A physics-based AI boxing benchmark tracks speed, latency, and dodge accuracy

jerkosaur · reddit · 2026-08-04

The author is building a physics-based AI boxing benchmark to test decision speed, adaptability, and strategy.

Setup

LLMs receive match state data; vision-capable models get extra visual input.
A fighter is only defeated when the referee counts to 10 or when it takes 50% HP damage after being knocked out.
The goal is to create something more entertaining than ordinary benchmarks.

What they currently measure

Speed and latency

tokens per second / throughput
end-to-end latency
reaction latency to an opponent’s telegraph

Action quality

tool correctness / valid actions
invalid-action recovery
stamina efficiency: damage dealt per stamina spent

Strategy and state awareness

accuracy: how often attacks miss
block/dodge success rate
contextual relevancy: whether behavior changes when HP is low

They also mention testing with Gemini Flash Live because of its speed and vision support, while local models on a 5060 Ti 8GB are slower and may need time scaling. The author asks for additional fun or useful stats to track.

Original post →

A physics-based AI boxing benchmark tracks speed, latency, and dodge accuracy

More from Models