Frontier models can read neural network weights and identify a Pac-Man net

JnBrymn · x · 2026-11-04

An Arcturus Labs report shows frontier models can read neural network weights as plain tokenized text, reason about the behavior they encode, and even update them directly—without running a training loop.

Given a 17,253-parameter net as raw text, Astra, Opus 5.5, and Muse Spark 1.3 all figured out it plays Pac-Man; Astra did it with minimal reasoning.
The key puzzle: almost no training data explains what specific weights do, so either labs trained this in deliberately, or models generalize far further than expected. The author leans toward the latter.
Analogy: humans and fruit flies share neuronal components, but even with a complete fly connectome our brains can't intuitively read it—being a neural net doesn't grant understanding of other nets' weights, and models must overcome tokenization hurdles since weights arrive as plain text.

The finding suggests models understand structured numeric data far better than commonly assumed, with implications for interpretability research.

Original post →

Frontier models can read neural network weights and identify a Pac-Man net

More from Models