Paper: Claude edits model weights directly, lifting Qwen3.5-0.8B MATH-500 by 51.4 points
udmrzn · x · 2026-11-17
The paper "Same Fix, Different Weights: Replicating Weight-Space Programming with Claude" (Subham Kundu) shows gradient-free "weight-space programming": Claude Opus 5.5 can read and edit small models' weights as text, modifying output scaling and token logits to raise Qwen3.5-0.8B's MATH-500 score by up to 51.4 points.
Key points:
- Turns weight optimization into an in-context reasoning problem instead of an expensive fine-tuning loop
- Suggests LLMs can perform direct, surgical intervention on smaller specialized models
- Fixes replicate across different weights, hence the title
More from Models
- Mistral makes Vibe free in Europe for 3 months, claims coding edge over GLM and DeepSeek — qtnx_ · 2026-11-17
- Study: 9 LLMs can write neural game agents, but they act like hand-coded programs — burkov · 2026-11-17
- Cheap tokens, expensive agent: Flash hits $11.97/task vs Luna's $0.05 in Databricks AI Index — krishnan · 2026-11-17
- Rumor: Gemini 4 to use element names, four tiers mapped to Claude lineup — koltregaskes · 2026-11-17
- Dev: LLMs seem stagnant or worse on semi-specified math problems, citing disappointing Astra — LucaAmb · 2026-11-17
- Musk: Grok 4.8 rolls out this week after full C/C++ stack rewrite — iamfakhrealam · 2026-11-17