BrowseComp

Browser operation competition benchmark

BrowseComp is a benchmark for browser operation competition tasks, evaluating agents’ comprehensive browser operation capabilities.

Overview

Attribute	Value
Task Type	Browser operations
Evaluation	Grader-based scoring
Difficulty	Medium-High

Features

Competition-grade Tasks

Tasks from browser operation competitions with high difficulty

Comprehensive Skills

Tests a wide range of browser operation capabilities

Quick Start

Run Tasks

# Run first 3 tasks
bubench run \
  --agent browser-use \
  --benchmark BrowseComp \
  --mode first_n \
  --count 3

# Run with Agent-TARS
bubench run \
  --agent Agent-TARS \
  --benchmark BrowseComp \
  --mode first_n \
  --count 3

Evaluate Results

bubench eval --agent browser-use --benchmark BrowseComp --model-id bu-2-0

Data Loading

BrowseComp supports local JSONL files or HuggingFace downloads. To use HuggingFace:

bubench run --agent browser-use --benchmark BrowseComp \
  --data-source huggingface

The HuggingFace parquet file is converted to JSONL in the HF cache before use.

Evaluation Metrics

Metric	Description
Task Completion	Percentage of tasks completed
Accuracy	Result accuracy

Data Format

Task data is stored in benchmarks/BrowseComp/data/:

{
  "task_id": "browsecomp_001",
  "task": "Navigate to the website and complete the registration form",
  "expected_result": "Registration successful"
}

Related Links

• BrowseComp Official