Autonomous agent edits just 1,024 weights to lift Qwen3.5 0.8B from 7.4% to 37.4% on MATH-500
VictorSanh · x · 2026-11-12
After a human fixed a looping model by editing weights, the withbeaker team gave the same task to Beaker, their autonomous agent-improver. Without ever running the model or using gradients, Beaker analyzed failure modes, formed hypotheses, and hill-climbed fixes: on Qwen3.5 0.8B (93% failure on MATH-500 due to "Wait" loops), it edited just 1,024 numbers—a millionth of parameters—raising MATH-500 from 7.4% to 37.4%. They call it weight-space programming.
Related event: AI Agent Fixes Broken Model by Editing Weights Without Running It(2 posts)→
More from coding & agent
- Ex-Meta CV researcher builds AdaL v2, a chat-driven video agent harness for GTM — Scobleizer · 2026-11-12
- chrisalbon: Show me 3 months and 1000 prompts, not one magic prompt — chrisalbon · 2026-11-12
- Agent autonomously replaces Gazebo's Ogre rendering engine with Filament in 36 hours — facontidavide · 2026-11-12
- Guardian: open-source authority layer pins human approval to each exact agent tool call — manofstyle04 · 2026-11-12
- Sentry open-sources tattletale: trace Claude Code and Codex agent sessions for token inefficiency — zeeg · 2026-11-12
- Hermes launches open-source workspace where Claude Code, Codex agents collaborate like teammates — ycombinator · 2026-11-12