Text skills can broaden the capabilities of tool‑using agents, but long skill descriptions increase computational cost. Soft prompts offer a compact alternative, yet learning from entire successful trajectories may introduce unnecessary behavioral changes. We observe that the beneficial effect of a text skill is highly concentrated in a small subset of trajectory positions. Building on this, we propose Behavioral Localization, which compares a frozen agent’s outputs with and without the text skill to pinpoint positions that both raise the probability of successful tokens and meaningfully alter the next‑token distribution. These localized positions are then used 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 a shared‑prompt setting, selective supervision improves success from 30.36% (full trajectory) to 36.79%. Under task‑specific conditions, using the top 10% localized positions achieves 57.38%, surpassing both the full text skill and the on‑policy OPCD baseline. The results demonstrate that a small set of behaviorally relevant trajectory positions can provide effective supervision of soft skills.
Review: Behavioral Localization focuses supervision on critical decision points, markedly improving the efficiency and efficacy of soft prompts and offering a scalable path for large models in tool‑use scenarios.