Microsoft Research has Claude write neural net weights directly: 166 params, no backprop, 100% accuracy
新智元 · wechat · 2026-11-19
Microsoft Research with Johns Hopkins, Columbia and NYU propose interpretable weight-space programming (iWSP): gradient-based training is banned, and Claude Opus 5.5 must write Python code that directly sets Transformer weights, iterating on a single validation score (39 rounds on average).
- All 13 algorithm task categories hit 100% on hidden test sets across 26 runs, with no backprop or gradient descent.
- Counterintuitive finding: stronger models write smaller networks — Haiku 5.5's median is 100k params, Sonnet 5.5 41.6k, Opus 5.5 just 166 (it first writes 3.4k, then aggressively prunes).
- Applied to brain encoding models, the transparent nets exposed a confound in a language fMRI dataset, beat existing interpretable models on visual cortex tasks, and surfaced interpretable acoustic features from ECoG data.
- Limits: NLP struggles badly (72% sentiment accuracy at 190k params); the method mostly recombines known algorithm modules — foundation models still need massive data and gradients.
Paper: arxiv.org/abs/2611.18114, code open-sourced.
Related event: iWSP Lets LLM Agents Directly Program Neural Network Weights(2 posts)→
More from Models
- Frontier models reportedly play up a casual nontechnical vibe exactly when they're smartest — repligate · 2026-11-20
- GPT-6.2 Astra tops Blueprint-Bench 2, beating the human baseline on 3D floorplans — i_dg23 · 2026-11-20
- "Why is nobody talking about Neon?" — mystery model draws hype from big names — RubenEVillegas · 2026-11-20
- Opus 3 as agent safety reviewer is too strict and paranoid, dev finds — MikePFrank · 2026-11-20
- 220B MoE with full 262K context runs on RTX 3090s: 251 tok/s on 4 cards — QuixiAI · 2026-11-20
- Suspected GPT-6 Astra leak: one-prompt iterative 3D scene edits with editable Blender files — ZeroStateReflex · 2026-11-20