reasoningcritiqueneutralChain-of-Thought rationales are often poorly aligned with efficient machine reasoning despite improving LLM performance on difficult tasksNeural and Evolutionary Computing27 Jul 2026http://arxiv.org/abs/2606.29354v1