Draft Proposes Self-Programmable LLMs: Let Models Read and Edit Their Own Weights
menhguin · x · 2026-11-03
The author explores using open-source LLMs of different sizes to edit model weights: large LLMs are smarter while small ones have simpler weights, and there may be a productive middle ground. The hypothesis is that LLMs may be especially capable of recognizing and editing their own weights.
A three-part recipe for unsupervised weight editing
- Leverage expert sparsity to improve monosemanticity;
- Use zeroth-order methods like ZOO that only need forward/backward passes;
- Apply automated interpretability to the weights.
The linked draft, Towards Self-Programmable Interpretable LLMs, argues finer expert granularity and lower active fractions may predict separability and cheaper, more precise edits, and outlines ablations over expert count, width, active fraction, routing, and shared experts — comparing sparse MoEs against dense models with SAEs and transcoders, since architecture may determine whether interpretability findings generalize.
Related event: Researchers explore LLMs editing their own weights for interpretability(4 posts)→
More from Research
- 'Text-to-neural-network' vision: describe a task and get a purpose-built tiny net — flowersslop · 2026-11-03
- CFSG fully formalized in Lean: 43.1M lines, and a flaw found in a published proof — ChrSzegedy · 2026-11-03
- Memo Akten's Ants simulation shows collective intelligence inspired by a PNAS study — memoakten · 2026-11-03
- Stanford launches Digital Brain Project to build predictive models of the human brain, trained on HAI's Marlowe GPU cluster — StanfordHAI · 2026-11-03
- Qwen base model looped on 93% of MATH-500, casting doubt on claimed 6x gain — paul_cal · 2026-11-03
- After two silent grant rejections, researcher posts federated learning paper on arXiv — wandedob · 2026-11-03