arXiv subjects

Jinqiao Wang

Publications and source records attributed to Jinqiao Wang.

At least 19 recordsLinked to original sources

arXiv2026

Data-Free On-Policy Distillation: How Far Can We Go Without External Data?

On-policy distillation (OPD) is increasingly applied to frontier foundation model post-training. Prior work in this area has largely focused on algorithmic advances, yet it remains unclear how much OPD depends on its training questions and, in particular, how far this dependence can be reduced. Across two representative single-teacher OPD settings, we find that training on 8 real prompts yields performance comparable to training on 17k problems, while datasets differing substantially in measured difficulty and initial distillation gap yield similar outcomes. Our analyses suggest two complementary explanations: repeated sampling could allow even a few prompts to expose substantial teacher supervision, while OPD transfers generalizable reasoning capabilities beyond dataset-specific knowledge. Building on these observations, we next propose a data-free on-policy distillation (DF-OPD) setting to investigate whether the system can supply the training questions itself, eliminating the need for external data. With 64 self-generated questions obtained without seed examples, DF-OPD yields performance comparable to full-data OPD in both single-teacher settings. This finding also holds in multi-teacher OPD: across mathematics, code, and instruction following, 1k generated questions achieve performance comparable to training on approximately 7k real post-training examples. We further explore whether OPD can operate even without explicit training questions. The experiments show that this is effective only in limited cases, where the student unexpectedly generates and answers its own questions, thereby reducing the process to an implicit form of DF-OPD. Together, these findings invite a reassessment of the role of training data in on-policy distillation. Code is available at \url{https://github.com/Ryuki661/DF-OPD}

cs.LG↗

arXiv2026

UniBYD: A Unified Framework for Learning Robotic Manipulation Across Embodiments Beyond Imitation of Human Demonstrations

In embodied intelligence, the embodiment gap between robotic and human hands brings significant challenges for learning from human demonstrations. Although some studies have attempted to bridge this gap using reinforcement learning, they remain confined to merely reproducing human manipulation, resulting in limited task performance. Moreover, current methods struggle to support diverse robotic hand configurations. In this paper, we propose UniBYD, a unified framework that uses a dynamic reinforcement learning algorithm to discover manipulation policies aligned with the robot's physical characteristics. To enable consistent modeling across diverse robotic hand morphologies, UniBYD incorporates a unified morphological representation (UMR). Building on UMR, we design a dynamic PPO with an annealed reward schedule, enabling reinforcement learning to transition from offline-informed imitation of human demonstrations to online-adaptive exploration of policies better adapted to diverse robotic morphologies, thereby going beyond mere imitation of human hands. To address the severe state drift caused by the incapacity of early-stage policies, we design a hybrid Markov-based shadow engine that provides fine-grained guidance to anchor the imitation within the expert's manifold. To evaluate UniBYD, we propose UniManip, the first benchmark for cross-embodiment manipulation spanning diverse robotic morphologies. Experiments demonstrate a 44.08% average improvement in success rate over the current state-of-the-art. Our project page is https://zhanheng-creator.github.io/UniBYD.

cs.RO↗

arXiv2026

From World Models to World Action Models: Rethinking Next-State Prediction

Predicting the next state is a core paradigm of World Models for modeling physical dynamics, emphasizing prediction fidelity. As World Models evolve into World-Action Models (WAMs), existing methods still fix the next state before training as RGB, a single latent feature, or a static combination of predefined targets, thereby constraining action learning to the inductive biases preserved by a particular representation. To address this limitation, we propose CF-WAM, a dynamic next-state prediction framework that samples visual, semantic, geometric, and interaction projections of the same future, standardizes them into a common video form, and supervises a unified WAM across these projections. The action-relevant constraints exposed by these projections accumulate across training steps, forcing WAM to capture the underlying state-transition structure that supports multiple projections of the same action-conditioned future. This dynamic mechanism also provides a natural cross-embodiment dynamics reference frame for Human and Robot learning. By jointly learning across different next-state parameterizations, heterogeneous Human and Robot experience can bypass appearance differences and directly contribute to shared state-transition learning, improving cross-embodiment generalization. Experiments show that CF-WAM improves both training efficiency and final control performance, while translating Human experience effectively into policy gains. CF-WAM achieves state-of-the-art performance on RoboCasa-GR1 with an average success rate of 82.50%, while reaching 82.65% on LIBERO-Plus and up to 84.00% in real-world evaluations.

