tl;dr: Large reasoning models (LRMs) produce impressive results—solving open math problems and winning IMO gold—but growing research shows their "chains of thought" often aren't faithful representations of internal reasoning: irrelevant or filler tokens work just as well, and 30-60% of "thinking steps" have minimal causal impact on outputs. Researchers like Subbarao Kambhampati argue LRMs are doing "approximate retrieval" rather than genuine step-by-step reasoning, with intermediate tokens serving to prime the model rather than narrate actual thought. The debate matters because trusting AI in non-verifiable domains requires knowing whether models are right for the right reasons.