CNY-7B

CNY-7B

Smaller reference checkpoint for CNY (Call Neighbours Yourself), a reinforcement learning framework that treats neighbour acquisition on a text-attributed graph as explicit graph-walk actions and supervises those actions with destination-conditioned on-policy self-distillation (OPSD).

Paper: Call Neighbours Yourself: Graph Walks with Destination-Conditioned On-Policy Self-Distillation (EMNLP 2026), arXiv:2608.29588. The 14B counterpart is Allen-UQ/CNY-14B.

What this checkpoint is

This checkpoint is also the OPSD-on arm of the ablation in the paper, trained against an otherwise identical beta = 0 control that uses the same reward and rollouts.

Zero-shot held-out accuracy

Evaluated on the full test split of each dataset.

How the model is meant to be used

This is not a plain chat model. It expects the CNY multi-step walk prompt, in which the model observes the target node text, the label descriptions and a short preview per neighbour, then emits <walk> actions to reveal a neighbour's full text before committing to an answer. Prompting it as a standard instruct model will not exercise the learned walk policy.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Allen-UQ/CNY-7B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

For the walk environment, the prompt templates and the evaluation harness, use the code release (see the repository linked from the paper).

Limitations

CNY trains walk selection rather than a general graph reasoner. The walk policy is learned under a bounded step budget and a fixed preview format, and accuracy depends on the ego-node text being informative enough to direct the first walk.

Citation

@inproceedings{liu2026cny,
  title     = {Call Neighbours Yourself: Graph Walks with Destination-Conditioned On-Policy Self-Distillation},
  author    = {Yilun Liu and Boyu Luo and Yanran Tang and Ruihong Qiu and Zi Huang},
  booktitle = {EMNLP},
  year      = {2026}
}