MIT researchers use GPT-6 Astra to directly edit Qwen open weights and boost math performance
🤖 AI-generated content — The title and summary were produced automatically by artificial intelligence, without human editorial review.
MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab used OpenAI's GPT-6 Astra on November 2 to perform "weight-space programming." By directly editing roughly 0.016% of Qwen3.5-0.8B's weights, they raised its MATH-500 score from 7.9% to 48.1% primarily by correcting repetitive reasoning.
Continue in the app — vote & join in ➔Source: VentureBeat · via ahirlevel.hu