Zero-Parameter Model Reveals Neural Network Structure
A_Freaky-Frog · reddit · 2026-07-15
This study claims to have built a new instrument to "read the internal structure of models," yielding results on day one under open, pre-registered, and reproducible conditions.
Core Method
The authors transform model weights into a "spectral basis" and compare them with randomly shuffled weights. Any preserved signal is considered a structural imprint written by the training process, rather than pure statistical noise.
Key Findings
- Embedding layers are crucial: All 11/11 tested models showed significant signals on token embeddings, covering models from 4B to 1T parameters with various training recipes.
- Minimal structural deletion severely damages the model: Removing the strongest 1.5% of spectral coefficients in GPT-2 cripples the model; deleting the same amount randomly has little impact, showing a difference of about 150x.
- Structure is written progressively during training: Using public checkpoints, the authors observed the signal going from almost nothing to the embedding "waking up" (around step 256), peaking at roughly step 4000, and then plateauing.
- Gradients carry structure very early: In controlled experiments, structural information can be seen in the gradients as early as step 4, indicating that the optimizer determines whether these structures are deposited.
- Training data is readable: The authors claim their probe can rank a model's actual training corpus first and identify verbatim memorization of public text; for example, the Gettysburg Address can be reproduced up to 9 words.
- "Thought text" differs from answers: The model's reasoning fragments show measurable differences from the final answer in this count signature, and attention patterns are closer to theoretically predicted cascade ratios.
Open Materials
The authors also provided links to the toolkit, guide, theoretical repository, and an updated v4.3 paper, noting they are continuing to work on the scaling ladder, fitting deposition curves, and probing the last blind spots of attention.
More from Research
- Structural ensembles beat single predictions in TCR:pMHC generalization study — quaidmorris · 2026-07-22
- RSS launches under OMSF to push structural biology data modeling at scale — MoAlQuraishi · 2026-07-22
- enFoldX tops 8 neoantigen scans and an unseen-peptide benchmark — quaidmorris · 2026-07-22
- enFoldX reaches AUC 0.82 on human VDJdb and transfers to mouse at 0.76 — quaidmorris · 2026-07-22
- enFoldX gains accuracy as AF3 ensemble disagreement rises for non-binders — quaidmorris · 2026-07-22
- A 3D ray plot shows how hard this Jacobian counterexample is to read — moultano · 2026-07-22