每天知道
你的领域
出了什么

Why does removing reference-policy KL regularization sometimes improve group policy optimization? This motivates studying how reference-policy information should enter group-relative updates. We analyze seven potential failure modes in the interactions between KL and rewards: residual KL updates after reward clipping, after gradient cancellation, and in groups with identical rewards; KL growth with response length and an imbalance in its relative contribution; KL concentration on a small number of tokens; and sampling noise when k1 is incorporated into rewards. We propose Zero-Sum Calibrated Policy Optimization (ZCPO), which uses relative drift measured by conditional KL to calibrate within-group reward coefficients and integrates them into the base surrogate. Mathematical reasoning experiments and ablations support this design's effectiveness in our settings.

Foundation models supply reinforcement learning (RL) with priors that mitigate its longstanding weaknesses in sample efficiency and transfer, but their token-by-token generation makes queries sequential and costly. Jev, a recently released decision model, generates nothing and returns calibrated, typed answers in a single forward pass. Existing work studies foundation models in RL either as models to be trained or as generators to be prompted, and Jev belongs to neither category, having so far served only as a black box in single domains. How well such a model decides on its own in RL environments, and how it can improve RL as a component of training, therefore remain unaddressed. To this end, in this paper we first examine the requirements that the objects of an RL system place on the answers they consume, and establish that Jev can fulfill all of them except the cardinal use of a value function. The remaining objects form positions that admit several roles each. We then construct algorithms that employ Jev at three of these positions, as a reference policy, an exploration judge, and a replay rater, to improve sample efficiency, exploration, and learning performance. Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, including where the learner makes no progress alone, while the model itself remains untrained. To our knowledge, we present the first use of Jev within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how Jev and other advanced decision models can improve RL.

Reinforcement learning (RL) is the central training paradigm for advancing large foundation models towards self-improvement. This report introduces the MiMo-V2.6 series, an omni-modal family that pushes the frontier of model intelligence by scaling RL compute. Prior to RL, we conduct mid-training on a broad multimodal corpus to provide ample exploration space, and build a solid infrastructure on the pretrained hybrid-SWA architecture to support subsequent scale-up. We scale RL compute along three dimensions: (1) larger batches and higher throughput, with an asynchronous training that consumes 1,568 samples and 2.7-3.7B tokens per step at context lengths of up to 1M; (2) more diverse and complex environments, spanning code, general, visual, and cyber domains under a mixture of agent harnesses; and (3) more grader compute, via groupwise agentic grading that yields more accurate reward signals for long-horizon tasks and steers the model towards shorter, more token-efficient solutions. To keep training stable at scale, we freeze the MoE router and establish a multi-layer defense against reward hacking. We further build infrastructure for mixed-task agentic RL, including a unified trajectory representation, high-concurrency multi-framework rollout, decoupled control and data planes, and training-inference consistency. We open-source the training dynamics, RL environments, and RL framework to facilitate reproduction and further research on scaled RL and model self-improvement.

Long-form video understanding often involves multiple questions about different aspects of the same recording. Yet existing video agents typically process each question through an isolated tool-use trajectory. This repeatedly restarts video exploration and memory construction, missing opportunities to acquire evidence jointly and progressively build a shared understanding that supports the complete question set. We introduce VAMR (Video Agent for Multi-Question Reasoning), which coordinates all questions about a video through one shared tool-use trajectory. At each round, a persistent policy model can invoke tools for one or more unresolved questions and submit answers for questions with sufficient evidence. Question-conditioned visual perception retrieves fine-grained clues for several questions in one call, while layered multi-question memory integrates reusable context into a shared video story and preserves separate evidence for individual questions. After supervised fine-tuning initializes this interaction protocol, we propose question-horizon policy optimization (\qhpo) to optimize shared trajectories in which questions progress and finish at different rounds. Specifically, a question-level critic estimates the value of each active question, while round alignment maps each question advantage to the rounds that directly serve it before the aligned advantages are aggregated to optimize the shared actor. Across LVBench, Video-Holmes, and LongVideoBench, VAMR achieves the highest accuracy overall and the fewest reasoning rounds among iterative methods. On LVBench, it reaches 62.1% accuracy, exceeding VideoARM by 4.3 points while reducing reasoning rounds and processed frames by 85.9% and 61.4%.

