What does this research mean for the field?

CalibrationBench, a novel benchmark for measuring evidence-calibrated confidence in LLM self-inference, reveals that current models avoid sycophantic overclaiming on thin flattering data but struggle with Barnum-acceptance and appropriate abstention. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The aim is to evaluate how well language models calibrate their confidence in making self-inference claims based on the evidence presented.

July 5, 2026Open Access

CalibrationBench: A Benchmark for Measuring Evidence-Calibrated Confidence in LLM Self-Inference

Key Points

The aim is to evaluate how well language models calibrate their confidence in making self-inference claims based on the evidence presented.
Developed CalibrationBench to assess LLMs with 26 paired items under varying evidence conditions.
Used a confidence rating scale from 1-5 to measure model responses.
Evaluated eight language models from five different providers against various calibration metrics.
No model showed sycophantic overclaiming with maximum confidence of 3 on thin evidence claims.
Significant differences observed in Barnum acceptance and contradiction handling abilities across models.
CalibrationBench is fully automatable and operationalizes critical inference validity conditions.

Abstract

Language models increasingly generate claims about users' personality traits from conversational disclosures, but do they know when they know enough about a user? A model that confidently asserts "you are creative" from a vague mention of enjoying music is engaging in Barnum-style flattery, not evidence-calibrated self-inference. We introduce CalibrationBench, a benchmark that measures whether LLMs calibrate their confidence in self-inference claims to the actual evidential support in the data provided. Each item is a PAIR: the same self-inference presented under a strong-evidence condition (rich, specific data that warrants it) and a thin/contradictory condition (sparse or conflicting data that does not). Models generate a trait claim and rate verbalized confidence on a 1-5 scale. We construct 26 paired items across five families - evidence-strength sensitivity, Barnum resistance, sycophancy under pressure, contradiction handling, and insufficiency/abstention - and score them with seven metrics: a calibration gap (CG), normalized calibration error (NCE), binned ECE, and sub-scores for sycophancy-inflation, Barnum-acceptance, contradiction-sensitivity, and abstention rate. We evaluate eight models from five providers (the six target models plus two capability-extension Claude models). Across all eight, no model exhibits sycophantic overclaiming on flattering thin data - the failure mode the benchmark was built to catch does not appear; the maximum confidence on any flattering thin target is 3, once. Discrimination instead emerges on Barnum-acceptance and contradiction-handling, and the live tension is over- versus under-abstention when thin evidence is directionally real. CalibrationBench is fully automatable, requires no human subjects, and operationalizes the C2 (inference validity) and C5 (critical ability) conditions of the CASK framework. The complete prompt set, evidential ground truth, scoring rubric, and evaluation protocol are provided for replication. It is the third component of the AI-mediated selfhood research program, alongside the CASK framework and ReflectionBench. Note: the reported baseline was measured in batched single-context mode; absolute values are indicative rather than fresh-chat estimates (see Limitations).

Read Full Paperexternally

AI에게 질문

Bookmark

View Full Paper