VPR-Qwen3-4B-Sokoban

This is a Qwen3-4B checkpoint trained with Verifiable Process Rewards (VPR) on Markovian Sokoban interactions. At each visited state, VPR samples four action responses, scores their parsed actions with a task-grounded search-based Sokoban oracle, commits one highest-reward candidate, and optimizes all eligible candidates using locally normalized advantages.

Reported result

Values are percentages reported under the evaluation protocol in the VPR paper. SR is success rate; CR is completion rate.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "nics-efc/VPR-Qwen3-4B-Sokoban"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

inputs = tokenizer("<current Markovian game prompt>", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Use the environment prompts, parsers, and action conventions in the VPR codebase for reproduction. These task-specific checkpoints are not intended as general-purpose assistants.

Resources

• Paper
• Code
• Project page

Limitations

Training relies on task-grounded oracle signals and specific Markovian prompts. Performance outside the documented environments and action formats has not been established. Evaluate safety and correctness before open-ended deployment.

Citation

@misc{yuan2026verifiable,
  title         = {Verifiable Process Rewards for Agentic Reasoning},
  author        = {Huining Yuan and Zelai Xu and Huaijie Wang and Xiangmin Yi and Jiaxuan Gao and Xiao-Ping Zhang and Yu Wang and Chao Yu and Yi Wu},
  year          = {2026},
  eprint        = {2605.10325},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2605.10325}
}