DeepSeek V4.1 Flash reads raw weights: 3 of 4 RL checkpoints identified correctly

poppear · reddit · 2026-11-04

Inspired by "Weight-Space Programming", the author tested whether DeepSeek V4.1 Flash 1017 can understand a small neural network by reading its weights alone.
Setup: four anonymized RL checkpoints, 51k parameters each with identical architectures, all weights rendered as text — no metadata, source code, or execution. Only the architecture outline and four candidate tasks were given.
Results: CartPole ✅, Flappy ✅, Catch ✅, BallBeam ❌. The model picked out patterns like relative bird/gap position and angle-based balancing; it confused BallBeam with CartPole — an interesting mistake since both share the same observation layout and similar balancing controls.
Caveats: one run per checkpoint, 1.58M tokens at maximum reasoning effort; the model even assumed ReLU where the net uses tanh. Code, checkpoints, and logs are open-sourced at github.com/galatolofederico/wsp.

More from Models