iWSP Lets LLM Agents Write Model Weights Directly, No Backprop Needed

Microsoft Research and collaborators introduced interpretable weight-space programming (iWSP), where LLM agents directly write model weights and features as code, using prior knowledge instead of gradient training. The method was applied to build interpretable Transformers and brain-encoding models.

2026-11-21 ~ 2026-11-21 · 2 related posts