UniSkill: Learning Actor-Aligned Skill Proposals for an Evolving Policy
UniSkill lets one policy act and propose skill edits using contrastive feedback without extra rollouts.
UniSkill trains a shared LLM-agent policy both to act in an environment and to propose skillbank edits labeled Add, Update, or No Edit. Environment rewards train the actor, while contrastive action feedback scores a proposal by how it changes the actor’s log-likelihood gap between stored successful and failed trajectories, so each candidate does not need a new rollout. Skill-edit support regularization keeps exploration when proposed skill content scores poorly. The paper reports 98.4% success on ALFWorld and 84.7% on WebShop, stable joint training, and effectiveness with a smaller policy backbone.
- A shared policy both acts and proposes Add, Update, or No Edit skillbank changes.
- Contrastive action feedback scores proposals without extra actor rollouts.
- Skill-edit support regularization keeps exploration when content scores poorly.
- Reports 98.4% success on ALFWorld and 84.7% on WebShop.
- Remains effective when the shared policy uses a smaller backbone.
From the original article
1 min read at huggingface.co
Large language model agents can improve across tasks by retaining reusable skills distilled from prior interactions. Recent work jointly optimizes task execution and skill extraction, enabling the policy and skillbank to co-evolve.
By Yifei Lu, Cheng Liu, Dianzhi Yu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.