Knowledge Distillation

Knowledge distillation can transfer reasoning from stronger teachers to frozen students through reusable prompts, but avoiding weight updates does not eliminate supervision. Without ground-truth answers, teacher solutions are unverified, and agreement with the teacher can reward shared mistakes. We introduce Knowledge-to-Prompt (K2P) for label-free knowledge distillation to prompts. K2P synthesizes reusable instructions from teacher solutions, refines them using paired teacher and student responses, and guides search and selection with answer agreement. It retains candidates that adaptive search may undervalue and selects on reserved questions. Deployment uses only the frozen student and selected prompt. Our theory separates generation and selection gaps and gives conditions under which agreement-guided construction yields accuracy guarantees despite imperfect teacher references. Across reasoning tasks and students, K2P outperforms label-free alternatives overall and remains competitive with supervised prompt optimization. Ablations and archive diagnostics assess the contributions of teacher solutions and refinement, while revealing the limits of agreement-guided selection.

Audio is increasingly used for human activity recognition (HAR) because it captures object interactions, environmental events, and contextual cues in everyday environments. High-resolution (HR) audio provides rich acoustic information for model development but incurs substantial energy and storage costs and may expose sensitive speech content. Low-resolution (LR) audio offers a more privacy-preserving and resource-efficient alternative for deployment, but reduced sampling rates can remove acoustic cues essential for activity recognition, leading to significant performance degradation. We formulate this training-deployment mismatch as sensor-resolution privileged learning, in which HR audio is available during training, while inference relies exclusively on LR audio. We propose RAST, a resolution-aware transfer framework that compresses HR teacher representations by preserving token-level information and neighborhood structure before performing localized HR-LR alignment. Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms LR-only training and direct teacher-transfer baselines, improving LR-only recognition by up to approximately 7.8% while requiring only LR audio at inference.

On-policy distillation (OPD) learns from teacher feedback on student-generated responses and has shown promise in reducing forgetting relative to supervised fine-tuning (SFT). However, its benefits and fragility remain incompletely understood. We study sequential distillation from multiple teachers, where the student minimizes its average divergence from the teachers. Forward Kullback--Leibler (KL) divergence yields a weighted arithmetic mixture, while reverse KL yields a normalized weighted geometric aggregate. We develop algorithms that learn these targets under off-policy and on-policy feedback, respectively, establishing logarithmic regret bounds in the tabular setting and extending the analysis to function approximation. By analyzing these aggregation targets, we identify mechanisms that help explain both the benefits and fragility of OPD. Relative to forward KL, reverse KL can better retain a confident expert's preferences under uninformative feedback, but is more sensitive to teachers that assign very low probabilities to correct responses. Its token-level conditionals also reveal a dependence on continuation distributions that can favor incorrect prefixes over long horizons.

On-policy distillation (OPD) has recently emerged as a promising post-training paradigm in which the student learns from trajectories generated by its own policy under dense teacher supervision. However, OPD introduces a fundamental asymmetry: although the sampled trajectories are on-policy for the student, they are off-policy for the teacher. The teacher is typically optimized to continue from prefixes generated by its own policy, but during OPD it must instead supervise prefixes generated by the student. Empirically, we find that its continuation performance degrades as these prefixes grow longer. To address this issue, we propose Student-COnditioned Updates of the Teacher (SCOUT), a co-training framework that adapts the teacher to student-generated prefixes. Alongside standard OPD updates, SCOUT periodically optimizes the teacher's conditional ability using reinforcement learning with verifiable rewards, where the teacher generates continuations from student prefixes and learns from outcome rewards. Controlled experiments show that SCOUT improves the teacher's ability to continue from student-generated prefixes, supporting the intended mechanism of student-conditioned teacher adaptation. Across multiple teacher--student configurations, model scales, and reasoning domains, SCOUT also consistently improves the effectiveness of on-policy distillation.

