2026-08-22 — views
Cross-Task Skill Transfer in LLM Agents: Subtasks and Text Win
Read this because Developers building autonomous agents can improve reliability by structuring memory as subtasks in text. The proposed utility score offers a pre-execution diagnostic to filter effective skills.
Study finds subtask-level induction and text formats enable reliable skill transfer in LLM agents, outperforming task-level and code-based methods.
The Challenge of Skill Transfer in LLM Agents
Large language model (LLM) agents possess the capability to induce skills from completed tasks and reuse them in future scenarios. This mechanism allows agents to grow more capable as they accumulate experience. However, in practical applications, the transfer of these induced skills is often unreliable. In some cases, retrieving and applying these skills can actively harm the agent’s performance rather than helping it. A critical open question in the field remains: under what conditions do agent-induced skills transfer reliably across different tasks?
Comparing Induction Methods and Formats
To address this, researchers conducted a comprehensive and controlled study examining how the method of skill induction shapes its transferability. The study focused on two primary axes along which existing methods differ: the granularity of induction and the format of the skill representation. Specifically, the research compared task-level skill induction against subtask-level skill induction, and text-based skill formats against code-based skill formats.
The results revealed significant performance disparities. Task-level skills mostly reduced the agent’s performance below its no-memory baseline, suggesting that broad, high-level memories can be detrimental. In contrast, subtask-level skills raised performance above the baseline on average. Regarding format, text skills transferred better than code skills. These findings highlight that the structure and representation of memory are crucial for successful cross-task application.
Specificity, Abstractness, and the Skill Utility Score
To further understand why certain skills transfer better, the study examined two complementary properties of induced skills: specificity and abstractness. Specificity measures how closely a skill matches real tasks, while abstractness measures how evenly its relevance spreads across various tasks. The researchers found that neither property alone predicts task success. However, their combined effect does. This combination led to the proposal of a new metric called the skill utility score.
This score correlates consistently with task success when skills are transferred. Notably, subtask-level and text skills received higher scores on this metric. A key advantage of the skill utility score is that it computes using only the skills and task descriptions, without requiring any task execution. This makes it a practical diagnostic tool for evaluating a skill memory before any new task is run, allowing developers to predict effectiveness without the cost of trial and error.
Practitioner note
Developers should prioritize inducing skills at the subtask level and storing them in text format to maximize transfer reliability. Implementing the skill utility score as a pre-execution filter can help identify high-value memories and discard those likely to degrade agent performance.