Large language models (LLMs) are commonly post-trained against LLM judges and other cheap surrogates because the true reward, such as human preference, is too expensive to query at scale. This practice often leads to reward hacking, where reinforcement learning against a miscalibrated surrogate leads to undesirable side effects. In this work, we study a setting in which a small number $n$ of model outputs are annotated with ground-truth labels (e.g., from expert review) and used to recalibrate the LLM judge before optimizing against it. Prior approaches to judge recalibration are costly or heuristic, and it is known that on-policy sampling fails when the surrogate is miscalibrated on a rare set of outputs. In this work, we propose envelope sampling, a theoretically-grounded method for judge recalibration that seeks to minimize an upper bound on the regret of the post-trained model under the assumption that the human reward and re-calibrated reward lie in an $L^2$ ball around the judge. We give practical algorithms to sample from the envelope by rejection or by fine-tuning against a modified reward, and experiments on clinical note generation and on a controlled sycophancy task show that recalibrating on envelope samples mitigates reward hacking where recalibrating on base-model samples does not.

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

A verified reference solution provides a correct trajectory for training a reasoning model. Alternatively, a prefix of the reference can guide the model in generating a trajectory of its own. How much reference guidance should we provide? We study this question through prefix continuation, where the model continues from a reference prefix and keeps the resulting trajectory if it passes verification, falling back to the reference otherwise. Since both procedures produce correct trajectories, we compare their distributions with the ideal distribution, the model's own distribution conditioned on successful verification. For one continuation, we derive the KL divergence in closed form, which, up to a bounded term, decreases with the product of the probability of generating a different correct trajectory and the reference surprisal, the negative log probability of the reference suffix given the prefix. Since a longer prefix tends to raise the former but lowers the latter, continuation success alone does not determine the preferred amount of guidance. From this analysis, we learn a prefix selector shared across training questions from continuation outcomes, without estimating success probabilities or additional generation. The resulting Adaptive Reference Guidance (ARG) constructs correct trajectories within a fixed generation budget, and we apply it to all-failure groups in Group Relative Policy Optimization (GRPO). Experiments on Qwen3-4B and Qwen3-8B across five mathematical reasoning benchmarks show that ARG achieves the highest aggregate pass@12 among the evaluated methods with competitive average sampled accuracy.

A language model (LM) can usually answer the same question in more than one way, but reinforcement learning with verifiable rewards (RLVR) is indifferent to which correct answer a model produces. A solution will earn the same reward whether it is the thousandth copy of a familiar answer or one the model has never produced before. Yet, there is potential value in having the model retain multiple correct solutions as it is trained. For instance, multiple modes may give users a choice and provide problem-solving strategies that improve overall model performance. Here, we introduce ModeBench, a benchmark of multi-solution tasks in which the verifier returns both correctness and mode discovered. We then use ModeBench to measure how solution diversity changes under RLVR post-training. We find that RLVR post-training concentrates probability onto fewer correct modes even as accuracy holds or improves, and moreover, that frontier models are already highly concentrated. We then introduce our solution, Re:Max, which stores one verified example per discovered mode in a replay buffer and trains on those stored modes uniformly. A solution found once is, therefore, practiced as often as one found repeatedly. Across three model scales, two RL objectives, and harder task constructions, replay improves both how often a policy succeeds and how many different ways it can succeed.

Large Language Models (LLMs) have shown strong reasoning capabilities when fine-tuned with reinforcement learning (RL), particularly through Group Relative Policy Optimization (GRPO). However, existing GRPO methods assume centralized access to training data, which may not hold in practice due to privacy or regulatory constraints. To this end, we propose Fed-GRPO, a federated GRPO training framework that addresses these privacy constraints by enabling collaborative reasoning training without sharing raw data, which leverages the reward statistics naturally produced during GRPO training as zero-cost signals to guide aggregation, local training, and communication. Fed-GRPO contains three reward-signal-driven mechanisms: (i) signal-weighted aggregation that weights clients by their reward standard deviation, prioritizing clients with stronger learning signals; (ii) global reward calibration that re-weights per-prompt objectives based on the local-global reward gap, steering each client toward its relative weaknesses; and (iii) adaptive sparse communication that allocates bandwidth based on the informativeness of each client's update. Extensive experiments on mathematical reasoning tasks demonstrate that Fed-GRPO achieves the best performance among all federated methods, clearly outperforms FedAvg and approaches centralized training performance, while losslessly reducing communication by $32\times$ and supporting up to $621\times$ compression under tight bandwidth budgets with only graceful accuracy degradation. Our code is available at https://github.com/HKU-HealthAI/Fed-GRPO.