cs.RO↗

arXiv2026

Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination

World-Action Models (WAMs) have emerged as a promising paradigm for embodied control by coupling future visual prediction with action generation. However, many existing WAMs rely on photorealistic future prediction, which incurs high inference latency and makes real-time robot deployment difficult. We introduce Efficient-WAM, a World-Action Model motivated by the idea that a compact future branch can still support effective action generation, even at lower visual fidelity. Efficient-WAM improves inference efficiency via a compact video expert transferred from WAN-2.2-5B, token-sparse video latents, and asymmetric video-action denoising that allocates fewer sampling steps to video than to actions. Instead of prioritizing visual fidelity, Efficient-WAM treats future video prediction as a compact guidance signal for action generation. Experiments on RoboTwin 2.0 and real-world manipulation tasks show that Efficient-WAM maintains strong action performance despite visibly coarse future predictions. Our 1B-parameter model achieves 98 ms per action chunk during physical deployment, over 30x faster than the Motus baseline with comparable task success.

cs.RO↗

arXiv2026

OmniMoE-VL: A Sparse Vision-Language Model with Coupled Visual-Depth Routing

Vision-language models (VLMs) increasingly use sparse mixture-of-experts (MoE) to scale language-side computation, yet visual information is typically routed only after passing through a fixed cross-modal interface. This leaves an important decision unresolved: which intermediate visual representations should be exposed to language computation for a given question? We introduce OmniMoE-VL, a sparse VLM with a coupled visual-depth routed projector. For each image-prompt pair, the projector selects a sparse set of intermediate visual depths and reuses the resulting global preference to guide both local patch fusion and dynamic visual injection into the language model. This design enables question-dependent visual access while preserving the native visual-token sequence, and complements token-level expert routing in the vision and language stacks. Across eight image-based benchmarks, OmniMoE-VL achieves an average score of 85.9 with 28B total and 9B activated parameters. Controlled comparisons show that the routed visual interface provides the dominant architectural gain, while matched route and component controls, same-image route analysis, and route interventions further support the value of coupling and question-conditioned visual access.

cs.CV↗

arXiv2026

CT-OPD: Counterfactual Trace On-Policy Distillation for Diffusion Vision-Language Models

Diffusion vision-language models generate answers by gradually resolving masked tokens, making accurate conditional prediction in partially resolved states central to post-training. Masking completed answers yields coherent contexts and targets, but prescribed masks do not reflect the model's reveal decisions. Its trajectories capture these decisions, yet their provisional visible tokens can conflict with the target response. Outcome-based reinforcement learning follows these trajectories but provides only response-level feedback, which loses contrast when sampled rewards tie. To align coherent token-level supervision with the model's reveal decisions, we introduce Counterfactual Trace On-Policy Distillation (CT-OPD), which combines completed teacher responses with trajectory masks from the current student. CT-OPD retokenizes each teacher response in the student's vocabulary and extracts unresolved-position masks at successive stages of the student's reverse process. For each mask, it discards provisional rollout values and reconstructs the partial state from the teacher endpoint, so the supervised positions follow the current trajectory while the visible context and targets remain consistent with the same response. The student is trained on these reconstructed states with its native categorical loss, and trajectories are refreshed as the model evolves. Across dense and sparse diffusion architectures, CT-OPD consistently enhances multimodal understanding and reasoning capabilities, with gains of up to 9.80 points on the nine-benchmark average. On the unified understanding-and-generation architecture, it also improves both visual understanding and image generation, showing that the same principle transfers across architectures and modalities. Ablations further attribute these gains to coherent reconstruction and current-model trajectory masks.

cs.CV↗

arXiv2026

On-Policy Distillation with Curriculum Turn-level Guidance for Multi-turn Agents

Multi-turn agents that plan, invoke tools, and interact with environments offer a promising paradigm for solving complex tasks, yet their capabilities typically rely on very large models whose inference cost is prohibitive in practice. On-Policy Distillation (OPD) is a natural recipe for transferring such capabilities to smaller students, but we find that it suffers a characteristic failure mode in this setting: small student errors compound across turns and push the trajectory out of the teacher's familiar state distribution, so the teacher's supervision becomes least reliable precisely where the student needs it most. We propose Guided On-Policy Distillation (Guided-OPD), a simple yet effective algorithm that mixes teacher- and student-generated turns within each rollout and schedules the teacher's intervention probability along a curriculum that decays to zero. Strong guidance keeps early trajectories close to the teacher distribution and is then gradually withdrawn to recover the purely on-policy regime used at inference. On ALFWorld, ScienceWorld, and WebShop, distilling Qwen3 students from a Qwen3-30B-A3B teacher, Guided-OPD yields average relative gains of 21.1\% in Score and 25.5\% in Success Rate over vanilla OPD, with larger gains on smaller students.

cs.LG↗

arXiv2026

G-ray: Ray-Level Relative Geometric Position Encoding in Multi-View Vision Transformers under Camera Heterogeneity

We study relative position encoding for multi-view vision Transformers under camera heterogeneity, including varying fields of view (FoVs) or projection models. Existing rotary relative position encodings commonly use image-plane positional coordinates, producing projection-dependent relative phases and inconsistent geometric cues for cross-projection attention. We introduce G-ray, a ray-level relative position encoding whose rotary phases are parameterized by camera-local ray angles. The same camera-local ray pair induces the same relative phase across projections, providing projection-invariant positional consistency. G-ray can be used directly or integrated with existing encodings, retaining complementary geometric cues without additional learned parameters. We validate G-ray in three host encodings, RoPE, GTA, and RayRoPE, across 3D reconstruction and novel-view synthesis (NVS). Across three heterogeneous 3D reconstruction benchmarks at 50 views, G-ray leads all six averaged metrics and reduces mean pointmap relative error by 45.8% over MapAnything, with calibration supplied to both. Trained exclusively on homogeneous pinhole images, the 3D reconstruction model handles mixed pinhole and non-pinhole inputs without retraining and remains competitive on homogeneous pinhole 3D reconstruction protocols. For NVS, GTA and RayRoPE improve with G-ray under joint viewpoint and FoV variation. The project's webpage is available at https://g-ray-project.github.io/.

cs.CV↗

arXiv2026

Routing Before Looking: Query-Adaptive Evidence Acquisition for Long-form Video Understanding

Long-form video understanding remains challenging for video agents due to the mismatch between query demands and evidence acquisition strategies. Although recent planning-before-perception methods outperform query-agnostic pipelines, they often rely on a single dominant strategy, either generation-based strategy or retrieval-based strategy, limiting their ability to handle diverse query demands. We propose Route2Look, a lightweight and model-agnostic framework for query-adaptive evidence acquisition in long-form video understanding. Route2Look operates in a Route-Look-Memorize loop with three tools: Global Browse for holistic context, Temporal Ground for explicit temporal cues, and Semantic Retrieve for semantic search. The core component is a routing policy that dynamically selects evidence acquisition tools based on the query. To build this policy, Route2Look adopts a two-stage design: first distilling the routing skill from differential contrastive analysis between generation-based and retrieval-based trajectories, and then applying the distilled skill with hard routing rules and continue-or-stop criteria during inference. Experiments on challenging long-video benchmarks show that Route2Look achieves state-of-the-art performance while maintaining strong frame efficiency across datasets and query types. Oracle routing analysis further reveals the potential of query-adaptive evidence acquisition for future long-form video understanding.

cs.CV↗

arXiv2026

ResMerge: Residual-based Spectral Merging of Large Language Models

Model merging offers a training-free way to combine multiple post-trained expert models, but merging experts obtained through reinforcement learning (RL) remains challenging. Existing spectral merging methods often assume that leading singular directions contain the main task signal, while lower-energy residual components can be compressed, selected, or attenuated to reduce interference. We find that this assumption does not hold for RL task vectors: after decomposing each task vector into a leading spectral head and a residual component, both parts can independently recover substantial behavior knowledge, while exhibiting different merging properties. The head is highly concentrated and informative but more prone to sharp cross-expert conflicts, whereas the residual component is more dispersed and provides a more stable basis for aggregation. Based on this observation, we propose ResMerge, a residual-based spectral merging framework for RL experts. ResMerge first constructs a stable residual backbone with Spherical Residual Consensus Adaptation, which estimates a reliability-weighted consensus direction on the Frobenius sphere. It then reintroduces leading-head information through a Lightweight Head Correction module gated by positive cross-expert agreement. Experiments across multiple RL expert groups and capability domains show that ResMerge better preserves expert capabilities than representative task-vector and spectral merging baselines.

cs.CL↗

arXiv2026

Continual Learning in Transition

Classical continual learning (CL) has primarily focused on enabling models to update and retain knowledge through parameter-centric mechanisms, e.g., training strategies, architectural designs, and weight adaptation. However, emerging paradigms are reshaping the scope of CL beyond this traditional model adaptation view. For instance, on-policy learning broadens the space of update mechanisms; test-time training extends CL from the training phase to inference; and external harness components such as memory, skill libraries, and interaction protocols extend the evolutionary boundaries of model capabilities far beyond the static parameter space. Collectively, these developments indicate a transition from parameter-centric learning toward system-level adaptation. To characterize this transition, we examine the evolution of continual learning through three dimensions: When, How, and Where learning occurs. The How dimension encompasses off-policy, on-policy, and beyond-gradient optimization mechanics. The When dimension captures evolution across pre-training, post-training, and inference-time stages. The Where dimension delineates updates occurring within internal parameters versus external structural constraints. Anchored by this tri-axial framework, we systematically survey representative methods, trace the ongoing transition of continual learning, and discuss the key challenges, broader implications, and future directions arising from this paradigm shift.

cs.LG↗

arXiv2026

HarnessWAM: Bridging Prediction and Deliberation in World Action Models

World Action Models (WAMs) jointly learn environmental dynamics and robot actions, introducing priors over physical evolution into embodied control. However, finite-horizon prediction and action generation are insufficient for complex embodied tasks that require global planning, cross-stage state maintenance, execution verification, and failure recovery. We refer to this mismatch as the prediction-deliberation gap of WAMs. To address this gap, we propose HarnessWAM, an agentic framework for WAMs. HarnessWAM employs a vision-language-model-based Task Manager to maintain an evidence-grounded scene belief and a structured task graph. A capability-conditioned executable-space projection further constrains open-ended semantic plans into sequences of atomic skills that satisfy task dependencies, embodiment-state constraints, and the capability boundary of the underlying WAM. During execution, HarnessWAM operates through an event-driven, dual-timescale feedback loop: a lightweight progress estimator continuously provides high-frequency execution evidence, while the Task Manager deliberates at salient milestones by jointly considering the current observation, task state, and interaction history to determine whether to advance the task, acquire additional observations, revise the plan, or initiate local recovery. This mechanism enables the robot to recover its state after a subtask failure and resume execution without discarding previously acquired scene knowledge. HarnessWAM achieves state-of-the-art full-task and subtask success rates of 59.6% and 69.9% on RoboMemArena, and an SR of 23.7% on RoboCerebra Ideal. These results demonstrate that model-external structured state maintenance and closed-loop agentic decision making can effectively extend the local control capabilities of WAMs into embodied task execution that is plannable, verifiable, and recoverable.

cs.RO↗

arXiv2026

Decoupled Similarity for Task-Aware Token Pruning in Large Vision-Language Models

Token pruning has emerged as an effective approach to reduce the substantial computational overhead of Large Vision-Language Models (LVLMs) by discarding less informative visual tokens while preserving performance. However, existing methods typically rely on individual attention sources from different LVLM components, resulting in incomplete and suboptimal pruning decisions due to biased attention distributions. To address this problem, we propose DeSAP, a novel Decoupled Similarity-Aware Pruning method for precise, task-aware token pruning within the visual encoder. Specifically, DeSAP introduces a decoupled similarity to capture fine-grained cross-modal relevance between visual features and text tokens, providing explicit task-related guidance for pruning. By integrating decoupled similarity with visual saliency signals derived from visual attention, DeSAP performs token pruning under the guidance of both task-related and visual cues, enabling robust pruning even under aggressive pruning ratios. Extensive experiments across diverse benchmarks and architectures show that DeSAP consistently outperforms SOTA methods in both accuracy and efficiency. On LLaVA-1.5-7B, DeSAP achieves a 10 times FLOPs reduction and a 2.3 times prefill speedup by retaining only 11.1% of visual tokens, while maintaining 98.1% of the original performance.

cs.CV↗

arXiv2026

DASH: Divergence-Adaptive Supervision Horizons for On-Policy Self-Distillation of Reasoning Models

Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models using automatically verifiable outcome signals, but these signals are typically sparse and at the sequence-level. On-policy self-distillation (OPSD) mitigates this sparsity by querying a privileged teacher at student-visited prefixes and providing dense token-level distributional supervision. Although this dense supervision alleviates signal sparsity, we find that standard OPSD still underexploits the temporal structure of the rollout. It assigns every local divergence the same coefficient, regardless of its position or the divergence sequence in which it occurs. In on-policy autoregressive generation, the same divergence magnitude can follow different discrepancy histories, reflecting different evolutions of the mismatch between the teacher and student. Since the local scalar alone cannot distinguish these temporal contexts, standard OPSD cannot adapt its token-level weights to the realized discrepancy sequence. To address this limitation, we propose Divergence-Adaptive Supervision Horizons (DASH). DASH maps the gap between each local distillation signal and the sequence-level mean to an adaptive propagation gate and then uses these gates to control backward multi-step aggregation. By doing so, DASH adjusts token-level supervision weights according to how local divergences evolve during generation. Experiments on three mathematical reasoning benchmarks across three model scales show that DASH improves over our matched vanilla OPSD reruns on every benchmark at all three scales. DASH reuses the teacher and student distributions that OPSD already computes, so the gains require no additional teacher or student forward pass. Code: https://github.com/DBtxy/DASH-OPSD

cs.AI↗

arXiv2026

Messages, Not Tokens: Grounded Coresets for Faithful VLM Compression

Modern vision language models (VLMs) turn high-resolution images into long sequences of visual tokens. Every token traverses the language decoder and persists in its prompt KV cache, inflating inference cost and motivating aggressive visual compression. Existing score-based methods assign each token an independent importance score and retain the Top-K. However, text queries consume collective, signed attention messages from the visual population, not isolated patches. Consequently, equally sized Top-K sets can repeatedly cover one salient region, omit sparse but complementary evidence and discard information carried by the removed population. We therefore formulate faithful visual compression as constructing a compact coreset for decoder messages, and introduce our training-free Grounded Message Coreset Pruning (GMC) which jointly allocates support across query-grounded, appearance, and coordinate-aware evidence, then transports discarded states into selected representatives at their original multimodal positions before physical compaction and native attention resume. This decomposes faithful compression into two coupled components, including selecting carriers that cover the required message modes and realizing the signed population message on those carriers. We further derive bounds connecting their errors to signed-message distortion, visual innovation, and candidate-margin stability. Experiments across multiple VLM families and diverse benchmarks demonstrate strong performance, with GMC-H2 retaining 97.78% Full-relative mean capability on Qwen2.5-VL-7B using 80.2% fewer visual tokens, while GMC-L16 reaches 100.36%. Controlled interventions verify that collective support and population realization jointly drive these gains.

