The authors trained independent vision-language experts via reinforcement learning on separate datasets for capabilities such as fine-grained perception, visual grounding, and complex reasoning.
Decoupling skills from expert policies improves multi-capability vision-language model training
Authors show that guiding expert supervision with explicit skills reduces gradient conflicts and yields consistent gains across eight benchmarks.
Chinese Tech
Yuqi Xiong · Chunyi Peng · Zhipeng Xu · Qing Yang · Zhenghao Liu · Zulong Chen · +2 more
Northeastern University · Alibaba Group
Research Digest··2 min read
The authors propose SCOPED, a framework that trains separate capability-specific experts via reinforcement learning, extracts explicit skills from their reasoning trajectories, and uses those skills to condition the experts during on-policy distillation into a unified student vision-language model.
Why this paper
From Alibaba Group and Northeastern University · Part of Reasoning Distillation Alignment, now 31 papers
In one line
SCOPED trains capability-specific experts and distills explicit skills into a unified VLM, outperforming mixed-capability RL.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§