Reasoning Distillation

Reasoning capabilities are critical for advancing Large Language Models, yet current approaches either require massive computational budgets or struggle to effectively distill reasoning to smaller models. Standard distillation methods rely on outcome-based rewards, failing to distinguish between sound reasoning and lucky guesses. We propose Collaborative Reasoning Distillation (CRD), a framework that enhances reasoning in compact models through three innovations: (1) interactive cross-feedback where teachers iteratively critique each other's reasoning, (2) fine-grained step-wise quality assessment capturing logical validity independent of final answers, and (3) coherence-aware step stitching that synthesizes complementary strengths. Students are trained via Reasoning Quality Optimization (RQO) with budget constraints. Our model, CRD-4B, achieves 97.3% on MATH-500 and 70.3% on AIME'25, surpassing baselines while using only 50K training examples, up to 12 times smaller than the datasets of comparable models.

On-policy distillation, where a student learns from a stronger teacher's feedback on its own outputs, is a common way to pass reasoning to smaller models. We analyze what it transfers at small scale, distilling Qwen3-8B into Qwen3 4B, 1.7B and 0.6B students, in thinking mode (reason at length, then end the reasoning and answer) and, for comparison, in non-thinking mode (no separate reasoning phase). Long reasoning needs two abilities, solving a problem and knowing when it is solved, and we find that distillation transfers the first, but in thinking mode not the second. Solving improves at every size, up to two ceilings, which we measure comprehensively across both modes and all student sizes: a student's single attempt never exceeds what it could already reach in many attempts before training, and the smaller the student, the further it stays below the teacher. Stopping is where the modes part. In non-thinking mode every student keeps stopping; in thinking mode students stop ending their reasoning early in training, and the smaller the student, the less of this ability survives: the teacher signals a stop almost only where a student already ends its reasoning, so distillation teaches no new stops; it only keeps the student's existing stops that land on a right answer, and a weak student has few such stops. The smallest students often reach the right value but do not commit to it: they either rarely mark it or mark it and write past it. Together, these results describe how small students behave under on-policy distillation, and a diagnostic that separates answer marking, correctness and stopping.

Large language models are increasingly expected to support diverse chemical reasoning capabilities within a unified model. One approach is to develop specialized capabilities separately and consolidate them through multi-teacher on-policy distillation, but this raises two questions: how should specialization be organized, and how should specialist guidance be integrated? We introduce ChemOPD, which addresses both. We estimate task affinities from supervised fine-tuning gradients and solve a constrained mixed-integer program(MIP) to construct partially overlapping specialist groups. During distillation, we retain a generalist teacher trained on all tasks so that specialist guidance supplements rather than replaces its supervision. Our anchor-residual objective gradually increases the routed specialist's contribution on student-generated responses. On ChemCoTBench, affinity-guided specialization produces task-dependent gains over the generalist teacher and improves several capabilities beyond semantic task grouping. Yet stronger teacher-side performance does not automatically yield stronger students: with the same specialists and routes, anchor-residual OPD improves most reported metrics over specialist-only distillation and realizes a larger share of the available teacher gains. These results highlight specialization and capability integration as connected but distinct design problems in chemical reasoning.

Reasoning distillation from powerful teacher models to smaller students faces the Gap Curse: as teachers grow more sophisticated, their complex distributions increasingly diverge from what students can approximate, causing performance degradation. Existing mitigation strategies either filter out challenging examples through data selection or introduce weaker intermediate assistant models, inherently compromising supervision coverage or quality. We propose Teacher Alignment, which directly adapts the teacher toward the student's distribution without discarding data or degrading reasoning quality. However, naive alignment through standard knowledge distillation triggers catastrophic collapse of the teacher's reasoning capabilities. To address this, we reformulate teacher alignment as reinforcement learning and introduce TeacherGRPO, built on Group Relative Policy Optimization with two key innovations: (i) Curriculum Selective Alignment applies dual token- and distribution-level curricula to focus rewards on high-signal reasoning gaps while filtering noise from trivial tokens and uncertain tail distributions, and (ii) Importance-Adaptive Length Regularization selectively penalizes verbose redundancy while preserving pedagogically critical reasoning steps. The aligned teacher then distills knowledge to students via standard pipelines. Extensive experiments show TeacherGRPO significantly outperforms baselines across diverse reasoning benchmarks and distillation methods. Our code is available at https://github.com/LzyFischer/TeacherGRPO.