Behavior Foundation Models (BFMs) give humanoids a promptable policy over a latent behavior space, enabling one single vector to represent a motion to imitate, a pose to reach, or a reward to maximize. Forward-Backward representations successfully produce such spaces, but at the cost of hundreds of GPU-hours for a single robot. Moreover, when the training process is repeated for a second robot, it produces a second space unrelated to the first, resulting in embodiment-specific latents that do not unify or transfer. We address these problems with CrossBFM, treating the latent space as the transferable asset for various embodiments. As retargeting provides frame-level cross-embodiment correspondence, we propose a unified encoder architecture with no robot-specific parameters for distilling the behavior space to address all training embodiments simultaneously in less than a GPU-hour. Following this encoder, latent-conditioned trackers turn the distilled latent into whole-body control in a conventional PPO training manner in just 10 more GPU-hours. On three distilled humanoids, all three prompting modes transfer: motion tracking with latent-conditioned policy losing only $0.025$ rad to its joint-conditioned counterpart, smooth goal reaching between poses with no falls, and reward optimization for all $41$ reward prompts. Our experiments further reveal that 1) regressing the encoder on a quarter of the motion corpus costs only $5\%$ of tracking performance and 2) training the encoder on a subset of robots and evaluating on an unseen one recovers up to $89\%$ of the tracking performance of seen robots, demonstrating cross-embodiment generalization to morphologically similar robots. We also verify the pipeline on real robots across all three prompting modes and with flow-based generated latents. Project website: https://dotandung.github.io/crossbfm/

On-policy self-distillation (OPSD) trains a student to match a privileged teacher distribution along its own sampled trajectory. Standard OPSD applies this supervision to unverified student rollouts while conditioning the teacher on privileged context, typically a reference solution. We separate these roles in a factorial analysis and find that scaffold correctness has a stronger effect on downstream accuracy than context correctness. Unverified scaffolds create an imitation gap because the teacher can use information unavailable to the student. This gap shrinks with model scale, yet OPSD continues to supervise mostly unverified trajectories. In contrast, verified scaffolds remain effective even when the teacher is conditioned on the student's own unsuccessful rollout. Based on this finding, we introduce OASIS, which retains the OPSD objective but supervises mostly verified by label on-policy trajectories and replaces written solutions with unverified model-generated attempts as the teacher context. OASIS therefore requires only final-answer labels. Across Qwen3-1.7B, 4B, and 8B on AIME 2024, AIME 2025, and HMMT 2025, OASIS improves over the base model by 3.2--3.8 points on average, while OPSD's gain falls from 3.05 points at 1.7B to 0.14 at 8B. At 8B, OASIS improves over OPSD by 3.05 points, showing that verified on-policy scaffolds preserve the effectiveness of self-distillation as models scale.

Off-policy and on-policy distillation have traditionally been formulated as separate paradigms, each favoring a different property of distillation trajectories. Teacher-generated (off-policy) traces are typically high-quality but lie far from the student's distribution, whereas student-generated (on-policy) rollouts are more learnable but often contain erroneous reasoning. We view these paradigms as the endpoints of a policy continuum and posit that a more effective rollout policy may lie in between. We introduce Interpolated Policy Distillation (IPD), which defines the next-token distribution at every decoding step as an explicit linear interpolation between the student and teacher distributions. The interpolation operates at the distribution level, token by token, and its coefficient provides direct control over the balance between trajectory quality and student learnability. Naively sampling from this policy would require sequentially querying the teacher at every token and is thus expensive. To make IPD practical, we accelerate it with a new speculative-decoding rule while exactly preserving the interpolated next-token distribution.At the trajectory level, the resulting rollouts naturally interleave student- and teacher-generated segments. Unlike recent heuristic segment-interleaving methods, however, this interleaving is induced by an exactly realized token-level interpolated policy rather than by hand-designed switching rules. Across text-only and multimodal reasoning benchmarks, IPD consistently outperforms both endpoint policies (SFT and OPD), their conventional two-stage combination (SFT-then-OPD), and recent heuristic segment-interleaving methods, demonstrating that token-level policy interpolation better balances trajectory quality and student learnability.

Recent event-based depth estimation methods successfully transfer geometric priors from vision foundation models via cross-modal distillation. However, their reliance on synchronized RGB-event pairs or depth annotations during training severely restricts practical deployment. To overcome this bottleneck, we propose SFE-VGGT, a novel source-free framework that distills the geometric priors of VGGT to the event domain without any paired RGB observations. Our core idea is to reconstruct surrogate frames directly from the target event stream to act as a frozen geometric teacher, entirely eliminating the need for genuine source RGB data. Crucially, as these surrogate frames inherently yield imperfect and spatially varying supervision, directly distilling from them propagates artifacts. To resolve this, we introduce a novel reliability-aware distillation strategy. This includes Density-Aware Feature Distillation to emphasize informative event regions, and Confidence-Weighted Depth Distillation to dynamically regulate supervision based on relative teacher-student prediction confidence. Meanwhile, we propose a Cross-Frame Relational Consistency loss that enforces temporal geometric stability using reliable inter-frame correspondences, bypassing the need for temporally consistent teacher's depth. Extensive experiments demonstrate that, despite source-free, our SFE-VGGT closely matches the accuracy of RGB-dependent baselines under standard conditions and significantly surpasses them in challenging nighttime scenarios. Across MVSEC nighttime sequences, SFE-VGGT reduces the average 10 m depth error by 15.3% compared with EventVGGT. Moreover, our method exhibits robust zero-shot generalization across real-world datasets, proving that highly effective geometric priors can be transferred to event cameras using strictly source-free supervision.

