quantization settings. Under W4A4KV4, RECAP improves accuracy over the baseline by 13.3 percentage points on MATH500 with DeepSeek-R1-Distill-Qwen-1.5B and 15.8 points on AIME-120 with Qwen3-4B. These results highlight the importance of preserving historical dependencies during calibration for low-bit reasoning.
Oct 8, 2026

REDaCT: Collapse-Free Unlearning in Large Reasoning Models

Unlearning in large reasoning models (LRMs) requires removing forget-set information from both the reasoning trace and the final answer. Existing approaches achieve this objective at the cost of reasoning collapse, where the model fails to produce a coherent chain-of-thought (CoT) and does not properly terminate its reasoning process. To address this challenge, we propose REDaCT (REgenerate via Detection and Corrective Tuning), an inference-time unlearning framework for LRMs. REDaCT consists of two key components: (1) a binary leak detector that identifies sentences in the reasoning trace that leak forget-set information, and (2) an embedding optimization module that learns a corrective embedding offline, such that when injected at the first detected leak during inference, it guides subsequent reasoning towards sentences that do not leak forget-set information. Experimental results show that REDaCT completely avoids reasoning collapse, substantially improving reasoning preservation compared to state-of-the-art unlearning methods while maintaining perfect answer forgetting and near-perfect CoT forgetting. Our code is available at: https://github.com/jmaccoby/redact.
Oct 8, 2026

Learning to Plan by Looking Back: Hindsight Hierarchies for Training Reasoning Models

We introduce a self-improvement loop for reasoning models based on the following observation: Even when the difficulty of a problem exceeds the model's current solving abilities, an additionally supplied solution might enable the model to extract useful solution ideas in hindsight. We operationalize this by jointly training the same model to exhibit the following three capabilities: predicting solution ideas from problems alone, reverse-engineering ideas from problems and known solutions, and solving problems using provided ideas. The loop alternates between reverse engineering such ideas from problems with supplied solutions and using these ideas as additional supervision for joint training of all three capabilities. We give a formal specification of our method and a concrete instantiation for interactive theorem proving in the Lean theorem prover; empirical evaluation remains future work.
Oct 8, 2026

RL-ARC: Calibrating Large Reasoning Models via Reasoning-guided Uncertainty

Language models (LMs) are commonly trained with Reinforcement Learning with Verifiable Rewards (RLVR) to enhance their reasoning capabilities. However, since RLVR does not explicitly account for calibration during training, it can lead to severe calibration degradation, including overconfidence. Recent calibration-aware training methods for LMs, which incorporate objectives for uncertainty estimation into training, improve calibration but still exhibit overconfidence under distribution shift, while sacrificing reasoning performance. To this end, we propose RL-ARC, a calibration-aware training framework that jointly leverages reasoning confidence and answer confidence. Specifically, RL-ARC leverages reasoning confidence as an auxiliary signal for calibrating answer confidence, applying it as reasoning-guided regularization for correct cases and as an overconfidence penalty for incorrect cases. Comprehensive results across ID and OOD settings show that, beyond improving calibration, RL-ARC enables reasoning models to adaptively estimate confidence based on the given question without substantially sacrificing reasoning performance, thereby highlighting the importance of reasoning confidence for training reliable reasoning models.
Oct 7, 2026

The Long Road to the Same Answer: Cognitive Bias Under Escalating Reasoning Budgets in Large Language Models

Reasoning models allocate extra computation at inference time and present their answers as the product of deliberate thought. If this deliberation works the way dual-process accounts of human cognition suggest, longer thinking should weaken the classic decision biases that fast, intuitive judgment produces. Using 30 vignettes covering six biases (anchoring, framing, loss aversion, escalation of commitment, availability, confirmation) from an established benchmark, we run a dose-response study across four model families, pairing each reasoning model with a matched non-reasoning sibling and requesting thinking ceilings of 0, 1,024, 4,096, and 8,192 tokens, for 12,350 API calls. Because a requested ceiling is not the same as realized deliberation, we use the reasoning tokens each call consumed as the dose. First, reasoning models are not less biased than their siblings; the point estimate leans the other way in every family, but the item-level pooled contrast is not reliable (Delta = +0.031, t(29) = 1.45, p = .157). Second, bias magnitude does not reliably fall as realized deliberation grows: no slope is significantly negative, and where anything moves it is the signed score drifting further from the human direction. Third, anchoring is the only bias in the human direction (d = 1.89). Four of the other five lean the opposite way in all seven models; with five items per bias, that reversal is reliable for framing and directional for escalation of commitment, confirmation, and loss aversion, while availability is absent. A one-line instruction to restate the anchor before answering lowered anchoring on all five anchoring items, which no amount of additional thinking did, although the effect does not reach significance (p = .057). The results argue against treating test-time reasoning as a rationality guarantee and for auditing deployed models bias by bias.
Oct 4, 2026