Although remarkable capabilities have been demonstrated by Large Language Models (LLMs) across scientific domains, feedback controller design remains underexplored. Existing benchmarks focus mainly on linear single-Degree-of-Freedom (DoF) systems and large API-hosted models, leaving performance on complex controller-design tasks and feasibility for edge deployment unclear. To address these limitations, we introduce the Complex Dynamics-to-Control Benchmark for Large Language Models (CoDyControlBench), comprising 132 system configurations across five evaluation dimensions: number of DoF, system type, coupling level, damping regime, and controller type. Six state-of-the-art LLMs were evaluated over three independent runs, including three commercial models (GPT, Gemini, and Claude) and three open-source models (GLM, DeepSeek, and Qwen). GPT achieved the highest design success rate at 94.8%, whereas Qwen showed the lowest rate at 50.0%. Across the benchmark dimensions, DoF and controller type exhibited the largest model-averaged variations in design success, with success-rate ranges of 36.3% and 17.6%, respectively, exceeding those associated with system type, coupling level, and damping regime. Comparison of GPT and Qwen showed that their performance gap arose mainly from the control-design knowledge, particularly gain selection and the use of transient-limiting mechanisms. For edge deployment, a specialized 1.5B-parameter model was developed through reasoning distillation. The reasoning-distilled model outperformed the answer-distilled and base model on CoDyControlBench, maintained stable performance across 1-6 DoFs, and achieved successful traget tracking in all three physical trials on a pneumatic-artificial-muscle-driven robotic arm. These results establish a benchmark baseline and highlight the potential of lightweight, edge-deployable controller-design models.

Language models are widely used for generating and otherwise processing code (e.g., identifying code hallucinations, possible inputs, or predicting outputs); however, LLMs can make mistakes, which can be serious. One key issue is that models are trained on (still) largely human-written, and thus imperfect, code; it's not easy to find sufficiently large code corpora that are entirely free of bugs. Thus, other inference-time ways of reducing LLM errors, without additional training, are desirable. "Reasoning" or "thinking" modes, exposed as a togglable feature by hybrid reasoning models, do reduce errors; however, reasoning consumes additional resources. This paper asks if better performance can be achieved without always incurring the cost of reasoning. Human students of programming learn to avoid mistakes by (a) identifying them, (b) reflecting upon the cognitive lapses that led to them (essentially, "thinking through" the errors), (c) inferring general rules or lessons from these reflections, and (d) internalizing these lessons into rules. In tutorial sessions with an instructor, this is a common Socratic interaction. Examples of such internalizable rules might include the nugget "Before coding, restate the requirements to clarify them." Inspired by this process, this paper describes an approach where we first identify examples in which "thinking mode" in a (low-resource) LLM avoids errors. These errors, and their avoidance via "thinking" in the same LLM, are then examined by a bigger LLM to generate summary explanations; these are then summarized by a large LLM into brief advisory prompts. This approach works on many modest-sized models; in some cases, the "advisory prompts" thus learned can also be gainfully transferred to other models. We also present investigations into the nature of coding errors that language models make, and a characterization of when this approach can be helpful.

Large language models deployed in open-world applications require safety guardrails that are both robust to complex risks and efficient enough for low-latency runtime moderation. Existing guardrails face a practical trade-off between lightweight classification-based models, which are efficient but often struggle with concealed intent, ambiguous semantics, and borderline safety decisions, and reasoning-based guards, which improve judgment quality but introduce additional token generation and inference latency. We present DT-Guard, a content safety guardrail model based on a Reasoning-Active Training, Reasoning-Free Inference paradigm. The key idea is to use reasoning supervision during training while emitting only structured safety labels at inference time. DT-Guard formulates safety judgment as a progressive decision process, Intent - Category - Safety, and constructs an intent-driven dataset with intent labels, risk categories, safety labels, and structured reasoning trajectories. To further improve hard-case robustness, we propose Rollout-Guided Progressive Hard-Case Optimization (RG-PHO), which uses multi-rollout consistency to identify stably mastered, persistently failed, and preference-unstable samples, and applies targeted supervised and preference optimization accordingly. At inference time, DT-Guard directly generates structured labels without explicit reasoning traces, preserving deployment efficiency. Experiments on prompt-side and response-side safety benchmarks show that DT-Guard achieves average F1 scores of 0.886 and 0.870, respectively. With only a 4B backbone, it reaches a dual-side average F1 of 0.878, outperforming strong 8B guardrail baselines. These results demonstrate that reasoning supervision can be effectively internalized into low-latency safety discrimination.

