Call Neighbours Yourself: Graph Walks with Destination-Conditioned On-Policy Self-Distillation
The paper makes neighbor retrieval an action the model learns during graph reasoning, not a fixed preprocessing step.
CNY lets an LLM inspect lightweight previews, choose graph-walk actions, and expand neighbors when it needs more evidence. Its training signal comes from “destination-conditioned on-policy self-distillation,” which scores a chosen neighbor after its content is revealed. On standard text-attributed graph reasoning benchmarks, the authors report consistent gains over fixed-context post-training baselines, with transfer to unseen graphs and one graph-level task outside training. ArXiv · AI/CL/LG's note
CNY lets an LLM inspect lightweight previews, choose graph-walk actions, and expand neighbors when it needs more evidence. Its training signal comes from “destination-conditioned on-policy self-distillation,” which scores a chosen neighbor after its content is revealed. On standard text-attributed graph reasoning benchmarks, the authors report consistent gains over fixed-context post-training baselines, with transfer to unseen graphs and one graph-level task outside training. ArXiv · AI/CL/LG's note
score 4