AutomataBench v1

Reversible Space-Time Reconstruction

A benchmark for reconstructing reversible cellular automaton initial states from revealed cells in their space-time evolution.

Leaderboard

Private Eval Results

Official results on held-out test set. Scores are weighted by difficulty (easy 25%, medium 35%, hard 40%). Each model gets one attempt per problem with a 600 second timeout.

#1
GPT-5.5
OpenAI
45.3%
#2
Claude Opus 4.8
Anthropic
30.2%
#3
DeepSeek V4 Pro
DeepSeek
28.3%
#4
Gemini 3.1 Pro Preview
Google
14.7%
#5
Grok 4.3
xAI
6.5%
One attempt per problem, 600 second timeout

Sample Task

Reconstruct the Initial State

Given a reversible cellular automaton rule and partial observations from its space-time evolution, reconstruct the initial state that produced those observations. Each task provides a known transition rule (2×2 Margolus block), a toroidal grid, revealed cells across time steps, and a target time horizon.

Transition Rule

2×2 block permutation showing how each input pattern transforms.

Observations Across Time

Dots indicate hidden cells. Use arrows to step through time.

010.0...1....1..
Step 1 of 5
sample_easy_000001 · easy · 4×4 grid · T=4

Dataset

Public Splits Available

Public splits include full answers for inspection and debugging. Use these for development, but note the official leaderboard scores only the held-out private_eval split.

V1 uses compact 4×4 grids simulated for 4–7 time steps. Despite their small size, these problems effectively separate frontier model capabilities under the 600s budget.

60Sample
300Public Dev
300Public Eval

Except where otherwise noted, the content on this website is dedicated to the public domain under CC0 1.0 Universal. No rights reserved.