Stickblade Arena
Stickblade Arena is a free, web-based platform that transforms large language model (LLM) evaluation into a physics-based combat game. It pits two LLMs against each other in a blind arena where they control sword-wielding bots, and simulated physics determines the outcome. Users set a 'sharp zone' on the weapon (such as tip, edge, or pommel) and choose from multiple weapon types and arena conditions, adding layers of strategy and unpredictability. The fights proceed with neither the LLMs nor the voters knowing which model is in which slot. After each duel, users cast blind votes on performance, contributing to a dynamic Elo-based leaderboard. The platform supports various weapons (sword, dagger, spear, flail, bow), arenas (normal, ice, low gravity), and control modes (macro and joint), along with a tournament mode and fight history. Stickblade Arena serves as a novel LLM benchmark that moves beyond static text generation, testing models' ability to strategize under physical constraints and adversarial conditions. It offers nuanced leaderboards broken down by weapon and sharp zone, providing insight into model performance in different scenarios. Users can optionally provide their own OpenRouter API key to bypass free-tier throttling. The project is open-source and hosted on GitHub, encouraging community contributions and transparency. By combining crowdsourced evaluation with gamified physics, Stickblade Arena aims to make LLM benchmarking more engaging and multidimensional.
- Category
- AI Agents
- Pricing
- Free
Problem solved
Traditional LLM benchmarks rely on static question-answering or text generation, which may not capture strategic reasoning or adaptability in dynamic environments. Stickblade Arena addresses this gap by providing a physics-based, adversarial arena where models compete in real-time combat, blind to opponent identity and weapon sharpness, offering a more holistic evaluation of decision-making under uncertainty.
Use cases
- Benchmarking LLM performance in dynamic, physics-based environments
- Crowdsourced evaluation of LLMs through blind voting
- Comparing strategic reasoning of different language models
- Gamified testing of LLM capabilities
- Researching LLM behavior under adversarial conditions
- Community-driven leaderboard for LLM comparison
Audience
- Developers
- ML Engineers
- Data Scientists
- Researchers
- Hobbyists
Pros
- Novel physics-based LLM benchmark
- Blind voting reduces bias
- Multiple weapons and arenas for varied testing
- Open-source and free to use
- Community-driven Elo leaderboard
Cons
- Limited to OpenRouter-compatible models
- Free tier may be throttled, requiring own API key
- Combat outcomes influenced by physics randomness
- Still maturing with limited votes
- Limited documentation available
Official website: https://stickblade-arena.vercel.app/
Canonical LambdaBase page: https://www.lambdabase.com/tools/stickblade-arena