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for improving large language models (LLMs) on various tasks, yet its sparse outcome rewards lack token-level credit assignment for intermediate steps. To address this, on-policy self-distillation (OPSD) leverages a self-teacher with privileged context to provide additional dense learning signals. However, because the self-teacher is often overconfident and imposes excessive penalties on long reasoning trajectories, OPSD frequently struggles in practice. To mitigate this, we propose self-instructing policy optimization (SIPO) with a contrastive self-teacher to provide dense credit. At each iteration, SIPO samples multiple rollouts per prompt from the current policy, scores them with environment rewards, and constructs two teacher contexts for each rollout by pairing the reference answer with mistakes made within the group. The model then re-evaluates its own responses under both contexts, using the difference between the two teacher log-probabilities as token-level feedback, so that biases shared by both contexts are expected to largely cancel. The resulting objective yields a token-level advantage for every rollout: the reward still sets the main direction of each update while the self-teacher redistributes credit across tokens. Even in groups where every rollout fails and group-relative advantages vanish, SIPO still provides a learning signal. By preserving direct optimization of the task reward while providing dense, token-level feedback, this approach bridges reinforcement learning and on-policy self-distillation. Extensive experiments across multiple reasoning and code-generation benchmarks demonstrate that SIPO outperforms both RLVR and OPSD baselines without an external teacher or additional generation.

Self-distillation turns knowledge distillation into a closed learning loop and offers a path toward recursive self-improvement. Without an external teacher, however, the model must determine both what information can improve its supervision and which induced changes should be learned. Existing methods typically improve teacher-generated data or select training examples in isolation, leaving the information transferred between these stages unmeasured. We introduce InFlow, a retrieval-guided on-policy self-distillation framework that models this process as potential-to-realized information flow. InFlow first retrieves potentially informative sources using certainty-calibrated hidden-state trajectories, then measures their realized effect through the Jensen--Shannon divergence between the teacher's initial and retrieval-conditioned answer beliefs. Examples with larger belief shifts are selected for on-policy distillation. Our analysis formalizes the information optimized by retrieval and selection and relates the answer-level shift to the teacher--student distillation gap. Across four open-weight language models and three knowledge domains, InFlow achieves the strongest cross-model average among the compared selection methods, with ablations supporting both stages of the framework. Our code is available at https://github.com/1240148048/INFLOW.

Language-model agents are usually trained by reinforcement learning from one reward per episode, and privileged self-distillation enriches it by letting the same policy, given a skill, teach its skill-free self through token probabilities. However, we identify two phenomena that question this channel. Invisible Advantage: a skill in context lifts WebShop success from 42.2% to 56.2%, yet changes the probabilities of fewer than a quarter of the sampled tokens. Much to Align: a skill changes the hidden states of over 80% of response tokens, in a way that linear probes can trace back to the specific skill. To exploit this, we propose Privileged Representation On-policy Self-Distillation (PR-OPD). After a GRPO warm start, the policy writes a hindsight skill for each trajectory, re-reads its own responses with that skill as a stop-gradient teacher, and aligns its projected hidden states to the teacher's at every layer alongside the reward objective, with no external skill library, separate teacher, or inference overhead. On ALFWorld and WebShop with two backbones, PR-OPD achieves the best overall results in every setting, improving over GRPO by up to 4.7 points in ALFWorld success and 14.0 points in WebShop accuracy. Code is available at https://github.com/balibata/PR-OPD.

