WeightCode installs new skills into trained neural networks without gradient training
A framework called Weight-Space Programming uses a frozen language model to write programs that modify network weights directly, adding capabilities while leaving existing ones intact, its authors report.
In plain words
Neural networks usually learn new skills through training, which adjusts their internal numbers (weights) using gradients. WeightCode instead has a language model write small programs that edit those weights directly, installing a new capability. Once installed, the new skill runs without needing the language model again.
What was announced
In a paper on arXiv, the researchers describe Weight-Space Programming, in which executable programs read, construct or modify a network's weights. WeightCode uses a frozen language model to write these weight programs from task knowledge and the network's existing weights, with no gradient-based training. The authors report that, given a grammar or eight demonstrations, weight programs lifted exact-match performance on unseen command compositions from 0% to 99.8% and a new pickup action from 0% to 98.4% success, while preserving the original command and navigation abilities.
On ImageNet, four weight programs produced net gains of 8.84 to 17.62 points while preserving clean accuracy, which the authors say outperformed the best of eight fine-tuning methods under matched examples, checks and retention constraints. In Vision Transformers, causal interventions linked installed noise-correction gates to robustness gains. When ten weight programs were rerun on five independently trained Vision Transformers without further language model calls, they retained 83-110% of their source-network gains on average, while copying fixed weight edits gave negligible gains. Six programs written for an already edited network added 5.23-14.50 points of contrast robustness while keeping the earlier noise gain.
What it does not show
- The evidence is the paper's abstract only; full methods, datasets and peer review status are not covered, and the reported results have not been independently confirmed.
For security and IT teams
- Scope
- The source is the paper's abstract, so claims reflect the authors' own reported results rather than an independent evaluation.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.