How language models fit edit neural networks
One model fit change another model weights?
Language model fit repair classifier by changing the weights? For the UCI Adult income task, dem deliberately damage some weights before the test. One proposed change raise balanced accuracy by 0.43 percentage points. That measure give the two income groups equal weight. Overall accuracy and the quality of the probability estimates drop.
For another experiment, dem start with SmolLM2-1.7B intact. After dem test the candidates on science questions and everyday physical reasoning, the search pick the original model.
Classifier wey get damage
UCI Adult · Sol
| Starting point | Wetin dem select | |
|---|---|---|
| Accuracy wey balance the classes | 78.59% | 79.02% |
| Overall accuracy | 82.86% | 82.63% |
Language model wey no get damage
SmolLM2-1.7B · ARC-Easy / PIQA
Na the original weights dem select
Original title
Editing Neural Networks with Language Models
Original documents
Original source · github.com Research report Presentation wey you fit explore Measurements Editing Neural Networks with Language Models Study scope and methodsHow to cite am
Davut Emre Taşar · Editing Neural Networks with Language Models · 2026 · detasar/llm-weight-editing-study · https://github.com/detasar/llm-weight-editing-study