iWSP: LLM agents write transformer weights by hand, no backpropagation needed
burny_tech · x · 2026-11-21
A new paper introduces interpretable weight-space programming (iWSP), where an LLM agent directly programs model weights in code without gradient training. Agents craft compact transformers hitting 100% hidden accuracy on algorithmic tasks, build brain encoding models that outpredict interpretable baselines in visual cortex, surface confounds in language fMRI, and discover interpretable acoustic features in auditory ECoG. Interpretability is audited by having a fresh agent rebuild the model from a ≤200-word description with near-identical performance.
More from coding & agent
- iWSP: LLM agents hand-code transformer weights without backprop, beat deep brain-encoding models — burny_tech · 2026-11-21
- Google launches Gemini Notebook API public preview with file uploads and MCP support — tokumin · 2026-11-21
- a16z's Martin Casado shares full AI smart home BOM, gives Grok bot a $20k budget — Scobleizer · 2026-11-21
- codehygiene: a 428K-param ternary model at 11KB that scores code cleanliness with zero training — CatAstro_Piyush · 2026-11-21
- How many tools is too many for a scraping MCP server? — Big_Barnacle_2452 · 2026-11-21
- ARTEX, an AI autonomous pentesting system, hits 13.8k stars on GitHub — AccBalanced · 2026-11-21