dspy-wsp open-sources weight-space programming: frontier LM edits model weights, no training, +40 pts on MATH-500
dbreunig · x · 2026-11-14
cmpnd.ai released dspy-wsp, wrapping MIT CSAIL's "Sparks of Weight-Space Programming" method as a DSPy optimizer: a frontier writer LM reads a student model's weights in a sandboxed Python kernel, runs budgeted forward passes on practice examples, and applies targeted weight edits — no gradients, no training.
- In the paper's Experiment 8, editing just 0.016% of Qwen3.5-0.8B's weights removed a looping-inducing direction and lifted MATH-500 from 7.9% to 48.1%
- Runs on MLX, PyTorch and vLLM backends; shines with decision models like Liquid AI's d1-2b and d1-omni
- Explicitly a prototype; the team solicits issues and PRs
More from coding & agent
- Qwen 3.5 4B scores 10/10 on hidden coding tests fully local on a 16GB MacBook Air — Ok-Brush8377 · 2026-11-14
- HQ rebuilds local bot runtime: self-host agents, chat from desktop, Slack or Telegram — jacob_posel · 2026-11-14
- HQ adds MCP signup: create an account inside ChatGPT, Claude or Grok — jacob_posel · 2026-11-14
- Coolify TLS cert rotation works, but CA API buckles under eMudhra fallout rush — tristanbob · 2026-11-14
- Delta editor adds Claude Agent via ACP, works with existing Claude subscriptions — dean_rie · 2026-11-14
- Claude Code 2.1.324 is about to be released — ClaudeCodeLog · 2026-11-14