Reward models are central to text-to-image post-training, but visual preference is subjective and better represented as a distribution over rubric scores than as a deterministic scalar. Existing scalar, score-token, and pairwise reward models over-compress uncertainty and fine-grained score differences, while reasoning-based generative rewards provide stronger judgments but are costly to deploy and difficult to use as direct optimization signals. We propose Z-Reward, a teacher-student reward modeling framework that decouples reasoning-heavy judgment from efficient reward deployment. The teacher is a large VLM that uses reasoning to infer rubric-aligned score distributions, and is trained with Group-wise Direct Score Optimization (GDSO), which combines policy-gradient rewards from distribution expectations with direct pointwise and pairwise supervision on score distributions and score gaps. The student is trained with Reasoning-Internalized Score Distillation (RISD), which transfers the teacher's reasoning-conditioned score distribution into a compact VLM without requiring explicit reasoning chains at inference time. On our internally annotated evaluation set, the 27B GDSO teacher reaches 89.6% human preference accuracy, outperforming SFT, RewardDance, and GRPO, while the 9B RISD student reaches 88.6%, outperforming the OPD baseline and closely matching the larger teacher. We further show that Z-Reward can serve as a differentiable reward signal for text-to-image optimization, yielding a 41.3% net human-preference improvement over the SFT baseline.

Reasoning distillation transfers complex reasoning abilities from large language models (LLMs) to smaller ones, yet its success depends on how well the training data align with the student model. This paper introduces the Data-Model Compatibility (DMC) metric, which can be used to assess the suitability of a dataset for reasoning distillation on a student model. DMC provides an assessment by jointly considering data quality, relative difficulty, and student capability. We validated the effectiveness of DMC from two perspectives: (1) DMC exhibits a strong correlation with reasoning distillation performance; and (2) using DMC as the criterion for data selection leads to improved reasoning distillation performance. Both findings are consistently demonstrated across multiple student models and tasks. Moreover, since the DMC of each dataset dynamically changes during training, our experiments demonstrate that dynamically selecting datasets based on DMC can further enhance performance.

Knowledge distillation is central to LLM post-training, yet its design space remains poorly understood, especially alongside reinforcement learning (RL). We show that the prevailing paradigms, off-policy distillation and on-policy distillation (OPD), implicitly couple two orthogonal choices: prefix source and token-level KL direction. This follows from decomposing sequence-level KL over autoregressive response distributions: forward KL pairs teacher prefixes with token-level forward KL, and reverse KL pairs student prefixes with token-level reverse KL. We argue this coupling is not intrinsic: decoupling the two axes yields four valid objectives. We establish gradient-level identities showing forward KL gives SFT-style cross-entropy matching with teacher soft targets, whereas reverse KL gives an RL-style policy-gradient objective with a dense teacher-student log-ratio reward, connecting them to off-policy SFT, DAgger-style on-policy SFT, offline-RL-style distillation, and OPD. We conduct an extensive controlled study on math reasoning, evaluating the four objectives both as standalone methods and as initializations for subsequent RL. The results reveal three tradeoffs: KL direction induces an accuracy-entropy tradeoff, prefix source a quality-compute tradeoff, and training length an accuracy-stability tradeoff. Motivated by these findings, we propose KL mixing and an entropy-gated length curriculum. KL mixing shows long-sequence distillation requires substantial forward-KL weight to prevent entropy collapse and length inflation without sacrificing accuracy. The entropy-gated length curriculum improves Avg@k and Pass@k by 3.6 and up to 5.8 points, and cuts average response length by roughly 3x versus fixed long-horizon training. Our results provide a framework and practical methods for designing reasoning distillation objectives that balance accuracy, diversity, compute, and RL behavior.

