GPT-6 Astra Improved Another AI Model by Directly Editing Its Weights In an experiment conducted by MIT researchers, GPT-6 Astra improved the mathematical abilities of the Qwen3.5-0.8B language model by directly modifyin

👨‍💻 GPT-6 Astra Improved Another AI Model by Directly Editing Its Weights In an experiment conducted by MIT researchers, GPT-6 Astra improved the mathematical abilities of the Qwen3.5-0.8B language model by directly modifying its parameters, without training it on new data. The smaller model's score on the MATH-500 benchmark jumped from 7.9% to 48.1% — a more than sixfold improvement. 🔍 Astra was given direct access to Qwen's weights and the ability to test its responses. It found that Qwen often got stuck in reasoning loops, repeating the same words instead of reaching an answer. Astra identified which components of the AI model were responsible for these loops and used Python to modify just 0.016% of its weights — 131,072 parameters out of 800 million. After the changes, the model got stuck in these loops far less frequently. 📌 On another math benchmark, GSM8K, accuracy rose from 13.8% to 67.1%, while its score on IFEval, which measures instruction-following ability, increased from 21% to 62.3%. The percentage of math responses that hit the 32,000-token limit dropped from 93% to 33%. Essentially, Astra fixed a problem that had been preventing the model from using abilities it alr