Distillation 进展

Algorithm Distillation (AD) has demonstrated the remarkable ability of Transformers to perform in-context reinforcement learning without explicit weight updates. However, capturing long-term learning progress necessitates expansive context windows, which incur prohibitive memory costs and limit scalability in complex, long-horizon tasks. To address this bottleneck, we propose Recurrent Algorithm Distillation (RAD). RAD employs a dual-component architecture: a Compression Transformer that distills extended interaction histories into compact latent tokens, and an AD Transformer that auto-regressively generates actions using a hybrid context of these compressed memories and recent transitions. By maintaining a fixed-size latent buffer, RAD decouples the effective history length from computational complexity, functionally providing the model with a long-horizon memory. Empirical evaluations across diverse environments demonstrate that RAD matches the asymptotic performance of standard AD with significantly reduced context window sizes, offering a scalable solution for efficient in-context decision-making.

Large Vision-Language Models (LVLMs) have shown strong promise for multimodal reasoning, yet often struggle with tasks requiring concepts beyond what is directly observable in the input image. Existing methods generate intermediate images or latent visual tokens to guide reasoning, but these representations can introduce errors and increasingly interfere with textual reasoning as reasoning progresses. We propose Visual-to-Text Chain-of-Thought Distillation (V2T), a framework that enables LVLMs to internalize visual reasoning without generating intermediate visual representations at inference time. V2T first trains a teacher LVLM using interleaved visual and textual chains of thought, and then uses knowledge distillation to train a student LVLM using the teacher's logits and cross-entropy supervision from ground-truth textual reasoning. When reasoning images can be mapped to the original image, V2T can additionally distill the teacher's attention to corresponding regions, while ground-truth bounding boxes can further guide a subsequent reinforcement learning stage. Experiments across multiple multimodal reasoning benchmarks show that V2T consistently outperforms the teacher and existing baselines, improving average accuracy by 14.3% on a held-out set and 2.7% on the broader visual evaluation suite. Moreover, lightweight SFT and substantially reduced RL make V2T up to 42x faster to train than state-of-the-art baselines.

Socially intelligent language agents must negotiate, coordinate, and resolve conflicting preferences while respecting the time and attention of both participants. Balancing these demands is challenging because agents must convey enough to address a partner's constraints and advance their goals without adding words that do not help the interaction. In this paper, we propose Teacher-Assisted Communication Training (TACT) to improve social goal attainment while reducing communication cost, making interactions with agents more productive and less demanding. We first characterize communication efficiency in terms of action strategy and expression, whose effects extend beyond the current utterance to the partner's response and subsequent exchanges. We design TACT to revise student-generated actions, test the revisions through partner responses, and distill useful feedback into the student. An expression specialist removes unnecessary detail while preserving the intended action, while a strategy specialist proposes alternatives that may better address the partner's constraints. To determine which revision helps, TACT samples a partner response for each candidate and selects a teacher reference by balancing local goal support against action-token cost. That reference guides on-policy distillation on the student's own generation prefixes, allowing the student to act independently at deployment. We evaluate TACT on SOTOPIA and AgentSense. On SOTOPIA, it achieves the highest Goal among the evaluated methods on All and Hard while using substantially fewer target tokens than SFT+SDPO. On AgentSense, it improves goal success over the initial student while reducing target tokens and interaction messages.

On-policy distillation (OPD) is becoming an important component of large language model (LLM) post-training for transferring the reasoning capability of a strong teacher LLM to a weaker student LLM. OPD trains the student by minimizing the reverse KL divergence between the teacher and the student via rollouts generated by the student's policy. However, estimating the gradient of the reverse KL divergence in OPD remains a challenge. Using only the sampled token from the student-generated rollout is computationally cheap but provides limited distributional supervision, which will degrade accuracy. In addition, using the full vocabulary provides complete distributional supervision but is computationally expensive. Therefore, recent works propose Top-$k$ OPD (TK-OPD) that use selected top-$k$ tokens, which provides richer distributional supervision than sampled-token estimation at substantially lower computational cost than full-vocabulary estimation. Unfortunately, using only the selected top-$k$ tokens induces bias, leading to accuracy degradation, as the probability mass outside the selected top-$k$ tokens is discarded. To address the bias of TK-OPD, we propose Tail-Corrected Top-$k$ On-Policy Distillation (TT-OPD). It preserves the advantages of TK-OPD, including rich distributional supervision and low computational cost, while providing an unbiased estimator of the gradient of the reverse KL divergence. The key insight of TT-OPD is to use not only the selected top-$k$ tokens, but also the sampled token from the student-generated rollout, thereby recovering the discarded probability mass in expectation, avoiding the bias. Experimental results demonstrate that TT-OPD significantly outperforms other tested OPD variants.

