VPR-Qwen3-4B-Base-Math-Mixed

This checkpoint is trained from Qwen3-4B-Base with mixed math and VPR game experience across Sokoban, Sudoku, and Minesweeper. VPR supplies action-level process rewards through task-grounded oracles and state-group rollout, while math training follows the mixed-training protocol described in the paper.

Reported results

Values are reported under the VPR paper's evaluation protocol. SR is success rate. The general-reasoning value is the macro average across the reported OOD benchmarks. These OOD results are not a claim of exactly matched total rollout compute across training methods.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

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

inputs = tokenizer("Solve the task step by step.", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Use the VPR codebase for the exact prompts, environments, and evaluation entry points.

Resources

• Paper
• Code
• Project page

Limitations

The checkpoint is shaped by the documented math distribution, task-grounded game oracles, prompts, and action formats. Performance and safety outside those settings have not been established.

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}
}