SCOPED: skill-decoupled expert distillation beats mixed-capability RL across 8 benchmarks
Yuqi Xiong · hf · 2026-10-16
Researchers from Northeastern University and Alibaba propose SCOPED (Skill-Decoupled On-Policy Expert Distillation), tackling cross-capability gradient conflicts when RL jointly optimizes heterogeneous visual objectives in VLMs.
- Problem: direct reward optimization entangles specialized knowledge with reward-hacking biases; training experts independently mitigates interference but makes integration hard.
- Method: trains capability-specific experts, extracts explicit skills via reward-guided reflection, and uses them in on-policy distillation so the unified student internalizes procedural knowledge without exposing explicit skills at inference.
- Results: consistently outperforms the backbone and mixed-capability RL across eight multimodal benchmarks, with analyses revealing interference during capability integration.
More from Multimodal
- OpenAI Astra tutorial shows PolyHaven-powered 3D scenes as GPT-6 docs page surfaces — jason_mayes · 2026-10-16
- Meituan LongCat's WorldCast benchmark: GPT-Image-2.5 tops at 72.8, open-source lags — meituan-longcat · 2026-10-16
- Three resources to get up to speed with FLUX 3 — ariG23498 · 2026-10-16
- Watch: Claude Opus 5.5 builds a Tesla Roadster from a single prompt — TAbrodi · 2026-10-16
- Midjourney v8.2 turns out striking indigo ink-wash art with sref 1848359068 — azed_ai · 2026-10-16
- "AI television from another dimension": robot grooms a chihuahua on air — BotLordBotflix · 2026-10-16