Fine-grained visual understanding requires models to recognize small details within complex images. Multimodal on-policy self-distillation (OPSD) addresses this challenge by using a teacher conditioned on evidence-centered crops to supervise a student conditioned on original images along student-generated trajectories. Ideally, teacher corrections, the distributional changes from the student toward the privileged teacher, should be driven by task-relevant visual evidence. However, the designs that make the teacher effective also introduce other interference. Using a lagged or frozen teacher improves training stability but introduces a model-state gap from the evolving student, while cropping enhances task-relevant evidence but also loses the visual context. These two sources of interference make the teacher corrections not purely rely on the visual evidence. We introduce Evidence-Aligned multimodal on-policy self-Distillation (EAD), which retains the crop-conditioned teacher as the target but constructs a separate evidence reference for weighting the corrections. To exclude the effect of lagged model-state from this reference, EAD measures prediction changes using the current student. To avoid crop-induced context changes, EAD masks the evidence region in the original image while preserving the other visual context. The change from the student's masked-image prediction to its original-image prediction provides a controlled reference for the direction in which the visual evidence shifts the student's prediction. EAD weights each teacher correction by its cosine alignment with the reference, i.e., retaining aligned corrections and downweighting the rest. Retaining only 6% of the supervision mass of dense OPSD, EAD consistently outperforms previous state-of-the-art methods.

Multi-teacher on-policy distillation (MOPD) combines independently developed domain teachers into a single student by distilling their predictions on student-generated samples. We study a setting where teachers share a reference model but undergo different post-training procedures, and find that MOPD can struggle to recover some teacher capabilities. Because distillation occurs on student-generated prefixes, the student initialization can strongly affect subsequent recovery. However, initial benchmark performance is not a reliable predictor of a good MOPD initialization. For example, merge initialization can start below SFT warm-up yet finish higher after MOPD. We further find that effective merging depends on both the relative teacher contributions and the overall merge scale, with some strong configurations lying outside the simplex of convex parameter averaging. Thus, selecting a good merge initialization requires evaluating not only its immediate performance but also the learning it enables under MOPD, making one-shot coefficient search difficult. We propose Iterative Merging for MOPD (IM-MOPD), which starts from a uniform merge and progressively adds task-vector increments for under-recovered domains during distillation. In a 5-domain setting, IM-MOPD achieves higher average normalized recovery than MOPD with either uniform merge initialization or SFT warm-up, showing that effective teacher contributions can be determined progressively during training.

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.

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.

LLM personalization aims to generate responses aligned with individual users' preferences and needs. User-specific rubrics make these expectations explicit, providing direct supervision on what a satisfactory answer should cover. Existing rubric-guided approaches, however, exploit such guidance only at a coarse granularity, either by using rubrics to supervise the prediction of relevant aspects for subsequent generation or by reducing aspect coverage to a single response-level reward for reinforcement learning. This leaves a gap between specifying what a personalized answer should contain and teaching the model how to generate it. To bridge this gap, we propose GRASP, a rubric-aware on-policy self-distillation framework for LLM personalization that turns user-specific rubric aspects into fine-grained, token-level supervision. Specifically, GRASP pairs a rubric-free student with a rubric-informed teacher that additionally receives the target user-specific rubrics. By aligning their next-token distributions along on-policy trajectories generated by the student, GRASP transfers the teacher's rubric-conditioned guidance into the student, translating user-specific semantic requirements into dense token-level supervision. Since rubric-informed teachers can still produce inadequate supervision, we further introduce Rubric-based Teacher Validation (RTV), which retains only instances where the teacher sufficiently covers the target aspects, improving both supervision quality and training efficiency. Experiments on the LaMP-QA benchmark for personalized question answering demonstrate that GRASP achieves state-of-the-art performance across multiple backbones, supporting the effectiveness of rubric-guided token-level supervision for personalization. To ensure reproducibility, our code is available at https://github.com/SnowCharmQ/GRASP.

