Astra reduces Qwen's looping without fine-tuning by changing 0.016% of its weights

In a Nov 2 experiment at MIT CSAIL, Astra changed 131,072 weights in a small Qwen model to make it less likely to get stuck in reasoning loops.