PeterGriffinJin/Graph-CoT

Implements an iterative reasoning framework where LLMs traverse text-attributed graphs through three-step cycles: reasoning about needed information, generating graph interactions (node lookups, neighbor traversals), and executing queries to retrieve context. Includes GRBench, a comprehensive evaluation dataset spanning five domains (academic, e-commerce, literature, healthcare, legal) with single-hop, multi-hop, and inductive reasoning tasks, plus baseline implementations for both open-source and GPT-based models with rule-based and model-based evaluation metrics.

300 stars. No commits in the last 6 months.