RealFeedApp GPT-6-Astra edited 0.016% of Qwen3.5-0.8B weights
Researchers demonstrated that GPT-6-Astra can fix errors in Qwen3.5-0.8B by modifying just 0.016% of its weights without gradients. The new dspy-wsp tool extends this weight-space programming to decision models.
- Field
- applying language models
- What they did
- The team developed a method called 'Weight-Space Programming' that allows language models to directly edit the weights of other models without using gradients or training, and released a prototype optimizer named dspy-wsp.
- Why it matters
- This approach enables models to modify the weights of others to solve decision-making tasks, which can improve the accuracy and efficiency of existing programs without the need for full retraining.
#dsp#weight-space#llm#optimization#gpt#qwen