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.

Verbal feedback can identify errors and prescribe corrections, providing rich supervision for language-model post-training even when reliable programmatic verifiers are unavailable. Such feedback, often generated by a capable model, can be used to condition the teacher in on-policy distillation, which trains the student to match the teacher's predictions on student-generated rollouts. However, this approach can transfer teacher preferences that the feedback did not motivate, while leaving much of the feedback's guidance unused. We find that both problems come from the standard on-policy distillation objective, specifically the divergence it minimizes and the distribution it uses as its target. Our proposed method addresses both limitations. First, to isolate the information conveyed by the feedback from the teacher's inherent preferences, we treat verbal feedback as evidence for or against the hypothesis that a particular token comes next at a given prefix. We then adopt a probabilistic confirmation framework which uniquely determines an ordering over the vocabulary based on the teacher's predictions before and after it receives feedback. Using a confirmation score consistent with this ordering, we construct a target distribution within a trust region of the student. Second, to learn from guidance that student rollouts can leave unused, we derive a simple shared-rollout estimator of a symmetric divergence between the student and target distributions over rollouts, reusing student and feedback-conditioned teacher rollouts in both directions through importance weighting. Empirical evaluations show that our method outperforms the common on-policy distillation recipe and a recent contrastive variant on knowledge-based and agentic benchmarks.

On-policy distillation (OPD) uses teacher correction on student-generated responses. Full-vocabulary correction can provide important corrections even for tokens that the student assigns low probability, but backpropagating through all token logits becomes memory-intensive for long sequences. Existing memory-saving approaches estimate corrections from sampled tokens or restrict supervision to the student's TopK tokens, introducing sampling noise or changing the full-vocabulary correction. We introduce SparseOPD, which uses full-vocabulary teacher correction to determine which corrections matter before selecting the token logits to differentiate. SparseOPD first constructs the full-vocabulary correction without retaining its backward graph, then selects tokens by correction magnitude rather than student probability. Signed residual compensation preserves the total promoting and suppressing correction mass, while correction-aware budget allocation distributes the sparse support across positions. Finally, the update backpropagates only through the selected token logits. Across six task--scale settings spanning mathematics, chemistry QA, and multimodal reasoning, SparseOPD outperforms Sampled Token and TopK in task-average accuracy and matches or exceeds Full Vocabulary. Gradient cosine similarity reaches 99% on 4B mathematics, while 8K full-parameter profiling shows 70.5% lower backward memory.

Maintaining persona consistency across multi-turn dialogues remains a core challenge for role-playing language models. Off-policy distillation from external teachers incurs distribution mismatch that compounds across dialogue turns, while reinforcement learning struggles with reward ambiguity inherent in subjective persona fidelity. We propose OSPD, an on-policy self-distillation framework where the same model serves as both teacher and student under asymmetric information: the teacher receives a complete character profile while the student sees only a brief summary, and the student generates trajectories from its own policy. We find that teacher confidence in role-playing dialogue exhibits a bimodal structure---sharply peaked at character-critical tokens yet diffuse at generic utterances---and introduce role-aware divergence switching to match this structure. A progressive trait masking curriculum further forces staged internalization of character knowledge along semantic dimensions. Experiments on CharacterBench, CharacterEval, and SocialBench show that OSPD substantially improves persona consistency over supervised fine-tuning and multi-turn RL baselines, without requiring any external teacher or reward model.

