WER Framework: Training a Skill Optimizer from Execution Feedback to Improve Tool-Using Agents

01 ABSTRACT

The paper notes that expert-written natural language skills can improve tool-using agents, yet agent-authored skills perform 8–11 points worse than using no skill; inference-time loops only repair individual skills per task and leave the underlying generator's refinement capability unimproved. The authors introduce WER, a multi-phase framework that organizes execution experience from intermediate skills into training states and trains a Skill Optimizer outside a frozen executor: the optimizer proposes skills, a frozen agent executes them repeatedly, and a programmatic verifier scores outcomes, which serve as group-relative rewards and filter for informative mixed-outcome instances; matched successful and failed trajectories form the next phase's refinement states. On BFCL v4 multi-turn and tau2-bench, WER improves average Pass@1 over the no-skill baseline by 7.80 and 6.06 points, respectively, and under an identical refinement workflow outperforms the same backbone without optimizer training by 10.02 and 12.50 points; the trained 4B optimizer reaches 76.63% on BFCL v4, outperforming all evaluated off-the-shelf general-purpose LLMs in the same role on average.

02 KEY FINDINGS

1. Agent-authored skills perform 8–11 points worse than no skill, exposing weak refinement in the generator.
2. WER trains a skill optimizer outside a frozen executor using group-relative rewards and mixed-outcome filtering from execution feedback.
3. On BFCL v4 multi-turn and tau2-bench, average Pass@1 improves over the no-skill baseline by 7.80 and 6.06 points.