Article by Shenchao TechFlowNo training required, no data needed, no gradient descent.Researchers at MIT fed the weights of an open-source small model into GPT. MIT researchers use GPT-6 Astra to enhance Qwen's mathematical ability by 6xTechFlowRelease Time: 2026/11/03 01:44:07ShareBOOSTalpha$0.0₄167933.6%SPACE$0.003342-4.89%OPENAI----TOKEN$0.001689-2.36%LLM----SummaryMIT researchers used GPT-6 Astra to enhance Qwen3.5-0.8B’s math performance, increasing it from 7.9% to 48.1% on MATH-500. The method, called Weight-Space Programming, corrected a neural network flaw without retraining.