Text-to-image users often provide concise and underspecified prompts, whereas generative models benefit from detailed textual conditions for reliable instruction following. Existing systems bridge this gap with Prompt Enhancers (PEs) that rewrite raw prompts at inference time, introducing additional latency and leaving prompt elaboration external to the generator. We instead view enhanced prompts as privileged training information and ask whether their benefits can be internalized. We propose Prompt-Enhanced On-Policy Self-Distillation (PE-OPSD) for text-to-image flow-matching models. During training, a raw-prompt student follows its own generation trajectory, while an enhanced-prompt teacher provides vector-field targets at the states visited by the student. This on-policy supervision distills the behavior induced by enhanced prompts into the raw-prompt student without requiring additional text--image pairs. At inference, both the PE and teacher are removed, and the student generates directly from raw prompts. Across multiple model families, PEs, and benchmarks, PE-OPSD achieves the strongest aggregate prompt fidelity among the evaluated baselines, yields positive aggregate visual appeal gains, and retains the base-model inference efficiency.

Modern agents increasingly rely on memories, tools, and execution logic, so their competence extends beyond model parameters. This shift exposes a limitation of conventional knowledge distillation, which asks how a student model imitates a teacher model. We define Agent Distillation as the persistent transfer of task-solving knowledge from a teacher agent to a student agent. Our study organizes the field by where transferred knowledge is retained: within the model, as artifacts, through the execution harness, or across substrates. This perspective separates transfer evidence from its outcome and clarifies how knowledge moves between agent components. We develop an evaluation framework that relates retention to causal contribution and deployed utility. Together, these contributions establish a foundation for the reliable, maintainable, and safe development of increasingly complex agentic systems.

On-policy distillation (OPD) reduces train-test state mismatch by training a student on its own generated trajectories, but weak students may visit teacher-misaligned prefixes where supervision is less representative. We introduce SAKI (Supervision Allocation with KL-constrained Interpolation), which combines a KL-constrained teacher-guided rollout with maximal coupling and reuses realized accept/correction events to route token-level supervision. Accepted positions retain sampled-token reverse-KL supervision, while correction positions receive direct supervision on the teacher's highest-probability token. Under maximal coupling, the correction probability is exactly TV(p_t, q_t), so the same trust-region radius controls rollout deviation and upper-bounds intervention and specialized-supervision frequency. We further implement an engine-resident speculative verifier that preserves the exact-q trajectory distribution and coupling semantics while improving matched-workload rollout throughput by 4.22x. Across seven mathematical reasoning benchmarks, SAKI improves the matched teacher-guided baseline in Mean@8 and Pass@8 for both 1.7B and 0.6B students. Placement controls and fixed-prefix analysis further support correction-triggered routing as a conflict-adaptive supervision signal.

Vision-Language-Action (VLA) models increasingly rely on action experts that generate short action chunks under receding-horizon control. While chunk-level training is convenient across robot embodiments, it optimizes local action likelihood without explicitly accounting for long-horizon task success. Sequence-level reinforcement learning can address this limitation, but typically requires policy rollouts and closed-loop interaction, which are costly for real-robot manipulation. We introduce DriftOPD, a teacher-free, rollout-free framework for sequence-level on-policy distillation of continuous VLA action experts. We show that the sequence-level reverse Kullback-Leibler (KL) divergence decomposes into a chunk-level reverse-KL term and a future-potential term that captures the long-horizon effect of the current action. DriftOPD optimizes these two terms using a one-step drifting objective and a Q-function critic learned from offline demonstrations, respectively, enabling sequence-level optimization with only offline data and one-step action generation. Across multiple VLA architectures in simulation and real-world manipulation, DriftOPD generally outperforms existing one-step distillation baselines while achieving task success performance comparable to multi-step teacher policies. These results demonstrate that long-horizon behavior can be effectively distilled into one-step VLA action experts without online interaction or a separate teacher.

Cross-resolution knowledge distillation aims to improve low-magnification whole- slide analysis by transferring high-magnification representations, yet the conditions for useful transfer remain unclear. We develop a decomposition-based analysis of teacher access, representation loss, and model excess, motivating three questions: whether (a) teacher targets help the task, (b) low-magnification students can predict them, and (c) slide models benefit from those predictions. We investigate them through controlled experiments across ten pathology cohorts spanning classifi- cation, grading, and survival prediction. In the main comparison, providing teacher regional means alongside native low-magnification features improves downstream performance in all ten cohorts. Direct prediction achieves lower reconstruction error than residual prediction, yet the predicted features underrepresent variation in the teacher targets. Moreover, better reconstruction does not consistently improve downstream scores, and retaining native features changes performance even when the predicted teacher features are held fixed. Together, these findings expose a gap between reconstructing teacher representations and realizing their downstream value. They challenge the sufficiency of reconstruction error as a measure of cross-resolution transfer and provide a diagnostic framework for examining where that transfer breaks down. Future distillation designs must account for both what students can predict and how slide models use those predictions.

We present OLIVE (OnLine InterVEntion). At each iteration, the evolving student policy generates a new prefix, the teacher continues it autoregressively, and the student is updated using cross-entropy computed on the teacher-generated tokens. Each design choice targets a corresponding limitation of existing distillation methods: (1) sequential covariate shift in offline supervised fine-tuning (SFT) on fixed teacher trajectories, (2) fragmented supervision under prefix failure in token-level on-policy distillation (OPD), and (3) the need for access to teacher token probabilities in distribution-matching distillation. OLIVE achieves higher reasoning performance than OPD (with a top-16 KL approximation) at comparable GPU-hour cost. Our asynchronous implementation further reduces OLIVE's total training time by 23.8%. We evaluate OLIVE on both hard reasoning tasks and agentic tasks which reflects modern post-training scenarios, and it consistently outperforms existing distillation methods under the same training budget. By regenerating prefixes from the evolving student, OLIVE continues improving after offline distillation plateaus while better preserving the general capabilities and plasticity of the student. Using only text from GPT-5.4-mini, continuously training with OLIVE outperforms offline SFT from the same teacher by 13% on ScienceWorld. These results support OLIVE as an effective and efficient approach to online language-model distillation.

Tampered Text Detection (TTD) is essential for safeguarding document authenticity in security-critical workflows. Existing expert models are effective at capturing subtle manipulation traces but often generalize poorly across diverse document domains, while Multimodal Large Language Models (MLLMs) offer stronger semantic understanding and transferability yet remain insensitive to fine-grained forensic artifacts. This complementarity motivates us to investigate how expert forensic perception can be internalized into an MLLM rather than merely accessed through an external module. We identify a fundamental Double Mismatch that hinders this goal: a Spatial Precision Mismatch between coarse visual tokens and tiny tampered regions, and a Perceptual Granularity Mismatch between semantics-oriented pre-training and low-level forensic perception. To address these challenges, we propose Expert Knowledge Internalization (EKI), a progressive two-stage framework that transfers forensic expertise into the MLLM itself. In Stage 1, Text-Focused and Image-Focused strategies establish precise spatial focus on small text regions. In Stage 2, the proposed Forensic-General Representation Alignment (FGRA) loss aligns shallow LLM representations with those of a pre-trained forensic expert, enabling the model to acquire fine-grained artifact perception before such cues are diluted by deeper semantic abstraction. Extensive experiments on multiple in-domain and cross-domain benchmarks demonstrate that EKI achieves state-of-the-art performance and stronger generalization than existing expert-model-based and MLLM-based methods. Moreover, the expert is required only during training, allowing the resulting MLLM to maintain inference efficiency nearly identical to the vanilla model without relying on any external expert at inference.

Linear Vision Transformers (ViTs) are designed to replace the attention in Softmax ViTs with the linear-complexity attention operator for more efficient token routing, but they require from-scratch pre-training and typically underperform the original Softmax version. How to initialize linear ViTs both efficiently and effectively still remains unclear. In this work, we explicitly ask: given that most foundation ViTs are built on the mainstream Softmax attention, can linear ViTs benefit from their pre-trained weights? Recent works on Attention Transfer show that attention is the effective transferable component between Softmax ViTs, suggesting attention alone suffices for such reuse. However, we find the opposite for Softmax-to-linear transfer. The attention weights are operator-specific: copying them barely helps, and is sometimes even worse than random initialization. Instead, the attention's token routing behavior can be recovered through distillation with a proper loss design, letting linear ViTs reduce the gap and even match Softmax ones. In contrast, the MLP weights, which carry the learned representation, are operator-agnostic: they can be transferred by simple direct copying, which already carries most of the benefit of the pre-trained weights. Thus, copying MLPs can serve as an effective foundation for Softmax-to-linear transfer: paired with the distilled attention, linear ViTs eventually close the remaining gap and even surpass Softmax ones. These findings hold consistently across various linear ViT variants, different model sizes, and diverse datasets. We hope this study deepens the understanding of reusing pre-trained weights across attention operators: copy what stays the same and distill what differs, to recover the benefit across the Softmax-to-linear boundary.

