LLM similarity scoring sensitive to document position and context, study finds

A new research paper, "Semantic Needles in Document Haystacks," explores the sensitivity of Large Language Models (LLMs) when used as judges for document similarity. The study, conducted by Sinan Aksoy, reveals that LLMs exhibit positional biases, penalizing semantic differences more harshly when they appear earlier in a document. Furthermore, the presence of topically unrelated context can inflate similarity scores, and while all tested LLMs penalize negation most severely among perturbation types, they do so on vastly different scales. The research proposes a framework for auditing and comparing LLM scoring behaviors across various models and prompt variations. AI

IMPACT Highlights potential biases in LLM evaluation metrics, crucial for developing more reliable AI systems.

RANK_REASON Research paper published on arXiv detailing findings about LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• DagsHub
• Gotit.pub
• Hugging Face
• LLM
• ScienceCast
• Semantic Needles in Document Haystacks: Sensitivity Testing of LLM-as-a-Judge Similarity Scoring
• Sinan Aksoy

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →