Published August 11, 2026 | Version v1

Soft Ranking in Large Language Models: Output Asymmetry, Epistemic Calibration, and Transparency

Authors/Creators

Description

Soft Ranking in Large Language Models: Output Asymmetry, Epistemic Calibration, and Transparency

Working paper, August 2026.

This paper identifies and names a recurring pattern in frontier large language models: soft ranking. Models frequently include multiple explanatory options for open questions while systematically lowering the status of some options through ordering, differential elaboration, and epistemic status markers. Formal openness is preserved, but practical salience and conversational friction are altered.

The central question is whether these asymmetries primarily reflect accurate epistemic calibration to existing research programmes, or whether they also contain an architectural residualisation effect that operates independently of the underlying evidence. Qualitative multi-model observations and a within-model intervention test are reported. Soft ranking is situated against existing literatures on position bias, framing effects, and preference collapse under RLHF. A controlled matched-question experimental design is proposed to separate calibration from residualisation, along with a transparency protocol (Dynamic Epistemic Unbundling).

The paper does not argue for the equal standing of any particular worldview. It identifies a measurable structural phenomenon in how models organise competing explanations and offers tools for making that structure visible.

Keywords: large language models, soft ranking, choice architecture, position bias, framing effects, RLHF, epistemic transparency

Status: Working paper. Core observational claims are supported by multi-model patterns and a qualitative intervention test. Causal architectural claims remain at the level of inferred mechanisms and hypotheses.

Files

Files (15.6 kB)

Name Size
md5:b29b45a69e08cfb6df8d324edc270ff4
15.6 kB Download