Back to AI Research

AI Research

Behavioral Localization compresses agent skills around the steps that matter

Key Takeaways

  • A SpreadsheetBench study selects influential trajectory positions to train short soft prompts, reporting gains over full-trajectory supervision.
  • Behavioral Localization investigates how to compress a written skill into a short soft prompt for a tool-using agent.
  • The paper’s premise is that a long successful trajectory contains many steps that do not explain why a skill helped.
  • Training on all of them may introduce unnecessary behavioral changes.
  • ## Find where the skill changes behavior

Behavioral Localization investigates how to compress a written skill into a short soft prompt for a tool-using agent. The paper’s premise is that a long successful trajectory contains many steps that do not explain why a skill helped. Training on all of them may introduce unnecessary behavioral changes.

Find where the skill changes behavior

The paper compares a frozen agent with and without the text skill. It selects positions where the skill both increases the likelihood of successful tokens and meaningfully changes the next-token distribution.
Those selected positions supervise a short soft prompt. The approach therefore tries to retain the useful behavioral effect rather than reproduce every detail of a successful run. It is a training method for the prompt representation, not a claim that a user can shorten any instruction arbitrarily without losing information.

What the spreadsheet experiment reports

Experiments use SpreadsheetBench with a frozen Qwen3.6-35B-A3B agent. The authors report that the top ten percent of trajectory positions capture 96.12 percent of the total positive skill gain under their analysis.
In a shared-prompt setting, selective supervision raises success from 30.36 percent with full-trajectory supervision to 36.79 percent. Under task-specific conditions, the top-ten-percent selection reaches 57.38 percent and exceeds the compared full-text skill and on-policy baseline.

Keep shared and task-specific results separate

The larger task-specific figure should not be presented as the result for one general prompt across all tasks. The conditions differ, and the distinction is necessary to understand what the compressed skill has learned.
For developers studying skill adaptation, the practical idea is to inspect which decisions a skill actually changes before distilling an entire run. Evaluation should still measure held-out task success and unintended changes elsewhere in the agent’s behavior. The paper supplies evidence in one spreadsheet-agent setup, rather than a universal compression ratio for all skills or tool environments.

Comments