Few-shot learning aims to recognize novel categories from limited labeled examples. Recent studies incorporate textual semantics to compensate for limited visual observations and improve class representations. However, high image-text agreement may reflect both intrinsic object properties and incidental context, making support prototypes susceptible to contextual contamination. To address this problem, we propose DVLA-RL++, which extends DVLA-RL with complementary semantic purification (CSP) and counterfactual reinforcement-learning gating (CRG). Specifically, CSP generates intrinsic and nuisance descriptions from labeled supports and compares their agreement with each support token. An ambiguity-dependent rejection margin guides sparse evidence allocation, while an intrinsic semantic anchor fills the unassigned mass to provide a fallback when visual evidence is unreliable. CRG learns layer-wise semantic fusion strengths using a reward that balances recognition performance and nuisance exposure. An independently executed reference trajectory on the same episode provides a paired learning signal. Theoretical analysis relates retained evidence and anchor quality to prototype stability and establishes conditions for unbiased on-policy gradient estimation. Experiments on standard, fine-grained, and cross-domain benchmarks show state-of-the-art accuracy, with an average gain of 1.4% over DVLA-RL. The project page is available at https://peacelwh.github.io/TPAMI27-DVLA-RLpp/.

Instruction-guided image editors have become highly capable, yet improving them further still depends on human-edited training pairs or external reward models. Such supervision is costly to obtain and can reward plausible failures: a realistic output may leave the requested change undone or alter content that should be preserved. In this work, we strive to improve a pretrained image editor using only its own generations, without human-edited targets or an external training-time reward model. To this end, we propose a self-evolving framework, named Rubric-CEPR, that verifies the editor's own samples with its internal representations through a rubric-augmented Contrastive Edit-Preservation Reward (CEPR). A Planner proposes structured edit instructions from unlabeled images, the Editor samples multiple candidate edits, and a frozen Critic scores each candidate with decomposed rubric checks for edit realization, removal of the old state, and content preservation, using features already exposed by the editor. Non-compensatory gates reject infeasible candidates, and the best verified candidate is distilled into the editor through lightweight adapter training. On Qwen-Image-Edit, Rubric-CEPR improves ImgEdit from 4.36 to 4.60 (+5.5%), with a +24.9% gain on object isolation, and transfers to GEdit-Bench and Complex-Edit. The same procedure also improves Step1X-Edit by +7.8% on ImgEdit. We hope our approach will serve as a solid baseline for image editors that improve themselves from their own verified samples. Our code is publicly available at $\href{https://riteshthawkar.github.io/Rubric-CEPR/}{\text{this URL}}$

Radiology report generation can automatically generate clinical descriptions from X-ray images, thereby significantly improving the efficiency of radiologists. This task is challenging because it requires medical knowledge to accurately identify diseases and describe them in a professional manner. However, existing methods often overlook the importance of enhancing medical knowledge in describing pivotal areas, a capability that requires models to effectively extract and aggregate knowledge at multiple levels of granularity. Accordingly, we herein propose a novel and compact Efficient Multi-Granularity Knowledge Transfer (EMGKT) method to address the above issues. First, we encode global knowledge embeddings using a medical vision-language model, which provides contextual medical knowledge. Moreover, we devise a novel Fine-Grained Knowledge Distillation (FGKD) training task which efficiently extract fine-grained knowledge. Specifically, the FGKD training task contains teacher embeddings and student embeddings. Teacher embeddings are encoded using extra priors; while student embeddings are learned from the teacher embeddings through knowledge distillation. During inference, the student embeddings are used to enhance fine-grained knowledge while the teacher embeddings are discarded, resulting in negligible computational costs and no need for extra priors. Finally, we further develop a mixture of disease diagnosis expert classifiers to enhance knowledge extraction. The classifiers are initialized using disease embeddings and are modeled as different experts to address various granularity features. Notably, EMGKT can be efficiently applied to most existing methods. Extensive experiments are conducted on two widely-used public datasets and various baselines, which demonstrates the effectiveness and transferability of EMGKT.

Unauthorized model distillation, in which a model is trained on the outputs of a proprietary large language model (LLM), is a growing threat to model providers. We study distillation inference: determining whether a suspect model was distilled from another model or trained independently. We formulate this problem as a hypothesis test and estimate the behavior expected under each hypothesis by training shadow models: distilled shadow models learn from the teacher's reasoning traces, whereas independent shadow models learn only from reference answers. The auditor measures how closely each model predicts the teacher's reasoning outputs and then uses the shadow models to convert the suspect's score into a calibrated p-value. In a preliminary study using Qwen2.5-7B as the teacher and Llama-3.2-3B for the suspects, our test achieves a true positive rate of 1.0 at a significance level of 0.02. These results demonstrate the feasibility of using distillation inference to detect distillation attacks.

Motion estimation in echocardiography is essential for quantitative assessment of cardiac function and myocardial mechanics, but remains challenging due to image artifacts, limited image information, speckle decorrelation, and the scarcity of ground-truth displacement fields. Anatomy-guided approaches can provide structural information, yet often rely on expert-labeled myocardial segmentations. We propose EchoDiST, a framework for unsupervised echocardiographic myocardial motion estimation that integrates self-distillation-based joint learning with a diffusion-conditioned motion estimation network. Here, unsupervised motion estimation refers to learning without ground-truth displacement fields. The self-distillation strategy jointly optimizes anatomical segmentation and myocardial motion estimation under limited anatomical annotations. Diffusion-based conditioning is used during training with stochastic perturbations, while inference requires only a single deterministic forward pass without iterative reverse-diffusion sampling. EchoDiST was evaluated on three echocardiographic datasets, including two external test datasets under cross-view and cross-dataset settings. Compared with seven representative learning-based methods, EchoDiST consistently improved anatomical alignment, myocardial strain assessment, and motion-derived functional and cardiac-phase assessment. These gains were statistically significant across the evaluated tasks and datasets. Overall, EchoDiST provides an effective approach for reliable myocardial motion estimation under limited anatomical supervision and supports downstream quantitative assessment of cardiac function.

Skills, reusable procedural guidance added at inference, can substantially improve LLM downstream performance (Li et al., 2026). Prior work retrieves skills from a bank by semantic relevance, then uses them as inference-time patches or for model distillation. The individual utility of each skill, however, is largely neglected. We first show that, in on-policy distillation where skill-conditioned policies serve as teachers, fewer than 25% of retrieved skills provide useful distillation signals. We then propose SGUID, a method for selecting a compact subset of skills for distillation. SGUID retains a skill only if it consistently yields effective learning signals during training. The selected skills are then distilled to produce a better model. Our results show that not all skills are worth distilling. Across four models from the Olmo and Qwen families, distilling 6 selected skills matches or exceeds full-bank distillation in mean avg@12 on three of the four models, and on all four after a second round that distills 3 newly selected skills, while the full banks are up to 11x larger. Importantly, SGUID supports stable model-skill co-evolution: after a distillation round, a new candidate bank is curated from the updated model's rollouts, and SGUID selects which skills to internalize next. In the second round, this loop selects 3 new skills and improves Qwen3-8B from 64.3% to 66.3%. The selection step is essential for stability: on Qwen3-4B, naively updating the model with unfiltered skills degrades performance, including a 0.3 percentage point drop on HMMT25, whereas SGUID improves HMMT25 by 0.5 points after the first round and 1.1 points after the second. These results identify skill selection as the key mechanism for stable model-skill co-evolution.

