iWSP: LLM agents hand-code transformer weights without backprop, beat deep brain-encoding models
burny_tech · x · 2026-11-21
A Microsoft Research-led team introduces Interpretable Weight-space Programming (iWSP), where an LLM agent iteratively builds interpretable models by explicitly programming weights and features in code — no backpropagation. The agent hypothesizes and refines human-readable computations across trials, mirroring a scientist's workflow, with LLM audits confirming interpretability. Validated on arithmetic and sentiment benchmarks, then applied to neural encoding across vision, language, and audition, where it outpredicts deep models in visual cortex while remaining interpretable. Authors include Jianfeng Gao and Nebojsa Jojic.
Related event: iWSP Lets LLM Agents Write Model Weights Directly, No Backprop Needed(2 posts)→
More from coding & agent
- Four DGX Sparks, one week of tuning: 26.85 to 56.71 tok/s — AIFlow_ML · 2026-11-21
- Lessons from shipping a no-auth read-only travel MCP server through Claude's directory review — Wooden_Body_7891 · 2026-11-21
- call.md: open-source desktop app turns meetings into live agent loops, 2.1k GitHub stars — tom_doerr · 2026-11-21
- Dev swaps GLM 5.3 Flash for Thinking Machines' Inkling 2 Small in Windows observability app — michellechen · 2026-11-21
- Dev fully automates Honkai Star Rail with Claude Code + GPT Sol on a Pixel 10 in ~5 days, ~$300 — Ok_Chipmunk_8734 · 2026-11-21
- I replaced automation chains with 31 tiny single-job agent apps — NexelythCollective · 2026-11-21