Reproduce “HyperSkill: Self-Evolving LLM Agents via Hypergraph-Structured Skill Memory” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruiyao Xu, Tiankai Yang, Wei-Chieh Huang
HYPERSKILL is a new experiential memory framework for LLM agents that uses a hypergraph structure to represent n-ary relationships between subtasks and skills from past trajectories. The framework integrates structure-informed retrieval and evolution mechanisms, consistently improving success rates by up to +11.18% over baselines on xBench, GAIA, and WebWalkerQA benchmarks while maintaining memory compactness and quality.
Key insights
- Representing an agent's experiential memory using a hypergraph effectively captures complex n-ary relationships between subtasks and skills, which is crucial for transferring compositional planning knowledge.
- A dual-path retrieval strategy, combining procedural and semantic matching with co-occurrence-based skill ranking, significantly enhances the relevance and comprehensiveness of retrieved knowledge.
- Structure-informed memory evolution mechanisms, including quality-driven pruning and structural merging, are essential for maintaining a compact, high-quality memory that supports sustained performance improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.