Monitorability Disposition in Large Reasoning Models

Monitoring the chain-of-thought (CoT) of large reasoning models (LRMs) is a common way to detect misbehavior in real-world practice. However, current monitoring is passive: a separate model inspects the session only after execution. This means harm may already have occurred before it is caught. An active alternative is to have the model self-report its misbehavior as it happens. Whether models are willing to do this, however, is unknown. We introduce "monitorability disposition": a model's willingness to make itself monitorable and stay monitored throughout inference when warranted. We measure it as the fraction of warranted cases in which a model self-reports its own misbehavior via tool calls to available monitoring channels. We evaluate four LRMs on three misbehaviors (sycophancy, reward hacking, and bias) while varying the available monitors (AI and human) and the pressure to use the monitoring tools. We find that when tool use is optional, models self-report in only about 16% of warranted cases on average. Increasing tool-use pressure does not improve reporting where it matters: high-severity misbehavior is never self-reported. Models also systematically select the monitor they perceive as least strict. Overall, we identify monitorability disposition as a new contributing factor to model monitorability: when sufficiently strong, it keeps models seeking monitorability throughout inference.
Oct 2, 2026

Reasoning Models Are Accurate but Unsound on Identification

A reasoning model asked whether a causal effect is recoverable from observational data can fail in two ways: it refuses an identifiable query or answers a nonidentifiable one. The latter is more consequential, as no observational data can validate the claimed formula. Measuring this failure requires queries that are provably non-identifiable, which prior evaluations lack, and grading that accepts correct formulas in any equivalent form, which string matching cannot provide. We build CERTID, a formal identification pipeline that addresses both limitations. CERTID uses the sound and complete causal identification algorithm ID to certify whether an effect is identifiable from a given graph and query, and verifies returned formulas against structural causal models whose interventional distributions are known exactly. CERTID further develops theoretical results to mitigate structural leakage, repair non-identifiable queries, and establish grading guarantees. We evaluate three frontier reasoning models (Gemini Flash, Gemini Pro, and GPT5.5) on 1,200 certified instances spanning 4 to 50 vertices. Accuracy proves a poor proxy for soundness: on identical instances, the false-claim rate on non-identifiable queries varies by seventeen-fold across models. We also find that models decide identifiability with 97-100% accuracy on graphs generated after the strongest model's training snapshot. Instances, the certification procedure, the verifier, and per-instance records are available at https://anonymous.4open.science/r/certid-D718.
Sep 30, 2026

Interpreting Reasoning of Large Language Models via Partial Information Decomposition

Large reasoning models (LRMs) have achieved substantial improvements in solving complex mathematical problems, but often produce lengthy, repetitive, or erroneous reasoning trajectories. In this work, we introduce a new interpretability framework, SLIDER, to evaluate the quality of the reasoning process. SLIDER leverages an emerging body of work from information theory called Partial Information Decomposition to disentangle the information about the final answer between two consecutive reasoning steps into non-negative components: unique information (in preceding steps or current step), redundant information, and synergistic information. Building on this decomposition, we propose the Step-wise Repetitive Reasoning Index (Step-RRI), a theoretically grounded measure that assesses whether the answer-relevant information in the current step SiS_i is predominantly redundant with the past steps S<iS_{<i}, relative to its unique and synergistic contributions. To evaluate the effectiveness of Step-RRI in detecting repetitiveness, we apply SLIDER to the redundancy class of the PRMBench dataset where Step-RRI improves step-level redundancy identification accuracy by over 1010 points compared to embedding-similarity and information-gain baselines. Next, we define Trajectory-RRI, an aggregate measure of repetitiveness for an individual reasoning trajectory. To demonstrate its practical relevance, we show that average Trajectory-RRI strongly correlates with actual reasoning length across QwQ-32B, DeepSeek-R1-Distill-Qwen-32B, and GPT-4.1, motivating its use as a signal for improving reasoning efficiency. Finally, we introduce Trajectory-RRI-guided data selection for fine-tuning, demonstrating that selecting training data based on Trajectory-RRI can improve a fine-tuned model's reasoning efficiency while largely preserving its task performance.
Sep 30, 2026

When a Kindergartener Solves Calculus: Measuring Capability Leakage in Role-Prompted Reasoning Models

