Report
GPT-6 Astra weight edit reportedly boosts Qwen3.5-0.8B math score in experiment
The edit reportedly changed just 0.016% of Qwen3.5-0.8B’s weights without gradient training in the researchers’ experiment.
TLDR
MIT CSAIL researchers report that, in their experiment, GPT-6 Astra changed 0.016% of Qwen3.5-0.8B’s weights without gradients or training. Under the same sampling settings, Qwen’s MATH-500 score rose from 7.9% to 48.1%, and its IFEval score tripled. The researchers attribute the improvement to correcting reasoning loops; the experiment does not show that the edit taught Qwen new mathematics.
Combined views
1.6K
7 Sources, first seen ago
