Faithful uncertainty in LLM agents: calibration vs utility tradeoff in practice[D]

The Google paper on metacognition for hallucination reduction makes a distinction that is underappreciated in benchmarks. Calibration is not about being right more often. It is about matching confidence to correctness. A perfectly calibrated model can still be wrong twenty five percent of the time. It just does not pretend otherwise. In agent systems this distinction matters more than in chat. A conversational model giving a hedged answer is slightly annoying. An agent with tool access acting confidently on