Large language models (LLMs) typically generate text autoregressively (AR), predicting one token at a time. Block diffusion language models (dLLMs) instead generate blocks sequentially while denoising multiple tokens in parallel within each block, offering a promising way to accelerate generation. Rather than training such models from scratch, recent work adapts strong pretrained AR models into block dLLMs through distillation. On-policy distillation (OPD) has been widely used for LLM training because it supervises the student on states generated by its current policy, rather than only on fixed offline trajectories. By training on the states the student actually visits, it reduces the mismatch between training and generation and can provide more relevant supervision as the student evolves. Recent work has extended this idea to AR-to-block-diffusion conversion. However, this setting introduces a fundamental mismatch in supervision: the block-diffusion student and the causal AR teacher condition on different information at the same training state. The student predicts from the entire partially denoised block, including visible future context, whereas the standard AR teacher target is defined only from the causal prefix. As a result, the teacher distribution used for distillation is not fully aligned with the information available to the student. We therefore introduce d-OPD, a future-aware on-policy distillation method that corrects the AR teacher distribution to better align with the student-visible state by incorporating visible future information within each block, providing supervision that better matches the information used by the student. Across Qwen3 models from 0.6B to 8B, d-OPD improves the six-benchmark average by up to $4.0$ points over OPDLM and reduces training time by $1.35$-$1.58\times$. The code is available at https://github.com/mit-han-lab/d-OPD.

Temporal video grounding is a key capability of advanced Multimodal Large Language Models (MLLMs) for the thorough understanding of video events, which is however often limited by repeated actions and visually similar contexts in long videos. In this paper, we study this issue from the perspective of On-policy distillation (OPD) and propose a new training regime for MLLMs termed Counterfactual Attention Policy Distillation (CAPD). In particular, OPD is a viable solution for MLLMs via providing dense teacher supervision on student-generated trajectories. But its next-token based teacher-student distillation is hard to identify the specific video segments supporting each predicted timestamp, which is critical for temporal grounding. In this case, CAPD measures how masking each temporal group changes the teacher's output distribution. The resulting counterfactual influence calibrates the teacher's attention and weights token-level distillation, allowing the student to learn the temporal evidence that affects boundary prediction. To validate CAPD, we trained it on Qwen3-VL-8B-Instruct using only 2,500 samples for one epoch, and evaluated it on the TimeLens and multiple general video benchmarks. Experimental results show that CAPD improves average recall by 12.0% relative to GRPO on TimeLens while preserving general video understanding, achieving comparable accuracy to the base model.

On-policy distillation (OPD) trains student agents through teacher supervision on their own interactions with an environment. However, in asynchronous multi-turn training, arrival-order batching can allow a few early or long rollouts to dominate learner updates while other valid rollouts become stale before being used, wasting already-generated experience. To address this problem, we introduce DivOPD, a simple learner-side batch-selection method that spreads a fixed turn budget across more rollouts and, within each rollout, prioritizes turns with larger cumulative teacher-student disagreement. Turns without usable teacher feedback are excluded. The per-turn loss and optimizer remain fixed; selection only changes which student-visited turns receive training weight. For no-progress rollouts, an optional extension briefly hands control to the teacher before returning it to the student. Across six teacher-student settings on the simulated ALFWorld, ScienceWorld, and WebShop benchmarks, with 1.5B-7B students, DivOPD raises cross-setting mean peak success rate from 77.4 to 84.4 and mean success over the last five evaluations from 71.5 to 78.6. It reaches all reported setting-specific targets with geometric-mean speedups of 1.84x in training tokens and 1.87x in learner GPU time relative to vanilla OPD. Teacher intervention further raises this last-five mean to 82.4 while retaining about 1.7x learner-GPU speedup over vanilla OPD. Code will be released at https://github.com/HanyangWang0418-oss/DivOPD.

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.

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.

Sampling latency compounds in diffusion workflows, where users generate and discard many candidates before keeping one. Surprisingly, the poor outputs of standard few-step samplers do not reflect a lack of reconstruction capacity: by optimizing only the initial noise, frozen 3-4-step samplers can closely reproduce their corresponding full-step outputs. Building on this finding, we learn corrections to the initial noise and denoising updates using endpoint supervision, improving correspondence with full-step outputs generated from the same noise and prompt. The resulting previews allow users to screen candidates cheaply and reserve full-step generation for promising ones. Input correction also transfers across sampling budgets without retraining. Experiments show substantial improvements in reference fidelity, including 53-78% lower reconstruction MSE than retrained LD3 on unconditional benchmarks, alongside improved ranking preservation and candidate selection on SD1.5, SDXL, and FLUX.1-dev.

