HiSkill: Empowering LLM Agents with Hierarchical Skill Graphs

TL;DR

HiSkill enhances LLM agents with hierarchical skill graphs, achieving a 17.33% success rate increase and 78.75% reduction in inference token consumption.

cs.AI 🔴 Advanced 2026-07-28 2 views
Yu Hao Jinxuan Cai Qi Zhang Yawen Li Zhiqiang Zhang Chuan Shi Cheng Yang
Hierarchical Skill Graphs LLM Skill Reuse Inference Efficiency Interactive Tasks

Key Findings

Methodology

HiSkill organizes interaction trajectories into a directed graph with skill nodes, AtomicOp nodes, and typed edges, connecting reusable high-level skills with executable action templates and capturing decomposition, temporal transition, compatibility, support, and recovery relations. During inference, HiSkill retrieves a task-relevant subgraph and performs subgraph-guided task execution.

Key Results

  • HiSkill achieved a 17.33% increase in success rate and a 78.75% reduction in inference token consumption in ALFWorld.
  • In WebShop and ScienceWorld, HiSkill improved average scores and success rates by 22.95% and 31.45%, respectively.
  • HiSkill consistently outperformed state-of-the-art baselines across all datasets in success rate and score.

Significance

HiSkill significantly improves the efficiency and stability of LLM agents in long-horizon interactive tasks by bridging high-level skills and executable actions. It addresses the underutilization of skill relations in existing methods, offering a new perspective on skill reuse.

Technical Contribution

HiSkill introduces a novel hierarchical skill graph framework, fundamentally different from existing flat skill collection methods. It provides new theoretical guarantees and engineering possibilities by structuring skill relations to enhance task execution efficiency.

Novelty

HiSkill is the first to structure skill relations into a hierarchical graph, significantly improving skill reuse efficiency and stability. Unlike existing methods, it focuses not only on skill extraction but also on the relationships between skills.

Limitations

  • HiSkill may face performance bottlenecks when handling highly complex tasks.
  • It requires a substantial amount of historical interaction data to construct effective skill graphs.

Future Work

Future work could explore applying HiSkill in more diverse environments and optimizing its performance in complex tasks.

AI Executive Summary

In long-horizon interactive tasks, existing LLM agents often struggle to effectively reuse historical experience, leading to inefficiencies. HiSkill addresses this issue by introducing hierarchical skill graphs that organize skill nodes, AtomicOp nodes, and typed edges into a directed graph.

During inference, HiSkill retrieves task-relevant subgraphs and performs subgraph-guided task execution. Experiments show that HiSkill outperforms state-of-the-art baselines across multiple interactive environments, achieving a 17.33% increase in success rate and a 78.75% reduction in inference token consumption.

While HiSkill makes significant strides in skill reuse, challenges remain in handling complex tasks. Future research could further optimize its performance and explore broader applications.

Deep Analysis

Background

In recent years, skill reuse has become increasingly important in long-horizon interactive tasks. Existing methods often organize skills as flat collections, failing to fully utilize relations between skills, leading to a gap between high-level skills and executable actions.

Core Problem

Existing trajectory-to-skill methods often produce flat collections of skills, failing to utilize relations between skills, leading to a gap between high-level skills and executable actions.

Innovation

HiSkill introduces hierarchical skill graphs, structuring skill relations into a directed graph, significantly improving skill reuse efficiency and stability.

Methodology

  • �� Construct hierarchical skill graphs: Extract skill nodes, AtomicOp nodes, and typed edges from interaction trajectories.
  • �� Subgraph retrieval: Retrieve task-relevant subgraphs based on task relevance.
  • �� Subgraph-guided execution: Use retrieved subgraphs for task execution.

Experiments

Conducted experiments in ALFWorld, WebShop, and ScienceWorld to evaluate HiSkill's success rate and inference token consumption.

Results

HiSkill achieved a 17.33% increase in success rate and a 78.75% reduction in inference token consumption in ALFWorld. In WebShop and ScienceWorld, HiSkill improved average scores and success rates by 22.95% and 31.45%, respectively.

Applications

HiSkill can be applied to tasks requiring long-horizon interaction, such as robotic control and virtual assistants.

Limitations & Outlook

HiSkill may face performance bottlenecks when handling highly complex tasks and requires a substantial amount of historical interaction data to construct effective skill graphs.

Plain Language Accessible to non-experts

Imagine a factory where workers need to complete multiple steps to produce a product. HiSkill is like a smart system that not only records each worker's tasks but also their collaboration. When a new product needs to be made, the system quickly finds the best workers and steps based on past experience and relationships, ensuring production efficiency and quality.

ELI14 Explained like you're 14

Imagine you're playing a complex game where you need to complete many tasks to win. HiSkill is like a super helper that not only remembers every step you've taken but also knows which steps can be done together. When you face a new task, it helps you find the best way to complete it quickly, so you can win faster!

Glossary

Hierarchical Skill Graph

A structure that organizes skill nodes, AtomicOp nodes, and typed edges into a directed graph.

Used to organize skill and action relations in interaction trajectories.

AtomicOp

Templates for executable actions, representing specific operational steps.

Serves as basic nodes in skill graphs, used for skill decomposition.

Skill Node

Represents reusable high-level skills, composed of multiple AtomicOp nodes.

Used in skill graphs to represent complex operation sequences.

Typed Edge

Represents relations between skills and actions, such as decomposition, transition, and compatibility.

Connects skill nodes and AtomicOp nodes, capturing their relationships.

Subgraph Retrieval

The process of retrieving subgraphs from skill graphs based on task relevance.

Used during task execution to select appropriate skills and actions.

Open Questions Unanswered questions from this research

  • 1 How to effectively apply HiSkill in more complex tasks remains to be explored.
  • 2 Further research is needed to reduce the data required for skill graph construction.

Applications

Immediate Applications

Robotic Control

HiSkill can help robots complete complex tasks more efficiently, improving production efficiency.

Long-term Vision

Virtual Assistants

HiSkill can be used to develop smarter virtual assistants, offering more personalized services.

Abstract

Skills have become an important abstraction for enabling large language model (LLM) agents to reuse past experience in long-horizon interactive tasks. However, existing trajectory-to-skill methods often produce flat collections of high-level textual skills that are stored and retrieved independently, leaving skill relations underutilized and maintaining a gap between high-level skills and executable actions. In this paper, we propose HiSkill, a hierarchical skill graph framework that organizes interaction trajectories into a directed graph with skill nodes, AtomicOp nodes, and typed edges. Specifically, the graph connects reusable high-level skills with executable action templates, while also capturing decomposition, temporal transition, compatibility, support, and recovery relations among them. At inference time, HiSkill retrieves a compact task-relevant subgraph and performs subgraph-guided task execution, where a symbolic task state, an active skill, and the retrieved subgraph guide the LLM agent to switch skills, select AtomicOps, and ground executable actions iteratively. Experiments on three interactive environments show that HiSkill outperforms state-of-the-art baselines while reducing inference token consumption, demonstrating the effectiveness of bridging high-level skills and executable action grounding through a hierarchical skill graph. Our data and code is available at https://github.com/BUPT-GAMMA/HiSkill.

cs.AI