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.

Original post →

More from coding & agent

coding & agent channel →