On-policy distillation (OPD) transfers knowledge between language models through teacher supervision on student-generated trajectories. With different tokenizers, a single teacher token may require multiple student tokens to generate, creating intermediate states where the event is entered but not yet completed. Existing cross-tokenizer methods align tokens or text spans to construct comparable prediction targets. We study a complementary problem after partial generation: once the student produces a prefix of a teacher token, multiple next tokens may complete the same remaining bytes, but the teacher only specifies the required completion rather than how probability should be divided among these valid continuations. We introduce Event-Set Completion Distillation (ESCD), which complements cross-tokenizer probability alignment with completion-set supervision. ESCD aggregates prefix-related teacher events and supervises the total probability of byte-compatible one-step student completions, avoiding tokenizer-dependent probability splits among individual tokens. The method reuses student trajectories and predictions, requiring neither additional rollouts nor changes to the student vocabulary. Experiments demonstrate consistent gains in mathematics, code, and scientific reasoning across model families and tokenizers, extending to large-scale MoE distillation from a 1T teacher to a 35B student. Local analyses show that retaining completion sets better matches the reference supervision, while one-step completion covers over 99% of observed compatible teacher mass after partial event entry in the studied tokenizer pairs. These findings support event entry and event completion as complementary supervision targets for cross-tokenizer knowledge transfer. Code will be released on GitHub.

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.

Effective multi-turn agents require interaction strategies that coordinate information gathering, actions, and feedback over long horizons. GRPO is a reinforcement learning algorithm used to train these agents, but sparse trajectory-level rewards limit early exploration in small models. Recent methods augment RL with on-policy distillation (OPD) from a stronger teacher. However, a fixed mixture assumes that teacher guidance and reward optimization should retain a constant relative role throughout training and across interaction turns. This assumption can fail at two scales. Globally, as training progresses, maintaining strong distillation pressure can constrain the model from moving beyond the teacher's capabilities. Locally, teacher--student disagreement identifies where the student departs from the teacher, but cannot tell whether that departure is exploration supported by better outcomes or low-quality policy drift. Our methodological insight is that teacher guidance and reward optimization should be dynamically rebalanced over training and jointly allocated across turns. We instantiate this insight in \tide. Globally, \tide uses the measured disagreement trend as a practical schedule signal, advancing an OPD-to-RL handoff when discrepancy reduction becomes slow but remains positive and progressively increasing the relative weight of RL. Locally, \tide jointly modulates teacher-guided and reward-driven updates: relative action value and disagreement prioritize the OPD signal, whereas relative action value supplies the RL advantage and normalized disagreement reweights it across turns. Coupled with the global handoff, \tide allocates stronger teacher guidance early and gives reward-driven updates greater relative weight later in training. Experiments across multiple benchmarks, student scales, and controlled ablations support the effectiveness of TIDE's adaptive OPD--RL coordination.

Reward-specialized post-training produces strong experts for flow-based generative models, while multi-teacher on-policy distillation (OPD) consolidates their capabilities into a single student. Existing methods, however, route each prompt to a single teacher according to its semantic category, implicitly binding the desired capability to prompt content. This coupling makes capability invocation vulnerable to prompt perturbations and prevents users from explicitly adjusting the strength of the desired capability at inference time. In this work, we introduce CapField-OPD, an OPD framework that integrates multiple teachers into a continuous capability field through explicit capability coordinates. We use teacher models as anchors to construct this field, with the coordinates determining how their outputs are combined. Each capability configuration thus receives a unique supervision target, and capability control no longer depends on prompt semantics. Since the training anchors may not be optimal at inference time, we further profile the learned field on a small calibration set. The coordinate with the highest mean reward serves as the recommended default, while coordinates that are frequently optimal offer a promising candidate set for test-time scaling. Extensive experiments on compositional generation, text rendering, and visual aesthetics demonstrate that CapField-OPD consolidates multiple specialized teachers into a single student while preserving or surpassing their performance, reliably invokes the desired capabilities under semantics-preserving prompt variations, and supports continuous capability control and coordinate-based test-time scaling.

