Beaker automates MIT's weight-space programming: one-neuron edit lifts MATH-500 from 7% to 37% for under $4
VictorSanh · x · 2026-11-12
After MIT CSAIL's weight-space programming paper went viral, autonomous model improver Beaker turned the mechanism into an automated loop: propose a weight edit, score it, keep it only if it measurably helps, and verify on held-out problems—no gradients, no training run.
- In about an hour, Beaker found a one-neuron edit lifting Qwen3.5-0.8B from 7.4% to 37.4% on MATH-500, for under $4
- Mechanism: the small model mostly fails by looping on "Wait" until it exhausts its token limit; Beaker traced this to a single neuron that suppresses "Wait" output and doubled it, using only weights and edit scores
- Complementary to MIT's result: their larger edit helps the model escape loops rather than stop them; stacking both reaches 58.8% vs 51.1% for MIT alone
- Limits: it shortens answers by default and can't hit every goal
Weights are open-sourced on Hugging Face with public run logs.
More from Models
- Claude Fable 5.2 tops Agent Arena at +30.4%, 1.5x the gain of runner-up Opus 5.5 — daniel_mac8 · 2026-11-12
- Polymarket Puts Only ~40% Odds on Next Gemini Flash (3.9+) Shipping by Year-End — Polymarket · 2026-11-12
- Agent Arena update: Muse Ember (Max) debuts at #25, tied with Muse Wonder — arena · 2026-11-12
- Blood on the Clocktower benchmark: Kimi K3.1 Pro tops open models, just below GPT 5.5 — eliebakouch · 2026-11-12
- Replicating all 9 weight-space programming experiments: big wins, one miss — burny_tech · 2026-11-12
- Open-source Iris-Z unifies image generation and depth in one backbone, beating MoGe-3 — victormustar · 2026-11-12