---
title: Semantic Calibration
type: vocabulary
url: "https://www.envisioning.com/vocab/semantic-calibration"
summary: "The property of a base LLM's confidence being accurate at the level of answer meaning, not just next-token probability, and the finding that RLHF and chain-of-thought reasoning break it."
year: 2025
generality: 0.40
---

# Semantic Calibration

The property of a base LLM's confidence being accurate at the level of answer meaning, not just next-token probability, and the finding that RLHF and chain-of-thought reasoning break it.
Semantic calibration is a language model's ability to assign accurate confidence to the meaning of its answers, not just to its next-token probabilities. Preetum Nakkiran, Arwen Bradley, Adam Golinski, Eugene Ndiaye, Michael Kirchhof, and Sinead Williamson, researchers at Apple, introduced the term in "Trained on Tokens, Calibrated on Concepts: The Emergence of Semantic Calibration in LLMs," posted to arXiv (2511.04869) on 6 November 2025 and accepted to ICLR 2026. Standard next-token calibration only requires a model's token-level probabilities to match observed frequencies. Semantic calibration asks more: when answers are grouped into equivalence classes by meaning (two differently worded correct answers count as one class), the model's distribution over those classes must also track the true probability of being correct.

The paper's theoretical contribution is a general notion it calls B-calibration: calibration parameterized by a choice of equivalence classes, of which semantic calibration is the case where the classes are defined by shared meaning. It argues that semantic calibration emerges as a byproduct of ordinary next-token prediction, via a known connection between calibration and local loss optimality, and predicts that a base model will be semantically calibrated whenever it can easily predict its own distribution over semantic answer classes before generating a response.

Testing three consequences of that prediction on open-domain question-answering tasks, the authors find base LLMs, pretrained only with next-token prediction and never explicitly trained on confidence, are semantically calibrated despite that. Reinforcement-learning-based instruction tuning systematically breaks this calibration, and so does chain-of-thought reasoning. These results are presented as the first principled account of when and why semantic calibration appears and disappears in LLMs, a question distinct from whether a model's calibration, once expressed, is faithful to its internal state.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/semantic-calibration)