Spatial reasoning remains a persistent weakness of vision-language models (VLMs), because RGB inputs do not directly provide geometric evidence. Existing remedies either inject 3D into the model at inference, paying architecture and latency costs, or train with outcome rewards that supervise only the final answer. Spatial errors originate in perception: a misjudged depth or direction can be corrected only by the scene's true geometry, which the 3D-scanned sources of spatial training corpora already provide. We propose GPD (Geometry-Privileged Distillation), which makes geometric evidence the privilege in on-policy self-distillation (OPSD). For each question, depth, semantic, and bird's-eye-view (BEV) cues are rendered as compact text and routed to the teacher alongside the reference answer; a privileged KL, applied only to incorrect trajectories, augments GRPO, and the deployed model remains RGB-only. On the 4B backbone, GPD achieves 57.1 on VSI-Bench and 37.6 average across MindCube, SPARBench, MMSI-Bench, and ViewSpatial, outperforming both GRPO and answer-privileged OPSD across spatial reasoning benchmarks. Ablations confirm the complementarity of 3D and answer privilege, the advantage of question-conditioned routing over full-context injection, and the benefit of restricting distillation to incorrect trajectories.

Vertical-domain few-shot classification remains challenging for small language models, as limited supervision makes it difficult to acquire domain-specific decision knowledge. On-Policy Distillation (OPD) can improve teacher-guided adaptation by supervising student-generated rollouts, while GRPO-based reinforcement learning can further refine downstream predictions. However, existing KD-to-RL pipelines typically rely on globally fixed transition schedules, ignoring that different samples may require different amounts of teacher-guided acquisition before reward-driven refinement. We propose PIVOT (Perplexity-Informed Transition Optimization), a dynamic transition framework that routes samples between OPD and GRPO according to teacher-evaluated sequence perplexity. PIVOT moves low-perplexity samples to GRPO for reward-driven refinement while keeping high-perplexity samples under OPD for continued domain knowledge acquisition. Experiments on Banking77 and HWU64 show that PIVOT consistently outperforms continued OPD and globally synchronized OPD$\rightarrow$GRPO baselines under the same number of post-warm-up student optimization steps, achieving stronger downstream performance and more stable training dynamics.

On-policy distillation (OPD) trains a student on its own generated responses using token-level teacher supervision. However, uniform weighting overlooks differences in token learning value, while existing weighting methods rely on predefined mappings from prediction signals to token weights. These mappings are not learned from the effectiveness of the resulting student updates, limiting their ability to adapt to evolving learning needs. In this paper, we propose MetaOPD, a bilevel optimization framework that jointly learns the student model and a lightweight token-weighting network. The inner objective updates the student through weighted OPD, while the outer objective optimizes the weighting network using validation loss on reference solutions after a virtual student update. Differentiating through this update connects weighting decisions to their effects on post-update performance, allowing the mapping from prediction signals to token weights to evolve alongside the student. Experiments on six mathematical reasoning and three out-of-domain datasets, covering two student scales and seven baselines, demonstrate the effectiveness of MetaOPD, with Avg@8/Pass@8 gains over OPD of 1.99/5.97 percentage points for the 0.6B student and 2.25/6.41 points for the 1.7B student.

Knowledge distillation (KD) transfers knowledge from stronger Teacher models to weaker Student models, but most methods require training the Student parameters, thereby binding the distilled knowledge to a specific architecture and checkpoint. This implicit representation is difficult to interpret or reuse across models and limits KD for API-only or costly-to-train models. This paper studies knowledge transfer for large language models (LLMs). We introduce Universal Textual Teaching (UTT), a parameter-update-free framework that distills observed Teacher-Student knowledge gaps into a textual, interpretable, and reusable natural-language artifact called Primer. Specifically, UTT first identifies representative gap cases through paired evaluations, and iteratively updates the Primer via multi-role interactions: the Student attempts each task, the Prompter turns evaluation feedback into a teaching instruction, the Teacher provides a targeted demonstration, and the Synthesizer consolidates validated lessons. Empirically, on the challenging math (Omni-MATH-2) and code generation (KernelBench) tasks, extensive results confirm the effectiveness of the method: UTT remarkably raises the Student's accuracy from 9.4% to 48.6% and Fast1 accuracy from 9% to 35% on KernelBench, while increasing mathematical reasoning accuracy from 27.6% to 51.7%. UTT also performs better than representative prompt engineering and parameter-based KD methods. Of note, UTT is shown to be generalizable across different Teachers and Students: a Primer synthesized for one Teacher-Student pair can generalize to other Students that do not participate in the synthesis.

