LLMs can natively edit neural network weights: 0.016% of changes boost small model math 6x
Dr_Singularity · x · 2026-11-03
A MIT CSAIL paper, Sparks of Weight-Space Programming, reports that LLMs can natively read and write the weights of other neural networks. GPT-6 Astra was asked to fix Qwen3.5-0.8B's repetitive reasoning loops by directly editing its weights — changing only 0.016% of parameters, with no gradients or training.
Results: math benchmark 7.9%→48.1%, grade-school math 13.8%→67.1%, instruction following 21.0%→62.3%. Authors include Omar Khattab.
Related event: MIT paper shows LLM editing 0.016% of weights boosts small model math 6x(27 posts)→
More from Models
- Sol 6.1 still best cost-per-intelligence worker; Fable 5.5 more well-rounded — haider1 · 2026-11-03
- Anthropic halves its Max 20x plan usage and renames it "Ultra" — EntityRemovalGuy · 2026-11-03
- Critics tear apart "Weight-Space Programming": Atari test spoon-feeds answers, Qwen case fails on Gemma — anshulkundaje · 2026-11-03
- Spider-Verse-style 2D animation built with Opus fuels rumors it routes to Fable 5.5 — minchoi · 2026-11-03
- More "Fable 5.5" demos: pelican, horse and bike scenes drawn and animated entirely in code — minchoi · 2026-11-03
- One prompt, four models: rumored "Fable 5.5" ships a full 3D open-world RPG while rivals stumble — minchoi · 2026-11-03