Weight-Space Programming: direct model-weight edits with a frontier LLM

Evidence record

Author report of GPT-6 Astra inspecting Qwen3.5-0.8B weights and editing a small subset; MATH-500 pass rates improved largely as chain-of-thought truncation decreased. The author notes non-generalization to a low-looping Gemma model.

Read the original source ↗

Claim-level citations (3)

• supportsA frontier model edited 0.016% of Qwen weights to shorten looping answers: The author reports a MATH-500 pass-rate increase from 7.9% to 48.1% after editing 131,072 parameters, approximately 0.016% of Qwen3.5-0.8B weights.

  MATH-500 evaluation and parameter-change accounting

• supportsA frontier model edited 0.016% of Qwen weights to shorten looping answers: The share of answers reaching the 32,000-token limit dropped from approximately 93% to about 33%, and a low-looping Gemma control did not show analogous improvement.

  Reasoning-loop analysis and Gemma 3 270M control

• supportsA frontier model edited 0.016% of Qwen weights to shorten looping answers: In the author's 50-forward-pass experiment, Qwen3.5-0.8B GSM8K increased from 13.8% to 67.1% and IFEval from 21.0% to 62.3%, while a low-looping Gemma 3 270M control did not show a material MATH-500 improvement.

  Weight-space programming blog, Section Fixing Reasoning Loops, Qwen benchmark table and Gemma control