Robust Soft Skill Adaptation via Behavioral Localization in Tool-Using Agents
Training on selected steps raised the agent’s shared-prompt success rate from 30.36% to 36.79%.
The researchers’ Behavioral Localization method identifies where a text skill helps produce successful actions and meaningfully changes the agent’s next-token predictions. Those positions then supervise a short soft prompt while the underlying agent stays frozen. On SpreadsheetBench with Qwen3.6-35B-A3B, the top 10% of positions captured 96.12% of the measured positive skill gain. In task-specific tests, training on that subset achieved 57.38% success, beating both the full text skill and the on-policy OPCD baseline.
ArXiv · AI/CL/LG's note
The researchers’ Behavioral Localization method identifies where a text skill helps produce successful actions and meaningfully changes the agent’s next-token predictions. Those positions then supervise a short soft prompt while the underlying agent stays frozen. On SpreadsheetBench with Qwen3.6-35B-A3B, the top 10% of positions captured 96.12% of the measured positive skill gain. In task-specific tests, training on that subset achieved 57.38% success, beating both the full text skill and the on-policy OPCD baseline.
ArXiv · AI/CL/LG's note
score 4