WebArbiter-3B

A principle-guided reasoning Process Reward Model for web agents

Published at ICLR 2026

Paper | Code | Website | Collection | Demo

Introduction

WebArbiter-3B is a 3B reasoning Process Reward Model (PRM) for web agents, built on Qwen2.5-3B-Instruct. 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.

Despite its compact size, WebArbiter-3B achieves an Avg. BoN Acc of 59.06% on WEBPRMBENCH, outperforming the previous SOTA WebPRM (WebShepherd-3B) by 15.5 points and surpassing all open-source LLM-as-judge baselines up to 70B parameters. For the strongest variant, see WebArbiter-7B.

Highlights

• Reasoning as reward: Generates structured <State>, <Criteria>, <Analysis>, and <Answer> outputs with auditable reasoning chains, instead of scalar scores or brittle checklists.
• Principle-inducing evaluation: Dynamically derives evaluation principles from user intent and page state, enabling robust assessment that generalizes across environments.
• Two-stage training: Reasoning distillation from o3 (SFT) followed by RL with Verifiable Rewards (GRPO) to correct teacher biases and align verdicts with ground-truth correctness.
• Efficient and deployable: Strong performance at 3B parameters, suitable for resource-constrained deployment scenarios.

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-mini	81.74	50.92	78.23	56.72	89.17	73.33	81.43	46.70	82.64	56.92
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
Open-source LLM-as-judge
Qwen2.5-3B-Instruct	76.46	36.93	60.32	15.42	75.83	33.33	64.45	19.34	69.27	26.76
Qwen2.5-7B-Instruct	77.79	39.18	74.88	42.79	84.17	53.33	77.58	35.85	77.61	42.78
Llama-3-70B-Instruct	80.55	49.36	77.36	50.75	85.83	70.00	79.08	40.09	80.71	52.55
WebPRMs (3B)
WebShepherd-3B	87.50	65.21	68.16	41.29	66.67	46.67	50.00	21.23	68.08	43.60
⋆ WebArbiter-3B	93.32	78.42	81.97	56.22	78.33	46.67	81.01	54.81	83.65	59.06
WebPRMs (7B+)
WebShepherd-8B	86.66	73.69	68.33	43.88	55.92	30.00	54.56	25.53	64.34	43.28
⋆ WebArbiter-7B	97.07	89.53	88.43	68.66	89.17	70.00	82.09	70.19	89.19	74.60

WebArbiter-3B outperforms WebShepherd-8B (a much larger 8B model) on Avg. BoN Acc (59.06 vs 43.28), demonstrating the efficiency of the principle-guided reasoning approach.

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "ZYao720/WebArbiter-3B"

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:

<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	Qwen2.5-3B-Instruct	Stage 1 checkpoint
Fine-tuning	LoRA (rank 128, lr 8e-4)	FSDP + LoRA (lr 9e-6)
Framework	LLaMA-Factory	veRL
Hardware	8 × NVIDIA A100-80GB	8 × NVIDIA A100-80GB
Source Data	WebPRM Collection (~30k step-level preference pairs from Mind2Web)

Intended Uses

WebArbiter-3B 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, useful for debugging and analysis.
• Resource-efficient deployment: Suitable for scenarios where 7B+ models are too large, while still significantly outperforming larger checklist-based WebPRMs.

Limitations

• Text-only observations: WebArbiter relies on accessibility tree representations without visual observations. In environments where layout, spatial arrangement, or visual cues carry task-relevant information, this text-only formulation may miss critical signals.
• English-only: Training and evaluation are conducted exclusively in English-language web environments.
• Safe-action bias: The model may sometimes overvalue cautious actions (e.g., hover over click) because the accessibility tree does not encode interaction effects.
• Element reference hallucination: When a candidate action's reasoning is strongly task-aligned, the model may trust the semantic signal over low-level bid verification, potentially missing incorrect element references.

License

This model is released under Apache 2.0, following the base model Qwen2.5-3B-Instruct.

Related Resources

Resource	Link
WebArbiter-8B-Qwen3 (strongest)	ZYao720/WebArbiter-8B-Qwen3
WebArbiter-7B	ZYao720/WebArbiter-7B
WebArbiter-4B-Qwen3	ZYao720/WebArbiter-4B-Qwen3
WEBPRMBENCH (benchmark)	ZYao720/WEBPRMBENCH
Training Data	ZYao720/WebArbiter-Data
Search Trajectories	ZYao720/WebArbiter-Trajectories

Citation

@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}, 
}