UniSkill: Learning Actor-Aligned Skill Proposals for an Evolving Policy
UniSkill trains the agent and its skillbank together without running fresh tests for every proposed skill edit.
The paper proposes a shared policy that both acts in the environment and suggests Add, Update, or No Edit changes to stored skills. Its feedback checks whether a proposed skill better aligns the current actor’s action likelihoods on prior successful versus failed trajectories from the same task. The authors say this avoids extra rollouts for each proposal while keeping joint training stable. Reported results are 98.4% success on ALFWorld and 84.7% on WebShop. HF Daily Papers' note
The paper proposes a shared policy that both acts in the environment and suggests Add, Update, or No Edit changes to stored skills. Its feedback checks whether a proposed skill better aligns the current actor’s action likelihoods on prior successful versus failed trajectories from the same task. The authors say this avoids extra rollouts for each proposal while keeping joint training stable. Reported results are 98.4% success on ALFWorld and 84.7% on WebShop. HF Daily Papers' note
score 4