Task Decomposition-Guided Reranking for Adaptive Agent Skill Retrieval

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Yanping Chen, Weijie Shi, Wen Yang, Jiajie Xu
• Venue: arXiv: cs.AI
• Published: 2026-07-07
• Collected: 2026-07-08
• URL: http://arxiv.org/abs/2607.06283v1
• PDF: https://arxiv.org/pdf/2607.06283v1
• DOI: UNKNOWN

Plain-Language Brief

The authors propose a method called SkillReranker to improve the ability of agent systems to complete complex tasks by accurately selecting skills from large libraries. The method involves semantic decomposition of both tasks and skills, constructing a directed acyclic execution graph, and using a cross-encoder to score candidate skills. Experiments on two datasets show improvements in task performance, fewer environment interactions, and lower token consumption compared to existing methods.

What the Document Says

• SkillReranker is proposed as an inference-time reranking framework for adaptive skill selection.
• Semantic decomposition is performed on both tasks and skills, resulting in informative subtask and execution-state descriptions.
• Transition-state descriptions are used to construct a directed acyclic execution graph.
• The graph models intermediate task states as nodes and candidate skills as edges, establishing a structured task-skill correspondence.
• SkillReranker identifies subtask intervals and uses a cross-encoder to score candidate skills for each interval.
• Experiments on ALFWorld and ScienceWorld show improvements in task performance, fewer environment interactions, and lower token consumption compared to existing methods.

Key Concepts Mentioned

• SkillReranker
• semantic decomposition
• directed acyclic execution graph
• cross-encoder
• ALFWorld
• ScienceWorld

Claims Made by the Document

• SkillReranker effectively improves task performance.
• SkillReranker reduces environment interaction steps.
• SkillReranker lowers token consumption.

Evidence Used

• Experiments on ALFWorld and ScienceWorld with three backbone LLMs

Methods / Approach

• semantic decomposition
• construction of a directed acyclic execution graph
• use of a cross-encoder to score candidate skills

Limitations Stated by the Authors

Not stated in source

Limitations Noted by the Agent

Agent interpretation — not from the source.

The effectiveness of the method may vary depending on the specific details and complexity of the tasks and skill libraries.

Open Questions

• What are the three backbone LLMs used in the experiments?
• How does SkillReranker handle tasks that do not have clear subtasks or transitions?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text does not specify the three backbone LLMs used in the experiments. It may be necessary to refer to the original paper for further details.

Raw Abstract / Excerpt

Skill usage can significantly enhance the ability of modern agent systems to complete complex tasks. However, the growing scale of skill libraries makes accurate skill selection increasingly challenging. In real-world scenarios, ambiguous semantic matching often arises between a specific task requirement and multiple generic yet semantically similar candidate skills. Moreover, existing methods tend to overlook the dynamic influence of task difficulty and skill applicability when selecting the optimal target skill set. To address these issues, we propose SkillReranker, an inference-time reranking framework for adaptive skill selection. Specifically, we first perform semantic decomposition on both the task and skill sides, yielding informative subtask and execution-state descriptions as well as transition-state descriptions that characterize each skill's functionality. These descriptions are then used to construct a directed acyclic execution graph, where intermediate task states are modeled as nodes and candidate skills as edges, thereby establishing a structured task-skill correspondence. On this basis, SkillReranker determines whether each state node satisfies the split condition to identify subtask intervals. For each task interval, we employ a cross-encoder to perform comprehensive scoring over candidate skills and select the most suitable ones to form the final target skill set. Experiments on ALFWorld and ScienceWorld with three backbone LLMs show that SkillReranker effectively improves task performance, reduces environment interaction steps, and lowers token consumption compared with existing skill selection baselines.