Qwen SkillDuet Co-Evolves Models, Skills
SkillDuet jointly evolves LLM parameters and reusable textual skills during reinforcement learning. On Qwen3-4B, it reports 98.57% success on ALFWorld and 92.97% on WebShop, outperforming a matched SEED-API baseline.
SkillDuet’s key idea is treating skills as a living learning substrate rather than static prompt memory.
- –Local skills guide on-policy distillation while the model trains
- –Its Strength score selects skills that most change model responses
- –Global consolidation preserves useful guidance as model behavior evolves
- –Reported gains are promising, but evaluation remains limited to agent benchmarks
- –The main open question is whether retained skills generalize without becoming stale or harmful
DISCOVERED
2h ago
2026-11-13
PUBLISHED
2h ago
2026-11-13
RELEVANCE
AUTHOR
Discover AI