Tool using agents improve efficiency by focusing on key behaviors
Robust Soft Skill Adaptation via Behavioral Localization in Tool-Using Agents
Artificial Intelligence
Summary
Tools that use text instructions to guide their actions can be slow when those instructions are long. The paper finds that the important parts of those instructions mostly affect just a small part of the agent’s behavior. The authors introduce a way to find and focus on these key moments, which helps the agents learn better and faster from shorter instructions. Testing showed this method improved success rates compared to using the whole behavior or other learning methods.
What this means in practice
- •For robotics engineers: Improve robotic tool use by identifying and training only key behavior moments to reduce computational effort and enhance task success.
- •For automation developers: Build more efficient soft prompt-based control systems by focusing supervision on small behaviorally important agent trajectories rather than full task execution paths.
Authors
Zijian Jiang, Xilu Wang
Abstract
Text skills can extend the capabilities of tool-using agents, but long skill descriptions increase computational costs. Soft prompts provide a compact alternative, yet learning from entire successful trajectories may also introduce unnecessary behavioral changes. We find that the useful effect of a text skill is highly concentrated in a small subset of trajectory positions. Based on this observation, we propose Behavioral Localization, which compares the frozen agent with and without the text skill to identify positions where the skill both increases the likelihood of successful tokens and meaningfully changes the next token distribution. We use these localized positions to supervise a short soft prompt. Experiments on SpreadsheetBench with a frozen Qwen3.6-35B-A3B agent show that the top 10\% of trajectory positions capture 96.12\% of the total positive skill gain. In the shared prompt setting, selective supervision improves success from 30.36\% with full trajectory to 36.79\%. Under task-specific conditions, using the top 10\% localized positions achieves 57.38\%, outperforming both the full text skill and the on-policy OPCD baseline. These results show that a small set of behaviorally relevant trajectory positions can provide effective supervision of soft skills.