MIT used GPT-6 Astra to boost Alibaba's Qwen math performance
MIT researchers just gave Alibaba's Qwen AI model a serious math upgrade using GPT-6 Astra.
By tweaking only 0.016% of its neural network, Astra helped Qwen ditch repetitive thinking and solve problems way better, showing how one AI can actually improve another from the inside out.
Astra raised Qwen scores without retraining
The team at MIT's CSAIL let Astra interact with Qwen in a Python setup; no retraining was needed.
After Astra fixed some internal issues, Qwen's math scores jumped big time: MATH-500 went from 7.9% to 48.1%, and GSM8K soared from 13.8% to 67.1%.
The researchers say there are still limits; Astra could not identify which of two models had been trained to agree with users until it was given access to information about where the models were focusing their attention, and this experiment hints at cool new ways AIs might help each other get smarter in the future.