Skill-Decoupled On-Policy Expert Distillation for Unified Vision-Language Models

Skill-Decoupled On-Policy Expert Distillation for Unified Vision-Language Models. By Yuqi Xiong and seven others, from Northeastern University and Alibaba Group. Published on the fifteenth of October. Its code repository is worked out: about four hundred and fifty source files, with a readme and a licence.

Reinforcement learning can draw specialized visual capabilities out of vision-language models. Optimizing several heterogeneous objectives at once, though, often introduces gradient conflicts between capabilities. Training a separate expert for each capability mitigates that interference, but direct reward optimization can entangle the specialized knowledge with reward-hacking biases. Solving the first problem invites the second.

Skill-Decoupled On-Policy Expert Distillation, or SCOPED, keeps the separate experts and adds a step between them and the final model. After the capability-specific experts are trained, explicit skills are extracted from them through reward-guided reflection. Those skills, rather than the experts alone, carry what is passed on to a single unified student.

The skills are used asymmetrically during on-policy distillation. They guide the experts' supervision, but they are never made available to the student. That gap is the point. It encourages the student to internalize the procedural knowledge itself, so that at inference time it works without any explicit skills.

Across eight multimodal benchmarks, SCOPED consistently outperforms both the backbone and mixed-capability reinforcement learning. Further analyses reveal interference during joint capability optimization, the problem the method was built around. They also suggest that skill augmentation produces more consistent expert trajectories, draws greater attention to task-relevant visual evidence, and better preserves specialized expertise in the unified student.

The design separates two things that usually travel together: the knowledge an expert holds and the explicit help used to hand it on. Experts are trained apart, so their objectives do not collide. Skills shape what the experts teach, while the student learns without them. That is how capabilities acquired independently end up inside one model. All code and data are released.