Structured pruning reduces the deployment cost of reasoning language models, but the resulting capability degradation can hinder subsequent on-policy distillation (OPD) recovery. Because OPD relies on student-generated trajectories, pruning damage that persists after offline distillation can limit its effectiveness. We propose RECAL, Recovery-Aware Calibration, a simple plug-and-play approach that improves OPD recovery by adjusting calibration before pruning. RECAL uses forward KL between an unpruned teacher and a pruned probe to identify teacher-supported predictions disrupted by pruning, then reweights calibration statistics to guide existing pruning criteria toward preserving these predictions. Across multiple models and pruning methods, RECAL consistently improves mathematical reasoning after OPD, achieving gains of up to 16.7 percentage points on AIME, alongside improvements in most code-generation comparisons. Further analysis shows that RECAL reduces residual damage at heavily affected tokens and establishes performance advantages that persist through recovery. These results demonstrate the value of recovery-aware calibration for improving on-policy distillation recovery of pruned reasoning models.

On-policy distillation (OPD) trains a student model by aligning its policy with a teacher model on trajectories generated by the student model itself. Through this process, the student policy moves toward the teacher on the prompts used for distillation. However, these prompts are often private and costly, creating a need for prompt-level membership auditing. Existing methods mainly rely on likelihood-based confidence signals or student policy drift between checkpoints, but they do not capture the teacher-induced direction of the student update. In this paper, we propose Policy Alignment Membership Auditing (PAMA), a new auditing framework tailored for OPD. Our key observation is that a member prompt directly contributes to the teacher-guided policy update, while a non-member prompt only experiences indirect effects through cross-prompt generalization. Based on this directional trace, PAMA measures whether the student update moves toward reducing the teacher loss on a candidate prompt. Specifically, we introduce Teacher Alignment Gain (TAG) to estimate the teacher-aligned update direction from model outputs, and further combine it with student drift and uncertainty alignment signals for reliable membership auditing. We evaluate PAMA on six datasets and three teacher-student model families. On MATH, the primary evaluation benchmark, PAMA achieves AUC values of 0.791--0.941, improving AUC by 14.6--20.6% over state-of-the-art baselines.

Reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) have become two main paradigms for post-training reasoning models. RLVR gives each response a single outcome label, leaving the steps inside it without separate credit. OPD provides token-level guidance at student-visited prefixes, but its pointwise signal does not directly reflect the pattern of teacher--student disagreement across the vocabulary. Dense, unbounded log-ratio supervision can amplify the teacher's influence, yet a strong solver is not necessarily a suitable guide when the student's solution paths depart from the teacher's. We propose Residual Advantage (\RA{}), which treats the teacher--student probability residual as a bounded one-step reward, subtracts the corresponding state value under the student policy to form a standard advantage, and centers the result within each response before adding it to the verifier advantage. The guidance term has zero mean within each response, so the verifier advantage remains the response's mean label and the teacher only redistributes credit among the steps within it. \CoRA{} further updates a teacher LoRA with verifier advantages on the same scored student batch and uses the updated teacher in the next iteration's residual, adapting guidance to the student's attempts. With Qwen3-1.7B-Base and Qwen3-4B-Base students and a Qwen3-8B teacher, \RA{} combined with GRPO or REINFORCE++ improves the underlying sequence-advantage algorithm in all 24 comparisons on three mathematical benchmarks, raising macro Avg@8 by 1.7--3.6 points and Pass@8 by 3.9--6.3 points. Both combinations surpass teacher-only OPD, and \CoRA{} adds a further 1.0--1.5 Avg@8 points.

