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

Source

Source name: arXiv
Source kind: paper
Authors / organization: Yiyang Feng, Biddut Sarker Bijoy, Niranjan Balasubramanian, Jiawei Zhou
Venue: arXiv: cs.AI
Published: 2026-08-20
Collected: 2026-08-21
URL: http://arxiv.org/abs/2608.20274v1
PDF: https://arxiv.org/pdf/2608.20274v1
DOI: UNKNOWN
Registry group: academic_apis_and_preprints
Evidence role: canonical_academic_index
Language: en
Country / region: UNKNOWN

Plain-Language Brief

This research study investigates how large language model (LLM) agents can transfer skills they have learned from completed tasks to perform better in new tasks. The authors compare two methods of inducing skills (task-level and subtask-level) and two formats of skill representation (text and code). Their findings suggest that subtask-level and text skills tend to improve an agent's performance, while task-level and code skills do not. The authors also propose a skill utility score that could help predict the success of transferring skills, based on the specificity and abstractness of the skills.

What the Document Says

The study focuses on large language model (LLM) agents and their ability to acquire and transfer skills.
Task-level and subtask-level skill induction methods are compared, as well as text and code skill formats.
Subtask-level skills tend to improve an agent's performance, while task-level skills generally do not.
Text skills transfer better than code skills.
The authors propose a skill utility score based on specificity and abstractness of the skills, which correlates with successful skill transfer.
Computing the skill utility score requires only the skills and task descriptions, without needing any task execution.

Key Concepts Mentioned

large language model (LLM) agents
skill induction
subtask-level
task-level
text
code
skill transfer
skill utility score

Claims Made by the Document

Subtask-level skills raise the agent's performance above its no-memory baseline on average.
Text skills transfer better than code skills.

Evidence Used

The study compares task-level and subtask-level skill induction, and finds that subtask-level skills tend to improve an agent's performance.
The study compares text and code skill formats, and finds that text skills transfer better.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Large language model (LLM) agents can induce skills from completed tasks and reuse them later to grow more capable with experience. In practice, induced skills may transfer unreliably and can even harm the agent that retrieves them. When agent-induced skills transfer reliably across tasks remains an open question. We conduct a comprehensive and controlled study of how the way skills are induced shapes their transfer across tasks. Specifically, we compare task-level with subtask-level skill induction and text with code skill formats, the two axes along which existing methods differ. Task-level skills mostly reduce the agent's performance below its no-memory baseline while subtask-level skills raise it above on average, and text skills transfer better than code skills. To further understand our findings, we examine two complementary properties of the induced skills: specificity, which measures how closely a skill matches real tasks, and abstractness, which measures how evenly its relevance spreads across tasks. Neither property alone predicts task success, but their combined effect does, which we propose as a skill utility score. The score correlates consistently with task success when skills are transferred, and subtask-level and text skills score higher. Computing skill utility only needs the skills and task descriptions but not any task execution, so our score serves as a practical diagnostic of a skill memory before any new task runs.