Two LLMs sword-fight in real physics while humans blind-vote: open-source Elo arena shares data

Time-Shelter-35 · reddit · 2026-08-29

An indie developer open-sourced STICKBLADE ARENA: two models each control a stickman in a pymunk real 2D physics arena (momentum, ragdolls, weapon collisions). Each turn a model reads a JSON world state (HP, positions, weapon reach, cooldowns) and returns a JSON action — no canned prompts, pure reasoning about spacing and timing.

Eval design:

Humans watch replays blind (fighters labeled A/B), vote who fought smarter, then Elo updates
Elo keyed on 6 axes: (model, sharpzone, weapon, mode, arena, blindfolded)
Two leaderboards: Perceived (human votes → Elo) and Objective (win/loss/damage/hit-rate)
4 non-LLM baseline bots (random / greedy / distance-holder / scripted pro) test whether models truly beat "always swing"

Roster & early findings:

24 fighters: 10 OpenRouter free models (gpt-oss, gemma-4, kimi-k2…), 6 on Groq, paid gpt-4o-mini
Bow matches are won by whoever waits for cooldown; most models dump the whole magazine on turn 1
Blindfolded mode collapses the top of the leaderboard
deepseek-r1-distill-70b times out on 15% of turns, losing to the clock
Scripted bot:pro currently beats 3 free-tier LLMs objectively

All matches export via /api/export as JSON/JSONL; source on GitHub; live site needs no login, fights take 1–3 min. Motivation: tired of leaderboards where another LLM grades "reasoning".

Original post →

Two LLMs sword-fight in real physics while humans blind-vote: open-source Elo arena shares data

More from coding & agent