LLMs Show Systematic Epistemic-Rhetorical Miscalibration
A recent study published on arXiv (2604.19768) presents a framework aimed at measuring the separation between rhetorical intensity and epistemic grounding within large language models. The researchers developed a taxonomy of triadic epistemic-rhetorical markers, operationalized through composite metrics: form-meaning divergence, the genuine-to-performed epistemic ratio, and rhetorical device distribution entropy. By examining 225 argumentative texts (approximately 0.6 million tokens) from human experts, non-experts, and LLM-generated sources, they uncovered a consistent epistemic signature across models. LLM texts utilize tricolon nearly twice as often as experts (Δ=0.95), whereas human authors use erotema at over twice the rate of LLMs. Additionally, performed hesitancy markers are found in LLM outputs at double the density of human texts, highlighting significant form-meaning divergence in LLM content.
Key facts
- Study from arXiv:2604.19768
- Proposes framework for quantifying epistemic-rhetorical miscalibration
- Triadic ERM taxonomy with FMD, GPR, RDDE metrics
- Analyzed 225 argumentative texts (~0.6M tokens)
- Compared human expert, human non-expert, and LLM sub-corpora
- LLMs use tricolon at nearly twice expert rate (Δ=0.95)
- Humans use erotema at more than twice LLM rate
- Performed hesitancy markers twice as dense in LLM output
Entities
Institutions
- arXiv