MIT researchers show GPT-6 Astra can improve other AI models by editing their weights
In an experiment, Astra raised Qwen3.5-0.8B's accuracy on MATH-500 from 7.9% to 48.1% by changing only 0.016% of the model's weights.
MIT researchers have shown that GPT-6 Astra can improve other AI models by directly editing their weights without additional training.
In the experiment, Astra raised the accuracy of Qwen3.5-0.8B on the MATH-500 benchmark from 7.9% to 48.1% by changing only 0.016% of the neural network's weights.
The researchers gave Astra access to Qwen's weights through a Python notebook and allowed up to 50 test runs on math problems.
During analysis, Astra found that the model repeatedly loops through its own reasoning until it hits the response-length limit.
The cause was traced to a single internal signal formed by two neural-network layers. Astra removed its effect by directly adjusting the model's weights.
After that, Qwen stopped looping and was able to finish solving problems it could already handle in principle.
As a result, one targeted change affecting only 0.016% of the weights increased math-task accuracy from 7.9% to 48.1%.
Why it matters
AnalysisIt suggests that AI systems may be able to debug and optimize other models through direct weight manipulation, potentially reducing the need for costly full retraining.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Claude Becomes Smarter
Claude becomes smarter, indicating possible improvements in the model's capabilities.
GPT-6 Astra reportedly edits another AI model's weights without retraining
GPT-6 Astra is claimed to modify only 0.016% of Qwen3.5-0.8B's weights without gradients or retraining, improving reasoning, math results, and instruction following.
Frontier Models
The most powerful class of artificial intelligence at the forefront of global research (Claude 3.7 Sonnet, OpenAI o3/GPT-4.5, Gemini 2.0 Pro), defining the limits of modern reasoning, autonomy, and coding capabilities.
Part I. AI Agent Architecture: The Foundations
The eleven foundational building blocks of modern autonomous agents: model taxonomy, the agentic runtime loop, advanced prompting, distillation, skills architecture, memory systems, function calling, and FinOps.