arXiv:2603.09309v2 Announce Type: replace Abstract: Verbalized confidence, in which LLMs report a numerical certainty score, is widely used to estimate uncertainty in black-box settings, yet the confidence scale itself (typically 0--100) is rarely examined. We show that this design choice is not neutral. Across six LLMs and three datasets, verbalized confidence is heavily discretized, with more than 78\% of responses concentrating on just three round-number values. To investigate this phenomenon, we systematically manipulate confidence scales along three dimensions: granularity, boundary place
This research provides timely insights into the fundamental workings and limitations of LLM metacognition, appearing as confidence in AI systems is increasingly critical for their real-world application.
Understanding how LLMs express confidence directly impacts system reliability, trust, and our ability to interpret and utilize their outputs in sensitive applications.
The explicit recognition of discretization and scale dependence in LLM confidence reports means that raw confidence scores can no longer be taken at face value without careful consideration of their elicitation method.
- · AI researchers focusing on interpretability
- · Developers of robust AI systems
- · Industries requiring high-assurance AI
- · Systems relying on naive interpretation of LLM confidence
- · LLMs with poorly designed confidence mechanisms
This work will lead to improved methods for eliciting and calibrating confidence in LLMs.
Better confidence calibration will enable more reliable AI agents and decision support systems.
Increased reliability and trust could accelerate the integration of AI into critical infrastructure and white-collar workflows.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI