#01Oct 15, 2026
cs.CV
MMSearch-OPD: Easy Made Hard, Hard Made Learnable for Long-Horizon Multimodal Search Agents
Huanyao Zhang, Jiepeng Zhou, Jiaye Lin and 9 more
Multimodal deep search is moving from short queries to long-horizon investigations that continually acquire evidence. However, training such agents remains constrained by limited challenging training questions and insufficient process feedback. We introduce MMSearch-OPD, combining controllable data refinement with joint reinforcement learning (RL) and on-policy distillation (OPD). Our Refine pipeline performs targeted expansion of reasoning chains derived from existing simple multimodal questions. Parameterized transformations control search depth and cross-modal evidence requirements while reusing existing image--question pairs. For training, answer rewards and lightweight behavior penalties guide RL, while the reward gap modulates token-level teacher guidance on imperfect student trajectories. An exponentially decaying distillation coefficient reduces teacher dependence as training proceeds. Experiments show gains from data refinement and further improvements from joint training after SFT, indicating stronger long-horizon multimodal deep search capabilities.
#02Oct 15, 2026
cs.CV
Which Tokens to Merge? Diffusion Dynamics for Efficient Image Generation
SeungJu Cha, Ye-Chan Kim, HyunGee Kim and 2 more
Token merging has emerged as an effective training-free strategy for accelerating diffusion models by reducing redundant token computation while minimizing generation degradation. A key challenge is determining which tokens can be safely merged and which should be preserved for subsequent computation. Existing method addresses this by preserving prompt-aligned foreground regions, thereby protecting semantically relevant structures. However, this text-relevance criterion overlooks background objects, local details that are not explicitly mentioned in the prompt, leading to degraded structural coherence. In this work, we propose Dynamics-aware Token Merging (DaTMe), a training-free framework that grounds token merging in the underlying diffusion dynamics. Our key insight is that token importance should reflect whether a token remains unresolved during denoising and whether it carries structural information. To this end, DaTMe constructs a dynamics- aware importance map from the Tweedie denoised estimate. The map combines a temporal signal, measuring inter-step changes in the predicted clean image, with a structural signal, capturing local high-frequency content such as boundaries and fine details. Using this map, DaTMe adaptively assigns tokens into merging roles under a given compression ratio: important tokens are preserved as anchors, while tokens that are both temporally stabilized and structurally homogeneous are selected as safe merge candidates. Experiments on PixArt-a and FLUX demonstrate that DaTMe maintains visual fidelity across the entire image, highlighting diffusion dynamics as an effective token-level merging decision in efficient generation. Our code is available at https://github.com/SeungJuCha/DaTMe.
#03Oct 15, 2026
cs.CV
Grounded Human-Attributed Description and Activity Recognition in Videos
Jay N. Paranjape, Jurijs Nazarovs, Rohan Sarkar and 5 more
Understanding characteristics and activities of individuals in complex multi-person environments is crucial for real-world applications. However, existing datasets often simplify this problem to grounded group activity recognition, single-person activity recognition, or close-set activity recognition, thus limiting the generalizability of models trained on these datasets. In this work, we introduce the task of Grounded Human-Attributed Description and Activity Recognition (GHADAR), which involves describing the characteristics and activity descriptions of \textbf{every} person in a video, provided the location of the person in the video, which is more practical for real-world applications. To facilitate this, we introduce a new dataset derived from AVA-Actions by generating open-set captions for the person description and activity. In addition, we propose a novel method to effectively utilize the information contained in grounding during training by constraining the cross-attention masks during training in VLMs to improve performance for this task. Our experiments show that our method outperforms SOTA VLMs on this task. Finally, we demonstrate the limitations of existing evaluation metrics, which are overly reliant on human-annotations and exact text-text matching. As an added video-based evaluation, we propose a holistic VLM-based evaluation schema that compares concepts {\em directly} between the video and the generated predictions. Thus, in this work, we develop a complete framework for GHADAR, including a dataset, a novel method and an evaluation schema, thereby establishing a strong foundation for future research in this domain.
#04Oct 15, 2026
cs.CV
Skill-Decoupled On-Policy Expert Distillation for Unified Vision-Language Models
Yuqi Xiong, Chunyi Peng, Zhipeng Xu and 5 more
Reinforcement learning (RL) can elicit specialized visual capabilities in vision-language models (VLMs), yet jointly optimizing heterogeneous objectives often introduces cross-capability gradient conflicts. Training experts independently mitigates this interference, but direct reward optimization may entangle specialized knowledge with reward-hacking biases. This paper proposes Skill-Decoupled On-Policy Expert Distillation (SCOPED), which trains capability-specific experts and extracts explicit skills through reward-guided reflection. During on-policy distillation, these skills guide expert supervision while remaining unavailable to the unified student, encouraging it to internalize procedural knowledge without explicit skills at inference time. Across eight multimodal benchmarks, SCOPED consistently outperforms the backbone and mixed-capability RL, demonstrating the effectiveness of skill-decoupled on-policy expert distillation for integrating independently acquired capabilities. Further analyses reveal interference during joint capability optimization and suggest that skill augmentation promotes more consistent expert trajectories, greater attention to task-relevant visual evidence, and better preservation of specialized expertise in the unified student. All code and data are available at https://github.com/NEUIR/SCOPED.
#05Oct 15, 2026
cs.CV
Velox: Spatial Memory Makes World Models Efficient
Qiuhong Shen, Shizun Wang, Xinchao Wang
Video world models turn actions and camera motion into long visual rollouts, making video a natural interface for interactive simulation. Such interaction requires responsive generation, yet coherent rollouts must retain visual memory of what has already been generated. Chunked video diffusion models usually keep this context in growing KV caches, so later chunks become progressively slower. To address this bottleneck, we introduce Velox, a training-free attention router for region-level context selection using frustum slab correspondences from streaming geometry. Velox represents historical blocks as camera frustum slabs with depth intervals and obtains current-block depth through recency-priority multiwarp. For well-covered regions, frustum slab overlap softly modulates each layer's model-derived block importance. Uncovered regions retain the model-derived route without geometric bias. The fixed token-block granularity matches FlashAttention-style tiled execution and reduces the attention term that grows with rollout length. Frustum slab construction runs asynchronously with denoising on the same H200. At chunk 48, Velox accelerates attention by $4.36\times$ and denoising by $2.43\times$. On the full 158-case WBench navigation split, it remains within 0.15 API-free average points of dense LingBot-Fast-Cam. Please see our project page at https://velox-world.com/velox/ for more details.