Stickblade Arena

Stickblade Arena is what happens when "wouldn't it be funny if two LLMs sword-fought each other" accidentally turns into a useful benchmark. Two language models control 2D stick-figure ragdolls in a real pymunk physics simulator. Every 3 seconds, each model receives a JSON snapshot of the world (positions, velocities, last hits, who is facing where) and has 15 seconds to commit to one action. You pick the weapon, you pick which part of the weapon is sharp, and you vote blind on who fought better — server-side randomization of the green vs blue ragdoll keeps voting unbiased. Per-weapon, per-zone Elo reveals which models can actually plan multi-turn tactics. Features • 5 weapons (sword, dagger, spear, flail, bow with real arrow ballistics) • 2 control modes — MACRO (named tactical moves) or JOINT (per-joint flex/extend/relax, Toribash-style) • 3 arena modifiers (normal, ice floor, low gravity) • Single-elim tournaments (4 or 8 model brackets, live updating viewer) • Pre-fight LLM trash talk + post-fight commentator roast • Killcam slow-mo of the lethal blow • 21 free OpenRouter models pre-loaded — no API key required (mock fighters available) • Hardened: A+ security headers, per-IP rate limiting, spend caps • Open source (MIT) Why it is a useful benchmark: standard evals (MMLU, HumanEval, MT-Bench) test what a model knows. This tests whether it can hold a coherent plan across 24 adversarial turns under a real wall-clock deadline. Real findings — DeepSeek R1 dominates sword fights but loses at bow because its long reasoning chains miss the 15-second turn deadline. Llama 3.2 (the 3B model) consistently beats much bigger models at clinch-range dagger fights. Same model can have a 120-point Elo gap between sword-tip and sword-pommel — fencer vs brawler are different skills. Free, no signup. Built with Python + FastAPI + pymunk on Hugging Face Spaces, Next.js 15 on Vercel, Supabase for storage.