The Epistemic Cost of Preference Optimization

Rian Atri

Track: tiny / short paper (up to 4 pages)

Keywords: Epistemics, Machine Learning

Abstract: Stateful LLMs often exhibit contradictions across related questions and struggle with logical deduction and abductive revision under explicit evidence. We study the influence of preference pressure and evidence access on epistemic consistency to avoid contradictions across multiple related questions. To quantify these effects, we propose a minimal “pressure ladder” and “evidence toggle” evaluation protocol. We demonstrate the feasibility of this protocol through an empirical pilot and a reporting-gap audit. Finally, we propose a reporting checklist for logical reasoning papers to track epistemic consistency, evidence sensitivity, grounding/citation integrity, and a calibration proxy alongside standard metrics.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 149