What Qwen is thinking when it gets stuck saying "Wait"
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A write-up by Alice Corp's researcher (alisorcorp.github.io) examines an edit made by GPT-6 Astra to Qwen3.5-0.8B that cuts its looping on math problems. The edit, a rank-one change to 64 neurons in two layers, flips a "doubt" direction, lifting MATH-500 accuracy from 7.9% to 48.1%.