Report

MIT researchers use GPT-6 Astra to improve a Qwen model’s math-test score with weight edits

MIT researchers report Qwen3.5-0.8B’s MATH-500 score rose from 7.9% to 48.1% after editing approximately 0.016% of its weights.

2 Sources, ,

'Like gene editing for neural networks': MIT researchers used GPT-6 Astra to edit Qwen's open weights directly, boosting math performance

TLDR

The team calls its method “weight-space programming”: Astra directly changed the model’s numerical parameters without using conventional training to generate the edits. Researchers say the improvement came mainly from correcting repetitive reasoning. They have not established whether such edits save time or computing resources.

Today's Rank

—

Not ranked yet

Combined views

2 Sources, first seen 2h ago