xiaohan-yi/GC_OPD

GC-OPD models

Graph-Conditioned On-Policy Agent Distillation from Off-the-Shelf Teachers

Paper · Code and evaluation instructions

This collection contains eight BF16 models: GC-OPD and GC-OPD+GA for ScienceWorld 1.7B/4B, ALFWorld 1.7B and WebShop 0.8B. GA uses planner/oracle executions to augment the training graph.

Environment	Student	GC-OPD	GC-OPD+GA
ScienceWorld	Qwen3-1.7B	46.18%	53.61%
ScienceWorld	Qwen3-4B	48.78%	54.68%
ALFWorld	Qwen3-1.7B	85.26%	93.47%
WebShop	Qwen3.5-0.8B	37.65%	39.90%

Success rates are the paper's four-seed means for the corresponding checkpoints; ALFWorld reports Unseen success. Individual model cards contain the full results and inference settings.

Model files

Each model is stored in the subdirectory linked above and includes BF16 weights, configuration, tokenizer and chat template. Individual model cards describe the corresponding checkpoint and inference settings.

Download and evaluate

Download the desired model subdirectory with the Hugging Face CLI. Set REPO_ID to this repository's owner/name:

REPO_ID="owner/GC-OPD"
MODEL="gc-opd-scienceworld-qwen3-1.7b"
hf download "$REPO_ID" --include "${MODEL}/*" --local-dir ./models

Pass ./models/${MODEL} to --model in the matching ScienceWorld, ALFWorld or WebShop evaluation command. Use the supplied tokenizer and chat template with the environment-specific prompts and evaluation settings in the code.

Citation

@misc{yi2026gcopd,
  title = {Graph-Conditioned On-Policy Agent Distillation from Off-the-Shelf Teachers},
  author = {Xiaohan Yi and Wen Luo and Yani Huang and Junfeng Zhan and Asher Qin and Peilin Zhao and Xi Xiao},
  year = {2026},
  eprint = {2609.37522},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.37522}
}

License

The models are released under Apache-2.0. See the license and attribution in each model directory.