Deconstructing and Inheriting: Cross-Task Skill Transfer in Large Language Model Agents
This paper systematically investigates how large language model agents acquire skills from completed tasks and reuse them in subsequent ones to continuously enhance performance, focusing on a central question: how can such skills reliably transfer across different tasks. Through controlled experiments, the authors compare two key dimensions affecting skill transfer—task-level versus subtask-level skill induction, and text versus code skill formats. The results show that task-level skills often push agent performance below the no-memory baseline, whereas subtask-level skills raise it above baseline on average, and text-based skills transfer better than code-based ones. To explain these findings, the authors introduce two complementary attributes: specificity, which measures how well a skill fits real tasks, and abstraction, which captures how dispersed a skill's relevance is across tasks. Neither alone predicts task outcomes, but together they do, leading to a proposed Skill Utility Score. This score can be computed from skill and task descriptions alone, without any task execution, offering a practical diagnostic tool for a skill memory bank before running new tasks.
Background and Context
Large language model agents gain power by归纳 reusable skills from completed tasks and reusing them later, but a persistent practical problem has gone underexplored: skills the agent learns for itself transfer unreliably across tasks and can even hurt the agent that retrieves them. This paper reframes the open question from whether agents can learn skills to the undervalued dimension of how skills are induced, then answers it through systematic, controlled experiments.
The authors' central contribution is shifting attention away from skill acquisition toward skill induction and its effect on cross-task transfer. Rather than proposing a new model, they offer an explainable, computable evaluation framework that converts intuitive observations into a quantifiable diagnostic tool. This repositions skill memory as a resource requiring active selection rather than passive storage, giving subsequent research a clear coordinate system.
Deep Analysis
The experimental design varies two independent dimensions. The first is induction granularity: task-level induction packs an entire task into a single skill, while subtask-level induction decomposes the task into finer steps and induces skills separately. The second is expression format: text skills describe abilities in natural language, whereas code skills present them as structured, machine-readable programs. Other variables are held fixed so each axis's isolated effect on transfer can be measured.
To explain the results, the authors introduce two complementary attributes. Specificity measures how tightly a skill fits real tasks, while abstraction captures how dispersed the skill's relevance is across tasks. Neither attribute alone predicts task success, but their joint effect does, which motivates the proposed Skill Utility Score that integrates both into a single computable metric.
The controlled comparisons yield several stable and counterintuitive findings. Task-level skills on average push agent performance below the no-memory baseline, meaning blind reuse of whole-task skills is actively harmful. Subtask-level skills, by contrast, raise performance above baseline on average, indicating that finer decomposition aids transfer. On the format axis, text-based skills transfer better than code-based ones, suggesting natural-language phrasing generalizes more robustly. Specificity and abstraction individually fail to predict outcomes, but their combination does, forming the empirical basis of the Skill Utility Score.
Industry Impact
For long-term memory systems, the study offers actionable engineering guidance: a skill memory bank should not be a place where experience accumulates undifferentiated, but rather a curated collection where each skill's usability is assessed. The Skill Utility Score's standout advantage is that it requires only the skill and task descriptions, needing no task execution to compute. This makes it a low-cost, front-loaded diagnostic that can screen a memory bank before a new task launches.
For the open-source community, this enables smarter skill-filtering mechanisms. For industrial deployment, it lowers the risk of performance regression that comes from reusing experience indiscriminately. For subsequent research, it charts a path from experience induction toward experience evaluation, advancing the classic question of how to learn skills into the more deployment-realistic question of how to assess and select them.
Outlook
The Skill Utility Score correlates stably with task success when skills are transferred, and subtask-level and text-based skills consistently score higher, reinforcing the finding that a skill's value depends not on how completely it was induced but on whether it both fits real tasks and avoids overdispersion into irrelevant contexts. Future work could operationalize this score as an automated gatekeeping layer in agent memory pipelines, deciding which skills to retain, prune, or re-induce. As agents accumulate larger experience repositories, such pre-deployment diagnostics may become essential to preventing the very performance degradation this study documents.