Overview

WebPRMBench is the first comprehensive evaluation benchmark dedicated to Web Process Reward Models (WebPRMs). It evaluates how well a reward model can judge the quality of web agent actions during long-horizon web navigation. Each instance presents a web state (page context, trajectory history, user intent) and two candidate agent responses (thought + action). The task is to identify which response better advances the user's goal.

The benchmark spans 4 diverse web environments with 1,150 step-level preference instances, each containing one environment-verified positive action and four negative alternatives.

Data Distribution

Each instance has one positive (expert-demonstrated) action and four rejected alternatives, yielding four pairwise comparisons per state (4,600 rows total).

Data Fields

Scoring

Given a web state, two candidate responses (one expert-demonstrated, one rejected), and a model output selecting one response, the scoring works as follows:

• The model generates a structured justification concluding with <Answer>Response 1</Answer> or <Answer>Response 2</Answer>.
• A prediction is correct if the model's selected response matches the ground-truth expected label.

Evaluation Metrics

We adopt two complementary metrics:

• Pairwise Accuracy (Pair Acc): Fraction of pairs where the model correctly identifies the better response. Measured per-pair.
• Best-of-N Accuracy (BoN Acc): Fraction of states where the model ranks the expert action above all Q=4 distractors simultaneously. This is strictly harder than Pairwise Acc — a model must be correct on all 4 pairs for a given state. BoN Acc provides stronger discriminative power and better alignment with downstream agent performance.

Leaderboard

Results from the WebArbiter paper (Table 2). Models marked with ⋆ are ours. Bold = best, underline = second best.

Benchmark Construction

WebPRMBench is constructed from successful trajectories in AgentRewardBench (Lù et al., 2025):

• Positive samples: Actions from expert-demonstrated trajectories verified to succeed in the real web environment. Each trajectory is validated for monotonic progress with minimal steps.
• Negative samples: Four rejected alternatives per state, sampled from a diverse ensemble of policy models (Qwen2.5-7B/72B-Instruct, Llama-3.3-8B/70B-Instruct, GPT-4o/4o-mini, Claude-3.5-Haiku/3.7-Sonnet, Gemini-2.5-Flash/Pro). Actions are filtered via rule-based checks and manual review to ensure they are genuinely incorrect.
• Positional balancing: The positive action is not fixed to a specific side and may appear on either side of the preference pair.

Prompt Structure

Each prompt contains:

1. Intent — The user's high-level goal
2. AXTree — Accessibility tree snapshot of the current web page
3. Trajectory — Sequence of prior thought–action pairs
4. Start URL / Current URL — Contextual URL information
5. Two Assistant Responses — Each with THOUGHT and ACTION

Expected output format:

<State>Summary of the current page state.</State>
<Criteria>Task-specific evaluation criteria and weights.</Criteria>
<Analysis>Detailed comparison of Response 1 and Response 2.</Analysis>
<Answer>Response 1</Answer>

Usage

from datasets import load_dataset

dataset = load_dataset("ZYao720/WEBPRMBENCH", split="test")
print(f"Total rows: {len(dataset)}")  # 4600 (1,150 instances × 4 pairs each)

# Filter by source environment
webarena = dataset.filter(lambda x: x["source_name"] == "webarena")
mind2web_task = dataset.filter(lambda x: x["source_name"] == "mind2web_test_task")
workarena = dataset.filter(lambda x: x["source_name"] == "workarena")
assistantbench = dataset.filter(lambda x: x["source_name"] == "assistantbench")

Evaluation Script

We provide a vLLM-based evaluation script in the code repository:

export MODEL=path/to/your/model   # local path or HuggingFace repo id
bash eval/WebPRMBench/eval_one_command.sh

Results (Pairwise and BoN Accuracy per environment) are saved to results/.

Related Resources

License

WebPRMBench is released under the MIT License. As an aggregated benchmark, users should also comply with the licenses of the underlying source environments:

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}, 
}