We investigate the problem of role-capability leakage (RCL), in which a role-prompted reasoning model generates convincing in-role text while continuing to exhibit capabilities on benchmarks that exceed those implied by the assigned role. For example, when a model is prompted to assume the role of a kindergarten student, one might expect its performance on a mathematics benchmark to reflect kindergarten-level ability rather than expert-level proficiency in solving calculus problems. We introduce RoleCapBench, a curriculum-grounded benchmark for evaluating RCL across six educational roles and four assessment levels spanning elementary school through A-level, and use it to evaluate three open-weight reasoning models. We find that although the models can generate stylistically convincing in-role responses, they consistently fail to align their underlying capabilities with their assigned roles. Naive role prompting yields strong role-voice scores of 1.218--1.389 while retaining above-role accuracy of 0.811--0.898. RCL persists across a range of prompting conditions, including prompts that explicitly instruct the model to match the role's capability level. To mitigate this problem, we propose Injection, an inference-time intervention that combines explicit, role-specific capability guidelines with a guiding prefilled response prefix. Injection improves role-capability alignment across models, reducing above-role accuracy by up to 0.562 while preserving in-role accuracy with a marginal drop of less than 0.058 across most models. All artifacts, including scripts and evaluation data, will be released upon acceptance.
Sep 30, 2026

Diversity Combining for Multi-Path LLM Reasoning

Multi-path reasoning methods such as self-consistency (SC) sample KK reasoning paths and choose the most frequent answer. However, their gains quickly plateau as KK increases, and locating this saturation point currently requires fitting a scaling model to accuracies measured at several values of KK or stopping online per query. We formalize multi-path LLM reasoning as a diversity combining problem from wireless communications: each path is a noisy channel observation, and the pairwise correlation of path correctness caps the design-effect effective sample size of the vote at a finite ceiling. Generalized least squares (GLS) analysis shows that, for paths with equal gains, the optimal linear combiner of latent embeddings is uniform exactly when every row of the path covariance has the same sum, as under exchangeability, supporting majority vote as the natural default in standard SC while leaving room for weighting or pruning under heterogeneous prompt-template branches. Across 5 models and 12 benchmarks, prompt-template diversity reduces path correlation in 5555 of 6060 model-benchmark cells, with a mean relative reduction of 16%16\% and reductions in every task domain. We derive an Adaptive-K rule that uses a four-path pilot to select K∗K^*, retaining 9696--103%103\% of MV@K=32K{=}32 accuracy across Math, QA, and NLU.
Sep 30, 2026

When Reasoning Goes Astray: Attention Dynamics of Uncontrolled Reasoning

Large reasoning models (LRMs) improve performance on complex tasks through extended reasoning, yet the same process can degenerate into redundant verification and persistent generation loops. Such uncontrolled reasoning increases inference cost and creates risks of resource exhaustion and service degradation. However, existing mitigations largely truncate long outputs or react to surface repetition, and thus fail to distinguish normal thinking from uncontrolled reasoning or explain how benign reasoning degenerates into harmful behavior. In this paper, we operationalize LRM generation as four states and further introduce Reasoning-state Analysis via Dynamic Attention Responses (RADAR), which identifies the current reasoning state in real time and characterizes how effective reflection can develop into uncontrolled generation. Guided by RADAR's analysis, we further realign abnormal attention distributions toward patterns observed in normal requests and examine how this correction affects excessive reflection and persistent looping. Temporal analyses show that uncontrolled reasoning is characterized by attention distributions that deviate from normal generation, with abnormal trends becoming detectable before repetition begins. Correcting these deviations through Attention Realignment consistently reduces looping while largely preserving benign performance. Together, RADAR provide a mechanistic account of how reasoning becomes uncontrolled, offering actionable guidance for identifying critical failure stages and designing targeted runtime interventions.
Sep 29, 2026