On-policy distillation (OPD) supervises student-generated trajectories with token-level teacher signals. Its sampled-token variant avoids the cost of full-vocabulary probabilities. Yet we find that training on fewer tokens can outperform full-token OPD, challenging the intuition that more supervision improves learning. This motivates selecting tokens by learning value. Existing disagreement-based criteria ignore probability scale: tokens assigned negligible probability by both models, termed low-low tokens, can receive large log-ratio rewards and hinder learning. We propose DIAL-OPD, a token-selection method that bridges log-probability and probability spaces by weighting reward magnitude with the logarithmic mean of teacher and student probabilities. A parameter beta controls this weighting, and the highest-scoring tokens are retained. Across 4 teacher-student pairs and 7 mathematical reasoning benchmarks, we compare DIAL-OPD with 9 baselines. Retaining only 40% of tokens, it outperforms Vanilla OPD and its full-token variants, with mean accuracy gains reaching 5.25 percentage points over Vanilla OPD, and doubles AIME25 Pass@16 from 13.33% to 26.67%. It also achieves up to an 18% relative improvement in mean accuracy over the strongest token-selection baseline at matched retention ratios. With a 4B teacher, DIAL-OPD surpasses the strongest full-token baseline using an 8B teacher at both student scales, showing that effective supervision allocation can outweigh teacher scaling. Further analysis shows that moderate beta balances suppressing low-low tokens against preserving useful disagreements. Token-level evidence reveals that DIAL-OPD filters high-reward tokens with limited reasoning value while preserving supervision critical to reasoning correctness.

On-policy distillation (OPD) has become increasingly popular for transferring teacher capabilities to student models. In this work, we ask a critical research question: Is on-policy sampling always beneficial for distilling arbitrary teacher-student pairs? We show that a simple alternative, Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency. Across 17 teacher-student pairs ranging from 1.5B to 235B parameters, Semi-OPD outperforms OPD in 14 cases, with up to +13.6% accuracy and 11.4x training speedup. We further find that the choice between OPD and Semi-OPD depends on the alignment between the initial teacher and student, quantified by an output-token overlap ratio: OPD is beneficial only when the two are highly aligned with high overlap ratios. Our deeper investigation suggests that effective distillation requires on-policyness w.r.t. both the student and the teacher. For misaligned pairs, student rollouts can become increasingly off-policy w.r.t. the teacher as context length grows, weakening the distillation signal. In contrast, Semi-OPD is often more stable, as it distills on shorter contexts while covering full trajectories and exposing the student to more teacher-preferred tokens. Beyond proposing Semi-OPD as an efficient alternative, our work motivates the community to rethink when to use OPD and to study stronger OPD variants with meaningful teacher-student pairs.

Generating physically plausible videos for solid-gas dynamics is challenging as different phases exhibit distinct dynamics yet remain coupled through physical interactions. We present PAVG, a Phase-Aware Video Generator for solid-gas dynamics and interactions. It employs a dual-branch architecture to explicitly model the distinct dynamics of solids and gases, while spatiotemporal cross-attention captures their physical interactions. This design enables PAVG to preserve phasespecific motion characteristics while producing physically consistent responses across phases. To facilitate this task, we further construct a simulation corpus comprising over 700K physical trajectories across diverse solid, gas, and solid-gas interaction scenarios. Extensive evaluations demonstrate that our PAVG produces videos with improved motion adherence, physical plausibility, and visual quality compared with existing approaches.

Learning to predict how the world evolves can provide vision-language-action (VLA) policies with predictive context for long-horizon control, but its effectiveness depends on what future representation is modeled and how it conditions action generation. We introduce PLaW-VLA, which models task-relevant future states in a pretrained prediction-oriented representation space, reducing the need to predict control-irrelevant visual details. Built on a Mixture-of-Transformers architecture, PLaW-VLA conditions action generation on observation history, current task semantics, and predicted future states through structured causal attention. Experiments show a +11.8 percentage-point (pp) gain over reactive policies on RoboTwin Hard Horizon III and a +1.77 pp gain over reconstruction-oriented latent prediction on zero-shot LIBERO-Plus, supporting improved long-horizon control and generalization under distribution shift, respectively. By avoiding low-level visual reconstruction, PLaW-VLA lowers the burden of future prediction, enabling a lightweight latent world model with parallel future prediction and about 1/19 the inference latency of generative world-action modeling at comparable policy performance.

