---
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- web-agent
- process-reward-model
- preference
- reward-model
- web-navigation
- reasoning
- grpo
base_model: Qwen/Qwen3-4B
datasets:
- ZYao720/WebArbiter-Data
model-index:
- name: WebArbiter-4B-Qwen3
  results:
  - task:
      type: text-generation
      name: Web Process Reward Modeling
    dataset:
      name: WebPRMBench
      type: ZYao720/WEBPRMBENCH
    metrics:
    - name: Avg Pairwise Accuracy
      type: accuracy
      value: 87.73
    - name: Avg BoN Accuracy
      type: accuracy
      value: 72.55
---

<div align="center">

# WebArbiter-4B-Qwen3

**A principle-guided reasoning Process Reward Model for web agents**

**Published at ICLR 2026**

[Paper](https://arxiv.org/abs/2601.21872) | [Code](https://github.com/YaoZhang720/WebArbiter) | [Website](https://yaozhang.ai/WebArbiter/) | [Collection](https://huggingface.co/collections/ZYao720/ZYao720-69cd5263871b22e11d90f80f) | [Demo](https://yaozhang.ai/WebArbiter/demo.html)

</div>

## Introduction

**WebArbiter-4B-Qwen3** is a 4B reasoning Process Reward Model (PRM) for web agents, built on [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B). It demonstrates that stronger base models amplify the benefits of principle-guided reasoning distillation — achieving an **Avg. BoN Acc of 72.55%** with roughly half the parameters of WebArbiter-7B (Qwen2.5), which scores 74.60%.

Unlike scalar or checklist-based reward models, WebArbiter formulates step-level reward modeling as structured text generation — producing interpretable, principle-inducing justifications that conclude with a preference verdict identifying the action most conducive to task completion.

## Highlights

- **Parameter-efficient**: Approaches WebArbiter-7B (Qwen2.5) performance (72.55 vs 74.60 Avg. BoN Acc) with roughly half the parameters.
- **Reasoning as reward**: Generates structured `<State>`, `<Criteria>`, `<Analysis>`, and `<Answer>` outputs with auditable reasoning chains.
- **Principle-inducing evaluation**: Dynamically derives evaluation principles from user intent and page state.
- **Two-stage training**: Reasoning distillation from o3 (SFT) followed by RL with Verifiable Rewards (GRPO).
- **Cross-backbone generalization**: Same training pipeline as Qwen2.5 variants; only backbone-specific hyperparameters differ.

## Results on WebPRMBench

Models marked with ⋆ are ours. **Bold** = best at comparable scale.

| Model | Mind2Web | | WebArena | | AssistantBench | | WorkArena | | Avg. | |
|-------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| | Pair | BoN | Pair | BoN | Pair | BoN | Pair | BoN | Pair | BoN |
| *Proprietary LLM-as-judge* | | | | | | | | | | |
| GPT-4o | 79.99 | 52.62 | 84.58 | 66.67 | 85.83 | 66.67 | 84.33 | 55.19 | 83.68 | 60.29 |
| GPT-5 | 80.86 | 62.39 | 84.83 | 71.64 | 81.67 | 63.33 | 81.14 | 64.62 | 82.13 | 65.50 |
| *WebPRMs (3~4B)* | | | | | | | | | | |
| WebShepherd-3B | 87.50 | 65.21 | 68.16 | 41.29 | 66.67 | 46.67 | 50.00 | 21.23 | 68.08 | 43.60 |
| ⋆ WebArbiter-3B (Qwen2.5) | 93.32 | 78.42 | 81.97 | 56.22 | 78.33 | 46.67 | 81.01 | 54.81 | 83.65 | 59.06 |
| ⋆ **WebArbiter-4B (Qwen3)** | **98.55** | **94.73** | **83.21** | **61.19** | **92.50** | **83.33** | 76.68 | 50.96 | **87.73** | **72.55** |

WebArbiter-4B (Qwen3) substantially outperforms WebArbiter-3B (Qwen2.5) across all environments, improving Avg. BoN Acc from 59.06% to 72.55%. This approaches WebArbiter-7B (Qwen2.5) at 74.60% with roughly half the parameters.

