Portfolio

Cross-Domain Limits of CoT Surface Features

A single-model measurement study showing that cheap reasoning-trace features are strong for math, partial for science, and weak for executable coding correctness.