On-policy distillation (OPD) trains language models using dense token-level teacher supervision on student-generated trajectories. However, its reliance on frequently refreshed student rollouts often incurs substantial generation cost. We introduce REVO, an off-policy distillation framework that improves rollout efficiency by reusing each student rollout for multi-step learner updates. REVO addresses prefix-level and current-token policy mismatch through stabilized prefix weighting and one-step resampling from the current student, which enables repeated updates without regenerating full trajectories. To prioritize informative token positions within reused rollouts, REVO uses the variance of the student-teacher log-probability ratio to quantify the remaining token-level learning signal and guide repeated optimization. Across multiple student-teacher scales, REVO with only 50 rollout iterations matches or exceeds OPD baselines trained for 200 iterations on both in-domain and cross-domain reasoning benchmarks.

On-policy distillation (OPD) trains a student language model with dense feedback from a stronger teacher on student-generated trajectories. Yet standard OPD weights token-level distillation terms uniformly, implicitly treating local teacher preference as a proxy for correction utility. A decision's task value, however, depends on how the student completes the subsequent reasoning. This mismatch can cause imitation to suppress viable student strategies or reinforce paths the student cannot reliably execute. Verified trajectory outcomes provide complementary evidence about continuation quality, but do not directly identify the utility of individual decisions. We introduce Reward-Aligned Reweighting for On-Policy Distillation (R$^{2}$-OPD), which uses outcome agreement and the magnitude of teacher--student disagreement to continuously reallocate teacher supervision. It gives reward-aligned corrections greater relative influence while retaining dense feedback, moving beyond uniform imitation and hard filtering. Our analysis formalizes the mismatch between local teacher preference and student continuation value and establishes sufficient conditions for reallocation to improve first-order task progress over uniform OPD. Across seven mathematical reasoning benchmarks, R$^{2}$-OPD achieves the highest average accuracy among the compared training methods in both cross-size and same-size distillation. It outperforms standard OPD on all seven benchmarks, with average gains of 3.5 and 2.4 percentage points for 1.7B and 4B students, respectively. An extension to code generation yields an average gain of 1.6 percentage points over standard OPD. These results highlight outcome-guided supervision allocation as an effective way to translate dense teacher feedback into stronger student performance across model scales and task domains.

We study on-policy distillation (OPD) through the lens of reinforcement learning, establishing a connection between the reverse-KL objective in OPD and KL-regularized policy optimization. Building on this connection, we introduce Least-Square Policy Distillation (LSPD), an RL-inspired framework that brings optimistic exploration and off-policy data reuse from value-based RL into policy distillation. LSPD preserves policy diversity through exploration while improving rollout efficiency by repeatedly learning from previously collected trajectories. Our theoretical analysis connects LSPD to optimistic value-based learning and shows that its idealized formulation achieves a sharp $\tilde{\mathcal O}(\log K)$ regret bound under online exploration. Empirically, LSPD consistently outperforms existing distillation baselines across six mathematical reasoning benchmarks and diverse teacher-student settings, with average gains of +1.59 points in Avg@16. Remarkably, through Pass@k evaluations up to k=64, we found that LSPD better preserves policy diversity by achieving stronger performance as k grows. Its fully off-policy variant achieves comparable performance to vanilla OPD using only the first 25% of rollout batches. Together, these results provide an RL perspective on OPD that offers both a principled interpretation and a practical route toward more effective and rollout-efficient language model distillation.

Reinforcement learning can turn one language model into several specialists, each excellent at a single skill such as mathematics, coding or following instructions, but users need one model with all of these skills. Multi-teacher on-policy distillation (MOPD) merges them by letting the specialists teach one student: the student answers each prompt, and the specialist for that prompt's domain gives feedback on every token. This routing decides which specialist teaches, but not how strongly its feedback moves the shared student. In Qwen3.5 models at three sizes, we find that MOPD's student does not beat one taught by the best single specialist and gains little of the mathematics specialist's advantage. The feedback is unbalanced: instruction-following feedback is several times more spread out than mathematics feedback and dominates the student's updates. We propose Domain-Normalized MOPD (DN-MOPD), which keeps the routing and rescales each domain's feedback by its measured spread. On six public benchmarks, DN-MOPD improves the average score over MOPD at every size, across three random seeds and under two answer-length limits, and recovers most of the lost mathematics gain. Controls with fixed domain weights show that the gain comes mainly from turning down instruction-following feedback rather than turning up mathematics alone, and that fixed weights close to those DN-MOPD measures perform comparably. Combining specialists therefore requires deciding not only which one teaches, but also how strongly its feedback counts.

