RealFeedApp Qwen3.5-0.8B accuracy rises from 7.9% to 48.1%
GPT-6 Astra improved Qwen3.5-0.8B by directly editing weights without training, raising accuracy from 7.9% to 48.1% on MATH-500. The autonomous agent Beaker found a single-neuron edit in an hour that cuts loops and boosts accuracy by 30 points.
- Field
- training language models
- What they did
- Researchers from MIT CSAIL demonstrated that a language model can improve another model by directly editing its weights without training runs or gradient calculations. As an example, model GPT-6 Astra increased the accuracy of Qwen3.5-0.8B by 40.2 percentage points, raising the result from 7.9% to 48.1% on the MATH-500 test.
- Why it matters
- This method allows models to autonomously correct each other's errors, which is a step toward creating self-initiated continuous learning rather than external training processes.
#llm#weight editing#gpt-6 astra#qwen3.5#continual learning
