Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents

My synthesis will be thorough, precise, and comprehensive to ensure no critical detail is lost.

Here is the detailed summary of the paper:

---

Comprehensive Research Summary: Cross-Task Skill Transfer in LLM Agents

This research investigates the phenomenon where Large Language Model (LLM) agents can induce skills from completed tasks and subsequently reuse these induced skills for future, related tasks to enhance their overall capability. A central, unresolved question is whether this agent-induced skill transfer is reliable across different tasks. The study conducts a comprehensive and controlled investigation into how the method of skill induction dictates its subsequent transfer reliability across diverse benchmarks.

Key Findings on Skill Induction Levels and Format

The study systematically compares two primary dimensions of skill induction: task-level skills versus subtask-level skills, and two formats: text skills versus code skills.

1. Induction Level Effect: The results demonstrate a clear performance differential based on the level of decomposition used for skill induction.

• Task-Level Skills: These skills generally reduce the agent’s performance below its baseline performance when performing new tasks without memory.

• Subtask-Level Skills: Conversely, these skills consistently raise the agent's performance above its no-memory baseline on average. The subtask-level approach is shown to be superior for skill induction and subsequent transfer, with the subtask-level agent reusing skills more densely than the task-level agent across all tested benchmarks (Figure 6).

2. Skill Format Effect: When comparing skill formats, text skills transfer better than code skills at both the task-level and subtask-level induction levels.

The Role of Skill Properties and Utility Score

The researchers move beyond simply comparing induction methods by examining the intrinsic properties of the induced skills themselves: specificity (how closely a skill matches real tasks) and abstractness (how evenly its relevance spreads across a wider range of tasks).

• Individual Property Limitations: Neither specificity nor abstractness alone is sufficient to predict task success.

• The Skill Utility Score: The study proposes that the product of these two properties—defined as the **skill utility score: utility(s) = specificity(s) times abstractness(s) **—is what accurately predicts task success when skills are transferred.

• Utility Correlation: The skill utility score correlates consistently and positively with task success gains observed during cross-task transfer. Subtask-level and text skills exhibit higher skill utility scores compared to their counterparts.

• Defining a Useful Skill: A reusable skill is deemed useful only when it successfully balances these two properties: it must be specific to real tasks and abstract enough to remain relevant across many of them. This joint requirement is what the utility score captures.

Synthesis and Practical Takeaways for Practitioners

The research yields three critical, actionable takeaways for practitioners working with LLM agents:

1. Decomposition is Key: Decomposing a complex task into smaller subtasks and inducing one specific skill per subtask significantly improves cross-task skill transfer reliability compared to inducing skills at the higher, task-level granularity.

2. Format Preference: Text-based skills are demonstrably more transferable than code-based skills across both induction levels.

3. Utility as a Diagnostic Tool: The skill utility score serves as a lightweight, powerful diagnostic tool to assess the inherent memory value of a skill before it is deployed in a new task. A high-utility subset of the existing skill library will outperform a low-utility subset for both task-level and subtask-level libraries on benchmarks like KramaBench.

Conclusion

In conclusion, the study firmly establishes that subtask-level skills are superior for inducing transferable skills, and text skills are preferred over code skills. Crucially, the mechanism driving reliable transfer is not merely the presence of a skill, but its inherent quality—its ability to balance specificity and abstractness, quantified by the utility score. A useful skill is thus defined as one that maximizes this product: utility(s) = specificity(s) times abstractness(s). This provides a rigorous framework for designing, inducing, and curating reusable knowledge bases for LLM agents.

1. Skill Induction Strategy: Shift from task-level to subtask-level induction when aiming for reliable cross-task transfer, as Subtask-level skills lift the agent above the no-memory baseline on average. This improves performance by inducing skills from decomposed subtasks rather than whole trajectories, which generalize poorly to new tasks when induced at the task level.

2. Skill Format Preference: Prioritize text-based skill induction over code-based induction because text skills transfer better than code skills at both skill induction levels. This suggests that storing procedural knowledge as natural language workflow notes provides a more robust and transferable memory for agent experience.

3. Skill Utility Scoring: Implement a skill utility score, defined as the product of specificity and abstractness, to serve as a practical diagnostic of a skill memory before any new task runs. Skills with higher utility correlates consistently with task success when skills are transferred, allowing practitioners to assess memory usefulness without running new tasks.