Skill-Decoupled On-Policy Expert Distillation for Unified Vision-Language Models
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 proce...
| Authors | Yuqi Xiong, Chunyi Peng, Zhipeng Xu et al. |
|---|---|
| Primary category | cs.CV |
| Published | 15 October 2026 |
| Last updated at source | 15 October 2026 |
Recent changes
- 16 Oct 2026 Description: 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 mi → 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 mi
- 16 Oct 2026 Topics added cs.CV, cs.AI removed hf-daily-papers, reinforcement learning, vision-language models, cross-capability gradient conflicts, expert distillation, reward-guided reflection, on-policy distillation, skill-decoupled, multimodal benchmarks
Changes detected by Silytics between refreshes of the source.
Silytics shows public metadata from arXiv and links to the original. Information may be incomplete or out of date. Last refreshed 16 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.