Few-step autoregressive video generation commonly relies on Distribution Matching Distillation (DMD), requiring a bidirectional diffusion teacher and an online fake-score model. We instead learn the rollout distribution directly from reference videos, eliminating both score models during post-training. Our framework minimizes maximum mean discrepancy (MMD) in frozen self-supervised video representation spaces, using a hybrid Nyström--Monte Carlo estimator to balance approximation bias and sampling variance. Memory-efficient replay and gradient subsampling make this objective practical. Using the same architecture and initialization as Self-Forcing, our 1.3B model improves the VBench Total score from 83.80 to 84.64 while retaining 17 FPS. Removing auxiliary score models also enables 14B post-training on eight H200 GPUs. Beyond distillation, learning from reference videos enables the acquisition of new visual styles, semantic concepts, and spatial priors without a target-specific diffusion teacher.

Interactive world models require responding in real time to versatile controls and maintaining long-horizon consistency. However, modeling heterogeneous controls remains difficult, while explosive contexts and unstable distillation impede achieving both long-horizon consistency and real-time responsiveness. In this paper, we present WorldPlay2, an interactive world model that couples a factorized hybrid control interface with a co-design of compressed memory and stable distillation. 1) Our factorized hybrid control interface integrates frame-aligned action control with structured semantic control that explicitly disentangles scene appearance, character identity, and dynamic semantic events, thereby facilitating effective control learning. 2) To achieve efficient long-horizon modeling, we compress historical contexts into compact memory tokens shared by the autoregressive student and the bidirectional teacher. This design enables clip-wise, memory-conditioned score evaluation instead of jointly processing an entire long rollout, substantially reducing distillation overhead. 3) We further propose Stable Forcing, which initializes the autoregressive student via a few-step strategy and leverages full-rollout replay to preserve the quality of long-horizon rollouts, ensuring robust and stable distillation. Extensive experiments demonstrate the strong generalizability of our model and its superior performance compared to existing methods.

Instruction-based image editing has achieved strong performance in single-turn settings, yet practical editing is often iterative, with each instruction applied to the output of the previous turn. We find that existing editing models degrade rapidly under recursive editing and attribute this failure to a train-test mismatch in the conditioning distribution: models are trained on clean source images but must repeatedly condition on their own imperfect outputs at inference time. To address this, we propose MT-OPSD, an on-policy self-distillation framework that trains the model on self-generated conditioning states with editing supervision from a clean-conditioned teacher, without requiring multi-turn annotations. We further introduce LME-Bench, a benchmark of 100 ten-turn editing sessions for evaluating long-horizon robustness. Experiments across three editing backbones show that MT-OPSD substantially improves long-horizon editing success and reduces multi-turn collapse while largely preserving single-turn editing quality.

On-policy distillation (OPD) specializes pretrained video diffusion models through teacher supervision along the student's own generation trajectory. Although large video models are natural teachers, developing specialized video experts can require costly video data and training, while querying them incurs substantially higher latency than querying image experts. More readily available and cheaper to query, image experts offer a cost-effective alternative, particularly for largely temporal-agnostic capabilities such as aesthetics and OCR that admit frame-level supervision. However, heterogeneous image and video latent spaces prevent direct supervision of intermediate student states, while image experts lack cross-frame motion supervision, making temporal consistency vulnerable to frame-level improvements. In this paper, we propose MILD, a Motion-Preserving Image-to-Video Latent Distillation framework that transfers specialized image expertise while preserving pretrained video dynamics. MILD uses a learnable linear connector that aligns student latent states and predicted updates with those of image experts, enabling supervision transfer across heterogeneous latent spaces. We further constrain image-guided corrections around the pretrained student's predictions to preserve video dynamics and incorporate an optical-flow-based motion reward to improve motion quality and temporal consistency. Across specialized image experts and multiple video-student backbones, our method consistently outperforms video-teacher OPD baselines, with further studies demonstrating effective transfer across connector designs and heterogeneous architectures. These results establish image-to-video distillation as an effective route to improving video generation by drawing on the diverse and evolving capabilities of the image-generation ecosystem.

