Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents
This study compares task-level and subtask-level skill induction, finding subtask and text-format skills transfer more reliably, and introduces a skill utility score.
Key Findings
Methodology
The authors designed two agent types: task-level and subtask-level, combining text and code formats for skill storage. They systematically evaluated performance across three long-horizon benchmarks using 11 models. By introducing properties of specificity and abstractness, they proposed a skill utility score to predict transfer success. Experiments confirmed that subtask and text skills outperform task-level and code formats, with the utility score correlating strongly with task success, especially in complex scenarios.
Key Results
- Subtask-level skills significantly improved task success rates, with an average increase of 10-20%, outperforming task-level skills which decreased performance by 1.2-4.1 points. Text format skills transferred better than code, especially in complex tasks. The skill utility score showed a correlation coefficient above 0.75 with success rates, validating its predictive power. Results were consistent across models and benchmarks, demonstrating robustness.
Significance
This research systematically analyzes how skill induction level and format affect cross-task transfer, addressing the unreliability of previous methods. The proposed skill utility score offers a practical diagnostic tool, enabling better skill management and improving model generalization. These insights advance the development of autonomous agents capable of efficient multi-task learning, with broad implications for AI deployment in industry and research.
Technical Contribution
The paper innovatively combines task decomposition with property-based skill evaluation, introducing the skill utility score based on specificity and abstractness. It contrasts task-level and subtask-level induction across multiple formats and benchmarks, providing a comprehensive framework for understanding and predicting skill transfer. The approach offers theoretical insights and practical tools for optimizing skill storage and reuse in large language models.
Novelty
This is the first systematic comparison of task-level versus subtask-level skill induction in multi-model, multi-task settings, introducing a quantitative skill utility score based on properties of skills. Unlike prior work focusing solely on success metrics, this study reveals the internal mechanisms of skill transfer, emphasizing the importance of task decomposition and skill representation formats, thus providing a new perspective on skill management.
Limitations
- The validation is primarily in simulated environments and limited benchmarks; real-world complexity may introduce additional challenges. The skill utility score, while predictive, requires further refinement to adapt dynamically to evolving tasks. Computational costs remain high, especially for large models and multi-modal data, limiting immediate deployment. Future work should focus on scalability, real-world testing, and adaptive skill updating mechanisms.
Future Work
Future directions include integrating multi-modal skills (visual, auditory), applying reinforcement learning to dynamically select and update skills, and extending the framework to real-world applications such as robotics and industrial automation. Additionally, exploring adaptive skill management strategies to handle environmental changes and task variability will be crucial for advancing autonomous systems.
AI Executive Summary
The rapid growth of large language models (LLMs) has unlocked remarkable capabilities in automation, research, and personal assistance. However, these models traditionally solve each task independently without leveraging prior experience, limiting their efficiency and adaptability. Recent efforts introduced skill storage and transfer mechanisms, aiming to enable models to learn from completed tasks and reuse skills to improve future performance. Yet, the reliability of such transfer remains inconsistent, especially when skills are induced from entire task trajectories, which often leads to overfitting and irrelevance in new contexts.
This study systematically compares two fundamental approaches: task-level skill induction, which summarizes entire task trajectories, and subtask-level induction, which decomposes tasks into smaller units. It also evaluates two skill storage formats: natural language text and executable code. Through extensive experiments across three benchmarks—AppWorld, OfficeBench, and KramaBench—and 11 diverse models, the authors demonstrate that subtask-level and text-format skills transfer more reliably and effectively. They introduce a novel skill utility score, based on properties of specificity and abstractness, to predict the success of skill transfer without executing tasks.
Results show that subtask and text skills consistently outperform task and code formats, with success rates improving by up to 20%. The skill utility score correlates strongly with actual task success, serving as a practical diagnostic tool. These findings highlight the importance of task decomposition and skill representation choices in designing robust multi-task agents. The research offers valuable insights for future AI systems, emphasizing scalable, interpretable, and transferable skill management.
Despite these advances, challenges remain in deploying these methods in real-world, dynamic environments. The computational costs and the need for adaptive skill updating are notable limitations. Future work should explore multi-modal skills, reinforcement learning integration, and real-world validation to further enhance autonomous capabilities. Overall, this work significantly advances understanding of skill transfer in LLM agents, paving the way for more adaptable and intelligent systems.
Deep Analysis
Background
近年来，随着大规模语言模型（如GPT-4、PaLM）在多任务、多领域中的应用不断扩大，研究者开始关注模型的知识积累与迁移能力。早期工作主要集中在模型单任务性能优化（如Fine-tuning、Prompt Engineering），逐步引入技能存储机制（如 Wang et al., 2024a; Sharma et al., 2022），旨在实现经验的积累与复用。技能迁移的核心目标是让模型在面对新任务时，能快速调用已有技能，减少训练成本，提高泛化能力。然而，实际应用中，技能迁移的效果常受限于技能的过度专属和不匹配，导致迁移效果不稳定甚至负向迁移。现有研究多集中在任务级技能（Zhou et al., 2025），但对子任务级技能的系统性分析不足，存储格式（文本与代码）也缺乏统一评估标准。这些问题限制了技能迁移的实用性和可扩展性。
Core Problem
核心问题在于，当前技能诱导多依赖完整任务轨迹，导致技能过度拟合源任务，迁移到新任务时表现不佳。如何设计既能保持技能实用性，又能增强其跨任务泛化能力的方法？此外，缺乏有效的指标衡量技能迁移效果，难以指导技能存储与选择策略。这些问题在多任务、多模型环境中尤为突出，阻碍了自主智能体的长远发展。解决方案需要在保持技能表达丰富性的同时，提升其抽象性和匹配度，确保迁移的可靠性。
Innovation
本文提出了两大创新：一是系统比较任务级与子任务级技能诱导，揭示子任务级在迁移中的优势；二是引入技能特异性与抽象性两个属性，结合成技能效用评分指标，能在不执行任务的情况下预测迁移效果。这一指标突破了以往只关注任务成功率的局限，为技能管理提供了量化工具。研究还首次在多模型、多基准环境中验证了方法的普适性，推动了技能迁移理论的深化。
Methodology
- �� 设计两类智能体：任务级（单轨）与子任务级（多轨），结合不同技能存储格式（文本、代码）进行比较。• 采用三大长远基准（AppWorld、OfficeBench、KramaBench）测试模型表现。• 引入技能特异性（匹配度）与抽象性（分布均匀性）两个属性，计算技能效用得分。• 通过多模型、多任务实验，验证技能诱导层级与格式对迁移效果的影响。• 利用技能效用评分预测迁移成功，进行指标验证与实用性分析。
Experiments
采用11个模型，覆盖不同规模和架构，评估在三大基准上的任务成功率。对比无技能、文本技能、代码技能在任务级与子任务级的表现。设置不同预算（依赖、延迟）条件，分析迁移效果的稳定性。进行多轮AB测试，验证技能效用指标的预测能力。实验还包括不同任务难度、模型规模的子集分析，以确保结论的普适性。
Results
子任务级技能显著提升任务成功率，平均提升10-20%，优于任务级技能，后者平均下降1.2-4.1个百分点。文本格式技能迁移效果优于代码格式，尤其在复杂任务中表现更佳。技能效用评分与实际成功率高度相关，相关系数达0.75以上，验证了指标的有效性。多模型、多基准实验一致支持这些结论，表明方法具有良好的泛化能力。
Applications
该研究为多任务智能体设计提供指导，特别适用于自动化助手、科研工具等场景。通过技能效用评分，开发者可以在模型部署前评估技能存储的实用性，优化技能诱导策略。未来，结合多模态数据和强化学习，有望实现更智能的技能管理体系，推动自主系统在工业和科研中的应用。
Limitations & Outlook
当前方法主要在模拟环境和有限基准上验证，实际复杂场景中效果可能受限。技能效用指标虽有效，但仍需结合动态环境调整。模型规模和计算成本较高，实际部署存在资源压力。未来需优化指标的适应性和效率，扩展到多模态、多任务环境中。
Plain Language Accessible to non-experts
想象你在厨房做菜，每次做完一道菜，你会记住用的材料、步骤和技巧。以后做类似菜时，你可以直接用之前的经验，省时省力。但如果你只记住整份菜的做法，遇到不同的菜就不太管用。这个研究就像在厨房里，把做菜的技巧拆成小步骤，存成便于复用的笔记或程序。这样，无论做什么菜，只要用这些拆分的小技巧，就能更快做好。研究发现，拆成小步骤的技巧更容易在不同菜肴间迁移，比记住整份菜的做法效果更好。还提出了一个“技巧评分”，就像给每个技巧打分，告诉你这个技巧在不同菜中用得是否合适。这样，厨师（模型）（像你）就能提前知道哪些技巧能帮忙，哪些可能会出错，做菜更高效。这就像给厨师装上了智能助手，帮他挑选最实用的技巧，厨房变得更智能、更快。
ELI14 Explained like you're 14
想象你在玩一款游戏，每次打boss都要用不同的技能。有时候，你会记住整个战斗的流程，但这样只适合那次战斗，换个boss就不管用了。更聪明的方法是，把每个战斗拆成几个小任务，比如躲避、攻击、回血，然后记住每个小任务的技巧。这样，不管遇到哪个boss，你都可以用这些技巧组合起来，打得更顺利。这个研究就像在游戏里，把技能拆成小块，存成便于复用的“秘籍”，让你在不同战斗中都能用得上。研究发现，用拆分的小技巧比记住整套流程更有效，特别是在复杂的任务中。而且，他们还设计了一个“技巧评分”，就像给秘籍打分，告诉你哪些秘籍在不同战斗中都管用，哪些可能会出错。这样，你就可以提前知道用哪个秘籍最靠谱，打游戏变得更轻松、更厉害！
Glossary
Skill Induction (技能诱导)
将完成任务或子任务的轨迹转化为可存储的技能，用于后续任务。技术上包括从轨迹提取描述或代码，存入技能库。
论文中描述的技能存储与提取机制。
Skill Utility Score (技能效用评分)
结合技能的特异性和抽象性两个属性，量化技能在跨任务迁移中的实用价值。
用以预测技能迁移成功率的指标。
Subtask-level (子任务级)
将复杂任务拆分为多个子任务，分别诱导技能以增强迁移能力。
论文中提出的技能诱导层级。
Text Skills (文本技能)
用自然语言描述的技能流程或注意事项，便于理解和检索。
存储格式之一。
Code Skills (代码技能)
用Python函数等代码形式表达的技能，便于自动调用和执行。
存储格式之一。
Open Questions Unanswered questions from this research
- 1 如何在实际复杂环境中持续优化技能的抽象性与特异性，以适应不断变化的任务需求。
- 2 多模态技能（如视觉、声音）在迁移中的表现与优化策略。
- 3 结合强化学习动态调整技能存储与选择机制的研究方向。
Applications
Immediate Applications
智能助手技能管理
通过技能评分评估存储的技能质量，优化多任务助手的技能库，提升任务完成效率。
科研自动化工具
利用子任务级技能拆分，增强科研流程的自动化与迁移能力，减少重复调试时间。
Long-term Vision
自主系统的泛化能力
实现跨领域、多模态、多任务的技能迁移，推动自主系统在工业、医疗等行业的广泛应用。
Abstract
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.