Consistent but Miscalibrated: Evaluating LLM Limitations for Risk Communication in Natural Language
arXiv:2607.03882v1 Announce Type: cross Abstract: LLMs are increasingly deployed as post-hoc explainers of AI-generated outputs, yet it remains unclear whether they can reliably communicate probabilistic information in natural language. For this role to be viable, models must produce identical verbal descriptions for identical inputs, and select descriptions that accurately reflect the magnitude of the underlying numerical quantities. We evaluate whether nine LLMs meet these requirements within a two-stage prediction pipeline, in which an upstream model has produced probabilistic outputs chara
The increasing deployment of LLMs as explainers for AI-generated outputs necessitates understanding their reliability in communicating probabilistic information.
Reliable risk communication by LLMs is critical for trust and effective human-AI interaction in high-stakes applications, potentially impacting adoption and regulatory frameworks.
Our understanding of LLMs' current limitations in accurately conveying nuanced probabilistic outputs, highlighting a gap between natural language consistency and numerical calibration.
- · AI safety researchers
- · Developers of robust LLM risk communication systems
- · Regulators focused on AI explainability
- · LLM deployment in high-stakes probabilistic explanation without guardrails
- · Systems relying solely on LLMs for critical risk communication
Increased focus on developing LLMs capable of precise and calibrated probabilistic natural language generation.
Demand for new evaluation metrics and benchmarks specifically designed for the fidelity of LLM risk communication.
Potential for specialized 'risk communication' LLM architectures or fine-tuning approaches to emerge, leading to wider adoption of AI in sensitive domains.
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