cs.CV↗

arXiv2026

PixVL: Self-Supervised Training of Pixel-Level MLLMs via a Unified Mask--Text Consistency Cycle

Recent studies develop pixel-level multimodal large language models (MLLMs) that support both Region Segmentation and Region Understanding, extending multimodal interaction from whole images to specific objects and regions. However, these methods face two fundamental challenges. First, the scarcity of high-quality mask--text pairs leaves abundant mask annotations without corresponding language supervision. Second, discrepancies in supervision formats and learning-signal densities induce optimization interference between Region Segmentation and Region Understanding. To address these challenges, we propose PixVL, a self-supervised post-training framework that introduces a unified Mask--Text Consistency Cycle, enabling pixel-level MLLMs to generate and self-verify regional descriptions and learn from unlabeled data. We found that direct cycle based solely on geometric reconstruction is unreliable because re-segmentation IoU does not faithfully reflect the semantic quality and referring sufficiency. PixVL therefore introduces confuser-aware semantic verification, which uses the model's confidence when it correctly chooses the target among highly similar candidate regions, and assigns zero reward to an incorrect choice. Meanwhile, PixVL performs cross-view verification using temporally separated video frames or geometrically transformed image views, preventing cyclic learning from collapsing to positional and shape shortcuts. Finally, a quality-coupled bidirectional learning strategy uses the highest-reward description to guide Text-to-Mask learning. This strategy transforms Region Understanding and Region Segmentation from competing tasks into mutual generators and verifiers. Experiments demonstrate that PixVL improves both region understanding task and segmentation task.

cs.CV↗

arXiv2026

Don't Mix Rewards, Mix Policies: Policy Decomposition and Optimization for Multi-Reward RL

Modern large language models (LLMs) are expected not just to answer correctly, but to adapt their behavior to different human values and use cases. As a result, multi-reward reinforcement learning (RL) has become an increasingly important problem for LLMs, where each reward captures a different aspect of desired behavior. However, optimizing with multiple rewards suffers from a more severe alignment tax issue, where different optimization objectives can trade off or even conflict with each other, leading to unstable and inefficient post-training. In this work, we propose PRISM, a new multi-reward RL framework built upon the idea of policy-space decomposition and composition. Instead of compositing different rewards, PRISM optimizes a set of standalone positive policies and a global negative policy. This alleviates the potential conflict during multi-reward policy optimization, while enabling controllability during inference by flexible policy composition. Experiments on scientific reasoning, tool-use reasoning, and helpfulness-safety alignment show that PRISM consistently outperforms existing multi-reward RL baselines, with extra controllability for inference-time preference control.

cs.AI↗

arXiv2026

Dichotomous Diffusion Policy Optimization

Diffusion-based policies have gained growing popularity in solving a wide range of decision-making tasks due to their superior expressiveness and controllable generation during inference. However, effectively training large diffusion policies using reinforcement learning (RL) remains challenging. Existing methods either suffer from unstable training due to directly maximizing value objectives, or face computational issues due to relying on crude Gaussian likelihood approximation, which requires a large amount of sufficiently small denoising steps. In this work, we propose DIPOLE (Dichotomous diffusion Policy improvement), a novel RL algorithm designed for stable and controllable diffusion policy optimization. We begin by revisiting the KL-regularized objective in RL, which offers a desirable weighted regression objective for diffusion policy extraction, but often struggles to balance greediness and stability. We then formulate a greedified policy regularization scheme, which naturally enables decomposing the optimal policy into a pair of stably learned dichotomous policies: one aims at reward maximization, and the other focuses on reward minimization. Under such a design, optimized actions can be generated by linearly combining the scores of dichotomous policies during inference, thereby enabling flexible control over the level of greediness.Evaluations in offline and offline-to-online RL settings on ExORL and OGBench demonstrate the effectiveness of our approach. We also use DIPOLE to train a large vision-language-action (VLA) model for end-to-end autonomous driving (AD) and evaluate it on the large-scale real-world AD benchmark NAVSIM, highlighting its potential for complex real-world applications.

cs.LG↗