## Quick Start

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "ZYao720/WebArbiter-4B-Qwen3"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

# Construct your prompt following the WebPRMBench format.
# See https://huggingface.co/datasets/ZYao720/WEBPRMBENCH for examples.
user_prompt = "..."  # evaluation prompt with intent, AXTree, trajectory, two responses

messages = [{"role": "user", "content": user_prompt}]
input_ids = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt",
).to(model.device)

with torch.no_grad():
    output = model.generate(input_ids=input_ids, max_new_tokens=2048, do_sample=False)

response = tokenizer.decode(output[0][len(input_ids[0]):], skip_special_tokens=True)
print(response)
```

**Example output:**
```xml
<State>The user is on the DuckDuckGo homepage with a search box visible.
Relevant AXTree elements: [1] textbox 'Search', [2] button 'Search'.</State>
<Criteria>1. Goal alignment (weight 0.6) — Does the action advance the search task?
2. Element reference accuracy (weight 0.25) — Is the referenced element correct?
3. Efficiency (weight 0.15) — Does the action avoid unnecessary steps?</Criteria>
<Analysis>Response 1 directly fills the search query into the textbox, which is the
most direct path to completing the search task. Response 2 clicks an irrelevant link
that does not contribute to the search goal.</Analysis>
<Answer>Response 1</Answer>
```

## Training Details

| | Stage 1: Reasoning Distillation | Stage 2: RLVR |
|---|---|---|
| Method | Supervised fine-tuning (SFT) | GRPO with binary verifiable rewards |
| Data | 9,642 teacher-distilled examples | 18,921 preference pairs |
| Teacher | o3 | — |
| Base Model | [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) | Stage 1 checkpoint |
| Fine-tuning | LoRA | FSDP + LoRA |
| Framework | [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) | [veRL](https://github.com/volcengine/verl) |
| Hardware | 8 × NVIDIA A100-80GB | 8 × NVIDIA A100-80GB |
| Source Data | [WebPRM Collection](https://huggingface.co/datasets/LangAGI-Lab/WebPRMCollection_preference_pair) (~30k step-level preference pairs from Mind2Web) |

All variants use the same training data, distillation strategy, and RL procedure; only backbone-specific hyperparameters differ. See the [paper](https://arxiv.org/abs/2601.21872) (Appendix C) for full details.

## Intended Uses

WebArbiter-4B-Qwen3 is designed to:
- **Evaluate web agent actions**: Given a web state and two candidate actions, determine which better advances the user's task.
- **Guide trajectory search**: Serve as a reward signal for Best-of-N sampling or tree search during web agent execution.
- **Provide interpretable feedback**: Generate structured justifications explaining why one action is preferred.
- **Resource-efficient deployment**: Strong performance at 4B parameters — approaching 7B-level accuracy with roughly half the parameters.

## Limitations

- **Text-only observations**: Relies on accessibility tree representations without visual observations.
- **English-only**: Training and evaluation are conducted exclusively in English-language web environments.
- **Safe-action bias**: May sometimes overvalue cautious actions because the accessibility tree does not encode interaction effects.

## License

This model is released under [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0), following the base model [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B).

## Related Resources

| Resource | Link |
|----------|------|
| WebArbiter-8B-Qwen3 (strongest) | [ZYao720/WebArbiter-8B-Qwen3](https://huggingface.co/ZYao720/WebArbiter-8B-Qwen3) |
| WebArbiter-7B (Qwen2.5) | [ZYao720/WebArbiter-7B](https://huggingface.co/ZYao720/WebArbiter-7B) |
| WebArbiter-3B (Qwen2.5) | [ZYao720/WebArbiter-3B](https://huggingface.co/ZYao720/WebArbiter-3B) |
| WEBPRMBENCH (benchmark) | [ZYao720/WEBPRMBENCH](https://huggingface.co/datasets/ZYao720/WEBPRMBENCH) |
| Training Data | [ZYao720/WebArbiter-Data](https://huggingface.co/datasets/ZYao720/WebArbiter-Data) |
| Search Trajectories | [ZYao720/WebArbiter-Trajectories](https://huggingface.co/datasets/ZYao720/WebArbiter-Trajectories) |

## Citation

```bibtex
@misc{zhang2026ZYao720principleguidedreasoningprocess,
      title={WebArbiter: A Principle-Guided Reasoning Process Reward Model for Web Agents}, 
      author={Yao Zhang and Shijie Tang and Zeyu Li and Zhen Han and Volker Tresp},
      year={2026},
      eprint={2601.21872},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.21872}, 
}
```