Modern video diffusion models require tens of denoising evaluations over long spatiotemporal token sequences. Distribution Matching Distillation (DMD) reduces the number of function evaluations (NFE) to just a few. However, DMD samples can degrade during training, exhibiting progressive oversaturation and artifacts. We trace this instability to critic errors, which enter successive student updates and accumulate over time. We introduce Projected Distribution Matching Distillation (PDMD) to filter critic errors. PDMD projects out the component of the DMD update parallel to the student-critic endpoint residual. At a fixed noisy query, we prove that this residual is an unbiased estimate of the critic's endpoint error. Under high-dimensional assumptions, this projection removes a constant fraction of critic error while discarding only a vanishing fraction of ideal DMD signal. Empirically, the projection stabilizes training and improves sample quality where DMD degrades and develops unnatural textures. PDMD requires only a one-line code change to DMD, with no extra loss, network, data, model pass, or multi-stage training. With Wan2.1, PDMD achieves a VBench total score of 83.73 at 4 NFE, surpassing matched DMD by 1.03 points. On MiniMax-H3 joint video-audio generation, PDMD achieves a VideoGen-Eval visual total score of 83.17, 0.41 points above the strongest distilled baseline. PDMD also achieves the best performance on all six audio metrics among the compared 4-NFE models. Qualitative comparisons and user studies favor PDMD over the distilled baselines in visual quality, motion, and audio quality. Code and models are available at https://pdmd2026.github.io/.

On-policy distillation (OPD) trains a student on its own trajectories with dense teacher supervision. Recent work on OPD for multi-turn autonomous agents often treats large teacher-student token-level distributional gaps as promising intervention points, linking larger gaps to a greater need for correction. Yet, our empirical analysis reveals a supervision-benefit mismatch: large gaps can be benign, while small gaps can be outcome-critical. Teacher-student gaps capture differences at the current turn, whereas the benefit of teacher guidance depends on how the current student interacts with the environment afterward. The student may still succeed despite choosing an action that differs from the teacher's, while a teacher-preferred action may lead to a state from which the student cannot complete the task. Local gaps alone are therefore not enough to determine whether teacher guidance benefits the current student. Effective supervision should instead emphasize guidance that the current student can translate into better final task outcomes. Accordingly, we propose Outcome-Guided On-Policy Distillation (OG-OPD), which applies trajectory-relative weighting to teacher supervision and calibrates these weights using final task outcomes from paired student continuations. This calibration selectively strengthens supervision on the student's original trajectories at turns where teacher guidance benefits the current student. Across ALFWorld, ScienceWorld, and WebShop, OG-OPD consistently outperforms baselines under diverse settings. It improves task success rates by 3.6-17.7 percentage points over vanilla OPD and by up to 7.0 percentage points over the strongest baseline.

Reinforcement learning with verifiable rewards (RLVR) via Group-Relative Policy Optimization (GRPO) is widely used for multi-turn VLM agent training, yet it suffers from zero-gradient silence on uniform failures and coarse episode-level credit assignment. While On-Policy Distillation (OPD) and On-Policy Self-Distillation (OPSD) mitigate sparse rewards using hindsight information, their underlying mechanisms remain poorly understood. Through controlled counterfactual rollback probes across five multi-turn VLM agent benchmarks, we reveal that performance gains in OPSD/OPD are largely driven by physical state rollback at the pivot step, defined as the first unrecoverable action without remaining step budget. However, physical state rollbacks are computationally prohibitive and infeasible in real-world environments. To bridge this gap, we present Pivot-Aware Internalized Visual On-Policy Training (PIVOT), an RL framework that internalizes pivot localization and state restoration directly into token-level parameter updates, eliminating environment rollbacks during RL training and additional skill hints at test time. PIVOT unifies three functional roles within a single architecture: a failure Analyzer non-invasively localizes the pivot step and diagnoses failure modes from visual trajectory collages and action logs; a detached Teacher re-scores failed tokens under this privileged diagnostic context; and a Student optimizes joint GRPO and confidence-gated OPD objectives. At test time, both Teacher and Analyzer branches are stripped. Evaluated on five multi-turn VLM agent tasks across cognitive grid puzzles, 3D embodied control and navigation, and generative reasoning, PIVOT achieves 0.90 overall accuracy on Qwen2.5-VL-3B (+8% over SFT+GRPO baseline and +5% over previous SOTA) and scales to 0.92 on Qwen3-VL-2B (+12% over SFT+GRPO baseline).