S3S^3: Spectral Null-Space Swap Makes Reasoning Models Efficient

LLMs trained with Chain-of-thought excel in reasoning capability, but often come with excessive token cost. We find that the core of reasoning capacity lies in the Thinking model's weight component within the null space of a projection defined by the corresponding Non-thinking model's dominant singular directions, and removing the subspace component can largely improve reasoning efficiency without hurting the accuracy gained during thinking-mode post-training. Unlike existing efforts that mostly operate within the dominant subspace, we are the first to unveil the critical role of the null space and harness it for model optimization. Motivated by this finding, we propose Spectral Null-Space Swap (S3S^3), a training-free composition of paired Non-thinking and Thinking checkpoints. Our method keeps the Non-thinking model inside its own dominant subspace and takes the Thinking checkpoint outside it, improving reasoning efficiency while maintaining accuracy. We extensively evaluate S3S^3 on 2B-30B dense and mixture-of-experts (MoE) architectures spanning 28 evaluation environments across mathematical, multimodal, and audio reasoning domains. S3S^3 establishes new empirical Pareto Frontiers among training-free model composition strategies: across all settings, it reduces inference token overhead by an average of 27.4% compared to full Thinking models while simultaneously improving overall task accuracy by 1.0 percentage point (e.g., yielding +8.3% accuracy on HMMT25 alongside a 33.0% token speedup). We further use attention entropy for explanation and find that the retained component produces more concentrated attention, and we use a simplified analytical model about optimization to demonstrate why null-space can effectively reduce attention entropy, thereby improving the efficiency of reasoning.
Sep 29, 2026

Hidden Reasoning Must Leak, but Need Not Be Readable: Fundamental Opportunities and Limits for Chain-of-Thought Monitoring

Can reasoning models trick chain of thought (CoT) monitors and perform hidden computation without revealing it in their thinking traces? We show that the answer depends on the underlying task difficulty and the model size. Simple computations can be performed covertly; however, beyond a threshold depending on model size, successfully solving the task necessarily leaks a near-linear amount of information about the covert task input into the CoT. Therefore, sufficiently complex hidden computation always leaves an information-theoretic footprint. However, concerningly, this leakage need not be readable: Under plausible cryptographic assumptions, even a one-layer Transformer can encrypt its reasoning online so that no polynomial-time monitor can extract information about the hidden computation. Overall, our theoretical and empirical results provide a holistic view of both the opportunities and the limitations of CoT monitoring.
Sep 29, 2026

MetaCtrl: Your Large Language Models Can Reason Better and More Concisely with a Metacognitive Controller

Large reasoning models improve performance on challenging problems by allocating additional computation before answering, but longer reasoning does not always lead to better results and can introduce substantial redundant reasoning on simple problems. Conversely, aggressively shortening reasoning can degrade performance on difficult ones. Effective reasoning therefore requires dynamically deciding when additional computation is useful based on the reasoner's capabilities and evolving solution state. Existing approaches often rely on predefined budgets or intervention rules, retrain the target reasoner, or require additional supervision. We introduce MetaCtrl, a lightweight controller that adaptively regulates a frozen reasoner without predefined token budgets or reasoner retraining. We formulate reasoning regulation as a sequential metacognitive control problem: MetaCtrl observes the evolving reasoning trace and decides whether to continue, simplify, skip redundant steps, or conclude reasoning. It is trained directly with reinforcement learning using a reward that prioritizes correctness while favoring shorter trajectories among correct solutions, requiring neither supervised intervention trajectories nor problem-specific budgets. Across seven benchmarks spanning mathematics, science, and code, MetaCtrl consistently improves the accuracy of LRMs while reducing their reasoning length. On DeepSeek-R1-Distill-Qwen-7B, it improves average accuracy by 4.7 points while reducing generation length by 53.3%. Without further training, the same controller transfers to an unseen reasoner (e.g., Qwen3-14B), improving average accuracy by 2.9 points and reducing generation length by 50.3%. These results establish MetaCtrl as a plug-and-play controller for improving reasoning accuracy while substantially reducing inference-time generation. The code is available at https://github.com/binbin2xs/MetaCtrl.
Sep 28, 2026

Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models

Large Reasoning Models (LRMs) are commonly trained with reinforcement learning (RL) to improve their generation of chain-of-thought (CoT) reasoning before producing final answers. However, RL rewards are typically assigned based on final answers, providing little or no direct supervision over intermediate reasoning. This can lead to deceptive safety alignment, where the reasoning trace and final answer convey inconsistent safety signals. To systematically investigate this phenomenon, we introduce DSAR (Deceptive Safety Alignment Rate), a metric that jointly assesses reasoning traces and final answers to quantify their safety inconsistency. Across multiple LRMs and benchmarks, we find that deceptive safety alignment is pervasive under standard prompting conditions and is substantially amplified under prefilling attacks. We further provide a hidden representation analysis showing that models exhibit stronger safety discrimination at the final-answer stage than during intermediate reasoning. To close this gap, we propose SARA (Safety-Aware Reasoning Alignment), an RL-based method that rewards both safety-aware reasoning and safe final answers, encouraging early harmful intent recognition and enforcing reasoning-answer consistency. Experiments show that SARA significantly mitigates deceptive safety alignment under both standard and adversarial settings while preserving helpfulness and utility. Code is available at https://github.com/xzhou98/SARA.