Knowledge distillation is a key technique for transferring the capabilities of large language models (LLMs) into smaller, more efficient student models. Existing distillation approaches often overlook two critical factors: the learning order of training data and the capacity mismatch between teacher and student models. This oversight limits distillation performance, as manifested by the counter-intuitive phenomenon where stronger teachers fail to produce better students. In this work, we propose Curriculum Learning-Guided Progressive Distillation (CLPD), a unified framework that explicitly accounts for both factors by aligning data difficulty with teacher strength. CLPD constructs an explicit curriculum by organizing training examples from easy to hard, while simultaneously applying an implicit curriculum over supervision signals by progressively scheduling teachers of increasing capacity. Our framework is modular and can be integrated into standard distillation algorithms with minimal overhead. Empirical results on the reasoning benchmarks demonstrate that CLPD consistently outperforms standard distillation, data ordering alone, and teacher scheduling alone across multiple settings. These findings highlight the importance of jointly considering data ordering and teacher capacity when distilling reasoning abilities into small language models.

Recent years have witnessed growing interest in applying Large Reasoning Models (LRMs) to Machine Translation (MT). Existing approaches predominantly adopt a "think-first-then-translate" paradigm. Although explicit reasoning trajectories significantly enhance translation quality, they incur prohibitive inference costs and latency. To address these limitations, we propose ReflectMT, a two-stage reflection internalization algorithm for machine translation that employs a "translate-first-think-later" paradigm. Our approach develops the model's "translate-reflect-refine" capability through reinforcement learning. In the first stage, we cultivate the model's capacity for high-quality reflection and refinement, thereby enhancing its semantic comprehension and task-specific knowledge. In the second stage, we train the model to internalize the knowledge acquired during reflection. As a result, during inference, ReflectMT operates in a direct translation mode, producing high-quality translations on the first attempt without any explicit reasoning steps. Experimental results on datasets such as WMT24 demonstrate that our model's first-pass translations during inference outperform multi-step reasoning LRMs such as DeepSeek-R1 in both automatic metrics and GPT-based evaluation, achieving a 2.16-point improvement in GPT-based translation quality evaluation while reducing token consumption by 94.33%.

Chain-of-thought (CoT) distillation transfers reasoning behaviors from a strong teacher to a smaller student, but prior work reports a capacity gap: distillation may fail when the teacher-student capability mismatch is large. We revisit the capacity gap from a practical perspective by re-examining commonly used experimental settings. Notably, we find that CoT distillation often degrades performance compared to the student's pre-distillation baseline, and that some settings used in prior work, while suitable for establishing the capacity gap as a phenomenon, do not reflect realistic deployment scenarios. Complementing prior work that establishes the capacity gap, we evaluate its practical impact under more realistic settings and find that it does not consistently dominate; stronger teachers tend to be preferable when candidate teachers differ substantially in performance. Our results offer practical guidance for selecting teacher-student pairs in CoT distillation.

Large language models are increasingly deployed in high-stakes domains, where confident yet incorrect inferences may cause severe real-world harm, bringing the long-overlooked issue of confidence faithfulness to the forefront. A promising solution jointly optimizes unsupervised Reinforcement Learning from Internal Feedback (RLIF) with reasoning-trace-guided Reasoning Distillation (RD), yet it faces three persistent challenges, namely the scarcity of high-quality training corpora, factually unwarranted overconfidence, and erroneous updates amplified by indiscriminate fusion. Inspired by how human confidence accumulates from uncertainty to certainty, we propose Progressive Reasoning Gain (PRG) to measure whether reasoning steps progressively strengthen confidence in the final answer. Building on PRG, we introduce HyTuning, a hybrid post-training framework that adaptively reweights RD and RLIF, using scarce supervised reasoning traces as a stable anchor while exploiting abundant unlabeled queries for scalability. Experiments on several domain-specific and general benchmarks demonstrate that HyTuning improves accuracy while achieving confidence faithfulness under limited supervision, supporting a practical ``Less Data Approximates More'' effect. Our code will be released upon acceptance.