On-policy learning has been argued to reduce catastrophic forgetting, produce sparser parameter updates, and improve generalisation. However, existing comparisons between supervised fine-tuning and reinforcement learning vary many factors simultaneously, making the contribution of rollout policy difficult to isolate. We study the effect of rollout policy in a controlled strong-to-weak distillation setting, by independently varying rollout policy, token-level KL direction, and learning rate across the Llama3 and Qwen2.5 model families and reasoning tasks spanning scientific, medical, and arithmetic domains. Our analysis reveals a nuanced picture of distillation dynamics in which rollout policy does not necessarily play a central role. Instead, token-level KL direction more clearly shapes task performance and output coverage, while learning rate governs forgetting and update sparsity. Analysis of KL gradients and experiments along a continuous student-teacher rollout-policy spectrum explain this pattern: forward KL is remarkably robust to rollout policy, with its performance stable and strong despite changes to the rollout policy, whereas reverse KL is substantially more sensitive and favours student-generated rollouts. On-policy data nevertheless improves generalisation to harder variants of the Countdown arithmetic task under both KL directions, although this advantage does not reliably persist after subsequent RLVR. Our broader conclusions remain robust to removing gradient clipping, using sampled KL estimators, and training on tasks requiring longer reasoning chains. Overall, our results challenge the view that on-policy rollouts are inherently preferable and show that their value depends critically on the objective, evaluation setting, and optimisation hyperparameters.

On-policy distillation (OPD) is a widely adopted post-training technique for LLM reasoning. It is commonly believed to transfer knowledge from a stronger teacher, yet what OPD actually distills into the student's internal representations remains unclear. We study this question with sparse crosscoders, which learn one feature dictionary shared by the student before and after OPD and the teacher. Standard crosscoder analyses, however, identify model-specific features but cannot tell how a model's use of its features changes, since all models are encoded into one set of feature activations. We therefore propose the swap readout, which reads each student checkpoint's feature activations on its own, measuring how training changes the student's use of each feature, even for checkpoints unseen by the crosscoder. Across three OPD settings, we find that OPD neither creates features nor passes on the teacher's own, and leaves the firing rates of over 98% of the student's frequently used features within 20%. We further examine the SFT warm-up on the teacher's rollouts that commonly precedes OPD and makes it more effective. Rather than adding features, the warm-up reweights the shared ones in two ways. First, it already raises and lowers many of the features that OPD later raises and lowers, doing part of OPD's work in advance. Second, it changes features that OPD alone would not, notably those for conversation format, reasoning style, and mathematical notation, and these changes persist through OPD. Imposing this reweighting on a directly distilled student's features, without changing its weights, brings its accuracy close to that of the warmed-up student, whereas the same change on shuffled features does not. Together, these findings suggest that OPD reweights existing features rather than acquiring new ones: the student learns from the teacher how to use the features they already share.

Nuclei instance segmentation is fundamental to quantitative pathology. Yet high-accuracy methods mostly build on large vision transformers (ViTs) or heavy convolutional backbones, whose parameter count and compute cost hinder deployment in clinical and edge settings. We present NucShare, a lightweight model whose shared decoder handles nuclear pixel segmentation, horizontal-vertical distance, and classification through learnable transposed convolutions for upsampling and three lightweight heads, using 37% of independent decoders' parameters with no loss in accuracy. Output-level distillation then transfers knowledge from NucShare-T (UNI2-h) to NucShare-S (ConvNeXt V2-Tiny), reducing 690.5M parameters to 34.7M. On PanNuke they reach mPQ 0.524 and 0.519, above the prior best CFR-SAM-H at 0.511, and NucShare-S surpasses HoVer-Net and Mask R-CNN by 5.6 and 15.0 points; on zero-shot MoNuSeg it attains bPQ 0.668 against CFR-SAM-H's 0.662. Code: https://github.com/L-iMeng-L/Newidea

