Researchers have developed a new framework called Evolve on Graph (EoG) that enables large language models (LLMs) to improve their knowledge graph reasoning capabilities without requiring question-answer annotations. The system uses a "Graph Self-Play" mechanism where one LLM acts as a question proposer and another as a solver. To ensure questions are verifiable, a technique called Symbolically Verifiable Topology Synthesis samples reasoning paths on the knowledge graph and generates correct answer sets. This allows the proposer to learn to ask multi-hop questions that align with the solver's current skill level, leading to consistent performance improvements across various benchmarks. AI
IMPACT Enables LLMs to improve complex reasoning skills on knowledge graphs without costly human annotation.
RANK_REASON Academic paper detailing a new method for LLM reasoning. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- Evolve on Graph
- Freebase
- Graph Self-Play
- Hugging Face
- knowledge graph
- Symbolically Verifiable Topology Synthesis
- Wikidata
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
