Definition
Plain language
A set of 2D physics puzzles built to test whether an AI agent can experiment its way to a solution.
As stated in the literature
A benchmark built on PHYRE for evaluating experiment-driven physics-reasoning agents, shipping level-specific analysis tools that provide cheap intervention primitives.
Why it matters: It matters because it measures whether an agent can figure things out by experimenting, not just by reasoning in one shot.
For example, a level might ask the agent to get a green ball into a cup, and it can nudge objects to test what happens before committing.
Heard on the show
“The puzzle is from a benchmark called Interphyre — built on top of PHYRE, which is a set of 2D physics puzzles where you place one object, hit go, and let gravity do the rest.”Episode 186 — How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining