Research
SkillDuet: Co-Evolving Models and Skills through Response Feedback
SkillDuet is a framework that jointly updates an LLM's parameters and evolves reusable textual skills extracted from on-policy trajectories. A response-based score called Strength ranks candidate skills by how much they shift the current model's responses, with top skills periodically consolidated into a persistent global skill set. Using Qwen3-4B, it reports 98.57% success on ALFWorld and 92.97% on WebShop, exceeding the matched SEED-API baseline by 2.86 and 5.47 percentage points. Ablations indicate that response-based skill selection and global consolidation improve task performance.
LLM agentsreinforcement learningskill acquisition