Accurate simulation of rigid-body interactions is essential for predictive physical world models. Despite recent progress in modeling object dynamics, capturing how local contacts between surfaces shape object motion remains challenging. While end-to-end world models predict interactions across entire scenes or objects, in practice, rigid-body contact is inherently local, and only nearby surfaces can directly exchange contact forces. Motivated by this observation, we introduce Rigid-body Contact Reasoning (RiCo), which represents interactions between objects through sparse neighborhoods of contact surface points. RiCo combines each point's state with the relative geometry, motion, and physical properties of nearby surfaces, then reasons across the object's points to determine how these local contacts jointly affect its motion. By confining cross-object reasoning to nearby surfaces while propagating contact information within each rigid body, RiCo retains fine-grained interaction details without the cost of modeling every pair of scene points. Such properties enable RiCo a higher accuracy and contact fidelity. Experiments on MOVi-benchmark demonstrate that RiCo reduces 100-frame position and orientation errors by 31-35% and approximately 38%, respectively, compared with baselines. Moreover, RiCo achieves high contact fidelity, with ground-truth-relative penetration-time and mean-depth differences of 11.0% and 2.22 mm, respectively. RiCo further generalizes zero-shot from small-scale training scenarios to scenes containing 270 objects. Our real-world multi-ball collision experiments further provide preliminary evidence of sim-to-real transfer.

World-action models have recently improved autonomous driving by jointly learning future scene prediction and trajectory generation. Most existing approaches model the future primarily through RGB appearance, and recent works have begun to incorporate geometric prediction to improve spatial understanding. However, dense geometry describes the spatial layout of the entire scene without indicating which parts are most relevant to the ego vehicle's action. For driving, the model must also identify and anticipate where it can safely move and which regions may pose collision risks. Jointly modeling action-relevant regions and future geometry can provide the policy with both driving-relevant cues and their corresponding spatial structure. We therefore propose AffordDrive3D, an affordance- and geometry-aware world-action model that jointly learns future action-relevant regions and spatial structure. In order to capture the scene semantics and driving context needed for driving affordance prediction, we build AffordDrive3D on a VLM backbone to forecast drivable areas and collision-critical regions that directly affect ego motion, while predicting future geometry from RGB world-model latents. On NAVSIM, AffordDrive3D achieves state-of-the-art performance with 91.3 PDMS and 89.9 EPDMS, demonstrating the effectiveness of jointly modeling future affordances and geometry for trajectory planning.

A login screen that appears after a tap on Sign in is expected; the same screen after a tap on View order is an attack. For GUI agents, safety is therefore a property of the transition rather than of the screen, and a monitor that inspects only screens can be defeated by reusing a legitimate one. Judging a transition requires an expectation of what should have followed the action. Existing GUI world models provide one, but they output it as text, code, or images, so checking it against the observed screen requires a second model to judge the two. We argue that a world model meant for verification should instead predict in the space in which observations are encoded, and present LGWM, a decoder-free, action-conditioned world model that predicts the representation of the next screen directly, trained without semantic annotation on 1.85M real GUI transitions. Verification reduces to a vector comparison, and the same signal reveals whether a mismatch is harmful. We evaluate on RSWT-BENCH, a diagnostic where each credential screen appears under both a legitimate and a hijacked transition, so detectors that see only the screen are at chance by construction. The training-free score reaches 0.987 AUC at 17 ms per decision, on par with the strongest closed-source VLMs and about ten AUC points above generative GUI world models at over three orders of magnitude lower latency. The residual direction reaches 0.953 AUC at separating harmful from benign violations, where prompted VLMs are near chance. Further analyses show that the prediction is a usable future state rather than an anomaly score. World models have mostly served as simulators or planners; our results point to a third role, verification, for which predicting in representation space is the natural design.

To stream long videos while maintaining visual quality and temporal consistency, Self Forcing mitigates exposure bias through self-rollout training on self-generated histories with key-value (KV) caching. To keep memory manageable, it detaches historical caches, preserving forward dependencies between chunks but severing the backward gradient paths. We introduce Connected Self Forcing, a training framework that reconnects gradient paths across autoregressive chunks, allowing feedback from later predictions to guide how earlier context is generated. These connections go beyond historical KV-writing: gradients pass through generated latents into the computations that produced them, linking the generation of earlier context to its use in later predictions. To make this connected training memory-efficient, we develop shortcut gradient replay, which recovers cross-chunk gradients without retaining the full rollout computation graph. Integrated with distribution matching distillation, Connected Self Forcing trains historical chunks according to both their direct supervision and their contribution to subsequent generation. Experiments on autoregressive video generation show improvements in long-horizon visual quality and temporal consistency, without changing the inference procedure.