Do language models think alike when their answers match?

Across 16 language models (1.5B–72B parameters) solving 800 reasoning problems, representations align most on questions all models get wrong, not right. Post-decision activations diverge sharply even when early layers match, and shared information rarely causally influences predictions (1.5–5.5% flip rate under ablation). Convergence reflects common input processing, not shared reasoning logic.