GPT-6 Astra edits 0.016% of Qwen3.5-0.8B's weights, lifting math accuracy from 7.9% to 48.1%

AIRohan Paul relays research from MIT CSAIL showing that GPT-6 Astra improved Qwen3.5-0.8B's math accuracy on MATH-500 from 7.9% to 48.1% without training. Astra edited only 0.016% of the weights after tracing a repetition loop to a signal written by two layers, then subtracting that signal.

Rohan Paul

@rohanpaul_ai

View on X

MASSIVE. MIT researchers showed that GPT-6 Astra can improve another AI model by changing its weights directly, without training.

Here, GPT-6 Astra raised Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 by editing 0.016% of its weights, without training.

Astra got Qwen’s weights in a Python notebook, plus a budget of 50 test runs on practice math problems. It saw the model kept repeating itself until it hit its length limit, traced that habit to a single internal signal written by 2 of its layers, and subtracted that signal out of the weights.

With the loop gone, the model could finish answers it was already capable of solving.

That single targeted edit changed 0.016% of the weights and lifted math accuracy from 7.9% to 48.1%.

Image from @rikiyatakehi's post — Rikiya Takehi@rikiyatakehi

Source details

Rohan Paul · x.comPublished Nov 2, 3:17 PM ET · added here 3:48 PM ET

Original title: MASSIVE. MIT researchers showed that GPT-6 Astra can improve another AI model by changing its weights directly, without training.

GPT-6 Astra edits 0.016% of Qwen3.5-0.8B's weights, lifting math accuracy from 7.9% to 48.1%

Topics