nics-efc/VPR-Qwen3-4B-Base-Math-Mixed

1---2license: apache-2.03library_name: transformers4pipeline_tag: text-generation5base_model: Qwen/Qwen3-4B-Base6tags:7- qwen38- reinforcement-learning9- agentic-reasoning10- verifiable-rewards11- vpr12- math13---14 15# VPR-Qwen3-4B-Base-Math-Mixed16 17This 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.18 19## Reported results20 21| Evaluation | Metric | Result |22|---|---|---:|23| General-reasoning OOD suite | Macro average | 52.75 |24| ALFWorld | SR | 16.12 ± 2.87 |25| WebShop | Score | 42.01 ± 2.14 |26| WebShop | SR | 1.33 ± 0.76 |27 28Values 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.29 30## Usage31 32```python33from transformers import AutoModelForCausalLM, AutoTokenizer34 35model_id = "nics-efc/VPR-Qwen3-4B-Base-Math-Mixed"36tokenizer = AutoTokenizer.from_pretrained(model_id)37model = AutoModelForCausalLM.from_pretrained(38    model_id,39    torch_dtype="auto",40    device_map="auto",41)42 43inputs = tokenizer("Solve the task step by step.", return_tensors="pt").to(model.device)44outputs = model.generate(**inputs, max_new_tokens=1024)45print(tokenizer.decode(outputs[0], skip_special_tokens=True))46```47 48Use the [VPR codebase](https://github.com/thu-nics/VPR) for the exact prompts, environments, and evaluation entry points.49 50## Resources51 52- [Paper](https://arxiv.org/abs/2605.10325)53- [Code](https://github.com/thu-nics/VPR)54- [Project page](https://thu-nics.github.io/VPR/)55 56## Limitations57 58The 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.59 60## Citation61 62```bibtex63@misc{yuan2026verifiable,64  title         = {Verifiable Process Rewards for Agentic Reasoning},65  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},66  year          = {2026},67  eprint        = {2605.10325},68  archivePrefix = {arXiv},69  primaryClass  = {cs.AI},70  url           = {https://arxiv.org/abs/2605.10325}71}72```73