MIT researchers propose 'weight-space programming': GPT-6 Astra improves math scores by sixfold with only about 0.016% of Qwen3.5-0.8B weights adjusted.
MIT proposes a new approach to weight-space programming that bypasses traditional gradient training. By relying on minimal weight fine-tuning alone, it significantly unlocks the potential of small models, offering a fresh direction for model optimization.
Continue looking at this topic
Related contentWhat is Qwen? Choosing between cloud interfaces and local models ↗Hotly discussed viewsToo abstract. Vice President of Alibaba Group and CEO of Qwen Office, Chen Yusen, will appear in Li Jiaqi's livestream tonight to sell his own A... ↗News and AnalysisJoe Tsai of Alibaba: Open source is the only way to advance AI adoption; Qwen open models now exceed 470 versions ↗Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open the original article on ITHome