Multimodal large language models (MLLMs) have approached image segmentation by reasoning about visual content and predicting target locations. Their generated responses contain reasoning traces and localization proposals that can serve as working memory when revisiting the same image and query. Our exploration reveals that MLLMs benefit from using this self-generated working memory as context, leading to enhanced reasoning segmentation. Motivated by this finding, we seek to strengthen the backbone model's reasoning segmentation capabilities by distilling the guidance gained from revisiting prior attempts, enabling it to benefit with or without working memory at inference time. To this end, we propose Reasoning Segmenter with Working Memory (SWiM), a working-memory distillation framework for reasoning segmentation. Specifically, SWiM selects rollouts based on segmentation quality to construct working memory and uses the memory-conditioned model as a teacher. The teacher provides token-level distributional supervision along student-generated trajectories, while the student receives only the original image and query. Joint optimization of on-policy self-distillation and outcome-based reinforcement learning combines working-memory guidance with direct feedback on segmentation quality. Extensive experiments on reasoning segmentation benchmarks demonstrate that SWiM achieves state-of-the-art performance, validating the effectiveness of working-memory distillation.

Underwater image enhancement is crucial for improving visual perception in marine applications. Existing underwater image enhancement studies mainly focus on enhancement quality and visual fidelity, while rarely considering real-time deployment capability, which is essential for resource-constrained underwater robots. To this end, we introduce a physics-guided spectral distillation (PSD) method, which reduces model capacity for real-time applications while maintaining the high performance of underwater image enhancement models. To decompose the outputs of teacher and student models, PSD adopts a multilevel Haar discrete wavelet transform. It transfers low-frequency color and illumination information as well as high-frequency structural details through band-specific objectives. Moreover, the distillation process of PSD is degradation-aware. We estimate degradation-aware weights through a physical head and combine them with ground-truth-guided reliability masks to selectively retain valuable teacher guidance. Experiments on the UIEB, LSUI, and EUVP datasets validate the effectiveness of the proposed method. Furthermore, we demonstrate the benefits of enhanced images for downstream perception tasks, including object detection. Deployment on a self-developed ROV further demonstrates its practical applicability in real-world underwater scenarios.

On-Policy Self-Distillation (OPSD) has emerged as a promising paradigm for safety alignment, delivering dense, token-level supervision by distilling from a teacher conditioned on refusal-oriented privileged prompts. However, we reveal that this paradigm suffers from critical inefficiencies that degrade both training efficiency and general reasoning capabilities. Specifically, we diagnose two fundamental bottlenecks: (1) supervisory collapse over extended rollouts, where the teacher's corrective efficacy degrades precipitously as the student's generation prefix lengthens, injecting noisy gradients into late-stage tokens; and (2) gradient dilution from stylistic shifts, where the distillation objective is dominated by safety-irrelevant stylistic discrepancies induced by privileged prompting, washing out genuine safety signals and impairing base reasoning. To resolve these issues, we propose Efficient On-Policy Self-Distilled Safety Alignment (EOPSA), which concentrates computational and gradient budgets exclusively on reliably supervised, safety-critical tokens. EOPSA incorporates two coordinated mechanisms: (i) Adaptive Rollout Scheduling, which dynamically bounds the generation horizon guided by a novel Teacher Rescue Rate (TRR) metric to operate strictly within reliable supervision regimes; and (ii) Selective Distillation, which filters out safety-neutral tokens to restrict gradient updates exclusively to safety-pivotal transitions. Extensive evaluations across reasoning models up to 32B parameters demonstrate that EOPSA slashes rollout computation by $\sim$50% and backpropagates through merely $\sim$2% of tokens, consistently outperforming full-token distillation baselines in both safety compliance and reasoning retention.

Data pruning reduces the training cost of knowledge distillation (KD). However, the preferred teacher capacity changes with the data budget: smaller teachers can outperform larger ones when limited training data are available. Understanding what drives this shift is important not only for teacher choice but also for identifying which samples are useful for distillation. We analyze teacher supervision by decomposing it into relational ordering---the ranking of classes---and score geometry---the magnitudes and margins of class probabilities---and show that the small-teacher advantage in the low-data regime arises not only from score geometry but also from relational ordering. Beyond understanding teacher capacity, our analysis reveals two properties of effective subsets: samples should match the difficulty appropriate for the available budget, and their relational signals should be diverse rather than redundant. Based on these findings, we propose DVA (Difficulty- and Volume-Aware data selection for KD), a training-dynamics-free method, which uses a small teacher as a proxy for budget-aware difficulty filtering and class-conditional relational volume maximization. Despite requiring no training dynamics statistics, our method remains competitive with training-dynamics-based methods while consistently outperforming training-dynamics-free baselines.