Query–Document Dense Vectors for LLM Relevance Judgment Bias Analysis

Samaneh Mohtadi, Gianluca Demartini · 2026

Read Query–Document Dense Vectors for LLM Relevance Judgment Bias Analysis by Samaneh Mohtadi, Gianluca Demartini (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-01751.txt.

Description

Search evaluation increasingly uses large language models to judge whether documents are relevant, replacing costly human assessors. Samaneh Mohtadi and Gianluca Demartini ask a sharper question than average accuracy: do these models make patterned mistakes, judging wrongly in predictable kinds of cases?

Their method embeds each query-document pair as a single point in a shared space. It then clusters the pairs. Relevance is treated as a relation, not a property of one text. On TREC Deep Learning data from 2019 and 2020, disagreements between models and humans clustered in specific groups, especially definition-seeking, policy-related, and ambiguous queries. Pinpointing these hotspots lets evaluation designers target model weaknesses instead of trusting a single average score.

Catalog record

CREATORSamaneh Mohtadi, Gianluca Demartini
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-01751

Browse the full library · About the Tower of Babel