Source-linked AI summary
Robust Soft Skill Adaptation via Behavioral Localization in Tool-Using Agents
Zijian Jiang, Xilu Wang
TL;DR
Long text skills increase computational costs, while learning from full successful trajectories may introduce unnecessary behavioral changes. Behavioral Localization identifies positions where a skill increases successful-token likelihood and shifts the next-token distribution, then uses those positions to supervise a short soft prompt. On SpreadsheetBench, the top 10% of positions capture 96.12% of positive skill effect, and task-specific selective supervision reaches 57.38%, outperforming Text Skill and OPCD.
Problem
Long text-skill prompts increase context use and inference cost, while soft prompts trained on full trajectories may introduce unnecessary behavioral changes.
Source: §1 · pp. 1–2Method
Behavioral Localization compares predictions with and without a text skill, selecting positions with positive gain for successful tokens and meaningful next-token distribution shifts to supervise a short soft prompt.
Source: §1 · p. 2, §3.2 · p. 3Results
57.38% under task-specific conditions with the top 10% localized positions, outperforming both Text Skill and OPCD; those positions capture 96.12% of total positive Skill effect.
Source: §4.2 · p. 4Takeaways & Limitations
A small subset of behaviorally relevant trajectory positions can provide sufficient supervision for learning effective soft prompts.
Source: §4.2 · p. 4
Abstract
from arXiv · showhide
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.
1 Introduction
The introduction presents text skills as a flexible way to adapt agents, but notes that long prompts are costly and soft prompts can introduce behavioral changes. It proposes Behavioral Localization to focus soft-prompt training on the trajectory positions where a skill matters, while limiting unnecessary drift.
1 Introduction
- Natural-language agent skills specify task goals, operational rules, and expected behaviors without changing model parameters, but lengthy prompts increase context use and inference cost.Their effectiveness can also depend on wording and example selection.Source: §1 · p. 1
- Soft prompts are compact, learnable alternatives, but their lack of clear natural-language counterparts can lead to unexpected behavioral changes.Sequential prediction shifts can propagate into incorrect tool selection, malformed outputs, or task failures, and gaining capabilities may degrade existing abilities.Source: §1 · pp. 1–2
- Behavioral Localization selects trajectory positions where a text skill meaningfully changes predictions, then uses them to train a compact soft skill while constraining selected positions toward the base distribution.The approach targets skill-useful behavior and reduces unnecessary behavioral drift.Source: §1 · p. 2
- The skill-induced behavioral effect is concentrated at a small number of decision points, while the base agent already handles much of the remaining trajectory correctly.Token-level distributional analysis finds that skill-induced behavioral changes are highly sparse.Source: §1 · p. 2
- Training a soft prompt on selected Behavioral Localization positions outperforms training on the full trajectory.The introduction reports this comparison without giving benchmark values in the supplied passages.Source: §1 · p. 2
2 Related Work
Related work studies continuous prompt representations, their reuse, and compression of natural-language contexts. For agent skills, the paper emphasizes preserving behavior and asks which skill-induced changes should supervise a compact representation.
2 Related Work
- Prompt Tuning, Prefix-Tuning, and P-Tuning adapt frozen models by optimizing a small set of continuous prompt embeddings; SPoT reuses prompts, while XPrompt removes prompt tokens that hurt downstream performance.The passage contrasts these representation and reuse studies with the paper’s focus on selecting behaviors for learning.Source: §2 · p. 2
- Gisting compresses long prompts into reusable gist tokens, while ICAE encodes long contexts into memory slots intended to preserve their semantics.These methods primarily compress or preserve textual meaning rather than target skill-induced behavior.Source: §2 · p. 2
- SoftSkill compresses textual skills into a continuous prefix using successful-trajectory next-token prediction, while LatentSkill converts them into task-specific LoRA adapters.Prior work encodes skills compactly; this paper instead studies which behavioral changes should be used to learn the representation.Source: §2 · p. 2
3 Behavioral Locator
Behavioral Locator formulates soft-skill training as finding sparse trajectory positions where a text skill matters. It scores positions by skill-related token gain and distribution shift, then selectively supervises a soft prompt while preserving base behavior elsewhere.
3.1 Problem formulation
- The formulation seeks sparse trajectory positions to train a short soft prompt replacing a shared text skill, while avoiding harmful updates elsewhere.Source: §3.1 · p. 2
- Positive gold-token gain is highly concentrated across 61 successful training trajectories under teacher forcing.Source: Figure 1 · p. 3
3.2 Method
- At each position, positive gain tracks the successful-token probability increase with skill, while JS tracks how much skill changes the full next-token distribution.The distributions are computed along the teacher trajectory under text-skill and no-skill conditions.Source: §3.2 · p. 3
- Positions are ranked by localization score and top k% selected; selected positions use full-vocabulary KL to match skill predictions, while unselected positions use no-skill predictions as behavioral anchors.Source: §3.2 · p. 3
4 Experiment
Experiments on SpreadsheetBench evaluate localized-position supervision for shared and task-specific soft prompts. The strongest reported comparison is for the task-specific setting.
4.1 Setup
- Experiments use SpreadsheetBench, GPT-5.5-generated successful teacher trajectories, and a frozen Qwen3.6-35B-A3B student with length-8 soft prompts.Evaluation uses greedy decoding; a task counts as successful only if its output passes all SpreadsheetBench checks.Source: §4.1 · p. 3
4.2 Main Result
- 57.38% is achieved by top-10% localized positions under task-specific conditions, outperforming Text Skill and OPCD.Table 1 presents the main SpreadsheetBench results; bold marks the best learned soft prompt under each setting.Source: Table 1 · p. 4, §4.2 · p. 4
- Selecting only 10% of trajectory tokens captures 96.12% of the total positive Skill effect.Source: §4.2 · p. 4
- Under the shared prompt setting, the method beats full-trajectory SoftSkill but remains below on-policy OPCD.Source: §4.2 · p. 4