RManLuo/graph-constrained-reasoning

Implements constrained decoding via KG-Trie to enforce knowledge graph structure directly into LLM token generation, eliminating hallucinations in reasoning paths. The framework uses lightweight KG-specialized models (Qwen2/Llama variants) fine-tuned for graph-grounded decoding, then applies general LLMs for inductive reasoning over multiple candidate paths. Supports efficient inference with Flash Attention and includes pre-trained weights on RoG-WebQSP and RoG-ComplexWebQuestions datasets.

238 stars. No commits in the last 6 months.