reasoningfactbearishIn LLMs, detecting a reasoning trap is often decoupled from actually avoiding itComputation and Language02 Aug 2026http://arxiv.org/abs/2607.28478v1