GraphDancer: Training LLMs to Explore and Reason over Graph-Structured Knowledge via Curriculum Reinforcement Learning
ORIGINAL / GraphDancer: Training LLMs to Explore and Reason over Graphs via Curriculum Reinforcement Learning
GraphDancer is a reinforcement learning framework that trains LLMs to interleave reasoning with function calls over heterogeneous graphs, using a graph-aware curriculum that schedules questions by structural difficulty. Trained only on academic and biomedical graphs, it outperforms same-backbone graph-agent baselines on held-out and unseen e-commerce, literature, and legal graphs; the 14B variant also surpasses baselines powered by GPT-4.1-mini and GPT-5-nano on average acros
01 ABSTRACT
The paper studies how LLMs can learn multi-hop evidence aggregation and precise function calling over heterogeneous graphs and transfer these skills to unseen domains. The authors propose GraphDancer, an RL framework with a graph-aware easy-to-hard curriculum for moderate-sized LLMs. On the multi-domain GRBench benchmark, training only on academic and biomedical graphs yields gains over same-backbone graph-agent baselines on held-out and unseen domains, with trajectory analyses showing more frequent retrieval of answer-bearing evidence and fewer invalid-protocol outcomes.
02 KEY FINDINGS
- GraphDancer teaches LLMs to interleave reasoning with function execution over graphs rather than relying only on similarity-based retrieval.
- A graph-aware curriculum samples questions from easy to hard by structural difficulty to improve RL training for moderate-sized LLMs.
- Trained only on academic and biomedical graphs, it outperforms same-backbone baselines on unseen e-commerce, literature, and legal graphs, and the 14B variant exceeds GPT-4.1-mini and GPT-5-nano powered baselines on average across unseen domains.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.CL