EvoAgentBench Website

Leaderboard and project website for EvoAgentBench — a benchmark for evaluating AI agent self-evolution across diverse task domains.

• Live site: https://evermind-ai.github.io/EvoAgentBench/
• Paper: EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer
• Dataset: EverMind-AI/EvoAgentBench on Hugging Face
• Benchmark code: benchmark/

Benchmark Code

The public benchmark implementation lives in benchmark/. After cloning this repository, enter that directory before following its setup instructions:

git clone https://github.com/EverMind-AI/EvoAgentBench.git
cd EvoAgentBench/benchmark

The website and benchmark are versioned together but remain separate projects: the website root uses the MIT License, while benchmark/ uses Apache License 2.0.

Features

• Interactive leaderboard with filtering and sorting
• Filter by agent (OpenClaw, Nanobot), model, domain, and evolution condition
• Sort by method score, transfer gain, or Vanilla score
• Derived Overall across both scaffolds and four equally weighted domains
• Separate turn-cost table reproducing the paper's All/Solved/Unsolved analysis
• Paper conditions: Memento, ReasoningBank, GEPA, and Anchor Skill†
• Domain overview table with Ability/train/test statistics

Tech Stack

• Next.js 16 (App Router, static export)
• TypeScript
• Tailwind CSS v4
• shadcn/ui components

Getting Started

npm install
npm run dev

Open http://localhost:3000 to view the site.

Updating Leaderboard Data

src/data/leaderboard.csv and src/data/leaderboard-cost.csv are the sources of truth for the paper's per-cell scores and turn-cost results. Overall leaderboard rows are derived from the per-cell scores by equally weighting both scaffolds and all four domains. The CSV files are converted to src/data/leaderboard-data.ts by scripts/csv-to-data.js, which runs automatically before dev and build (or manually via npm run sync-data). Edit the CSV files, not the generated result arrays in the TS file.

Project Structure

src/
├── app/
│   ├── layout.tsx               # Root layout with Navbar + Footer
│   ├── page.tsx                 # Home: hero, features, results, domains, methods
│   └── leaderboard/
│       └── page.tsx             # Full leaderboard page
├── components/
│   ├── Leaderboard.tsx          # Main leaderboard with filters, sorting, overall rows
│   ├── BenchmarkDomains.tsx     # Domain overview table
│   ├── DomainDivergingBars.tsx  # Per-domain improvement chart
│   ├── ResearchNarrative.tsx    # Research findings section
│   ├── SkillMethods.tsx         # Self-evolution method cards with GitHub links
│   ├── FilterButton.tsx         # Reusable filter toggle button
│   ├── Navbar.tsx               # Top navigation bar
│   ├── Footer.tsx               # Page footer
│   └── ui/                      # shadcn/ui primitives
└── data/
    ├── leaderboard.csv          # Benchmark results (source of truth)
    ├── leaderboard-cost.csv     # Turn-cost results (source of truth)
    └── leaderboard-data.ts      # Domain info, method info, generated results
scripts/
└── csv-to-data.js               # Syncs result CSVs into leaderboard-data.ts
benchmark/                       # Public benchmark runner, adapters, and docs

Deployment

GitHub Pages

Push to main branch — GitHub Actions will automatically build and deploy.

Vercel / Netlify

Import the GitHub repository. No extra configuration needed.

Citation

@misc{gao2026evoagentbench,
  title={EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer},
  author={Xingze Gao and Chuanrui Hu and Hongda Chen and Pengfei Yao and Zhao Wang and Yi Bai and Zhengwei Wu and Yunyun Han and Xiaofeng Cong and Jie Gui and Yafeng Deng and Teng Li},
  year={2026},
  eprint={2607.05202},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2607.05202}
}

License

The website is released under the MIT License. The implementation under benchmark/ is released under the Apache License 2.0.