A physics arena tests LLMs as fighting agents
Several LLMs can now compete in a weapon duel with real physics instead of being compared only by text answers. On each turn, each model receives the fight state as JSON, including health, distance, the opponent’s last move, and what hit on the previous turn. The model then chooses an attack and footwork.
A physics engine calculates momentum, joint limits, collisions, damage, and which part of the weapon made contact. The user can decide which weapon zones are dangerous, such as only the sword tip or only the handle, and those rules are placed in the . People vote without seeing s, then the names and Elo scores are shown afterward.
Free models such as Llama 3.3 70B, GPT-OSS, Qwen3, Nemotron, and Gemma can be used at no cost, and any can also be pasted in. A joint-control mode lets the model directly move 10 body joints, but current models are still poor at controlling bodies.
Key points
- LLMs receive fight state as JSON and choose an attack plus movement each turn.
- A physics engine calculates collisions, speed, body limits, weapon zones, and damage.
- The user sets which parts of a weapon are dangerous, and the rule goes into the .
- Voting hides s until the result, then Elo scores are revealed.
- Free models can be used, so experiments can run with no model cost.