SkillTTA: Adaptive Skill Synthesis Boosts LLM Agent Performance

A novel approach known as SkillTTA tackles the issue of transforming extra compute during testing into enhanced capabilities for LLM agents. Rather than merely increasing context or incorporating rollouts, SkillTTA focuses on retrieving training trajectories pertinent to the task and creates a temporary skill based on the observable target context, utilizing a solver with fixed parameters. To boost performance further, it incorporates meta prompt optimization (MPO) to refine the policy responsible for generating these skills. MPO assesses candidate prompts across paired tasks, prioritizes informative transitions, and restricts updates to benchmark-specific atomic slots, minimizing variance from indirect observations. SkillTTA surpasses leading reuse methods across benchmarks such as ALFWorld, SpreadsheetBench, BigCodeBench, and WebShop.