EvoGuard improves agent skill evolution by 11% on ALFWorld
Original title: EvoGuard: Mitigating Degradation in Online Skill Evolution
EvoGuard uses pairwise rejection sampling on private test sets to select non-degrading skills at low latency.
Existing online skill evolution methods can degrade performance over time. EvoGuard builds skill pools, scores them on private test sets, and applies rejection sampling to retain pairs with clear utility gaps. The selector is trained on logits to pick better skills in a single forward pass, improving success by 11.0% on ALFWorld, 5.9% on ScienceWorld, and 2.7% on LiveCodeBench.
- Improves success by 11.0% on ALFWorld
- Improves 5.9% on ScienceWorld
- Improves 2.7% on LiveCodeBench
- Decides in single forward pass after training on label token logits
From the original article
1 min read at arxiv.orgSkills let language agents reuse experience across tasks as natural-language documents. Recent work evolves skills from interaction, but existing methods have limitations when skills evolve online. Prior methods update a skill offline from a batch of trajectories collected over many tasks.
By Yanli Wang · arxiv.org
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.