Skills on the Fly: Test-Time Adaptive Skill Synthesis for LLM Agents

In three linesSkillTTA synthesizes task-specific textual skills by retrieving relevant training trajectories, with no parameter updates to the solver model. Evaluated on SpreadsheetBench, ALFWorld, and BigCodeBench: SpreadsheetBench improves from 0.397 to 0.505 Pass@1, BigCodeBench from 0.517 to 0.651.

Summary generated by Claude — human-verified