From Semantic Alignment to LLM Hallucination Origins: An Algorithmic Approach
Abstract: We introduce a new text alignment algorithm that can produce fine-grained alignment between a query document and database documents. Our work explores two under-explored directions: i) alignment granularity at text segment level as opposed to traditional entire document retrieval, and ii) alignment directed by semantic similarity instead of exact matches.
We utilize text embeddings produced by Large Language Models (LLM) and perform efficient queries through nearest neighbor data structures. We also introduce an attack strategy exploiting temporal inconsistencies to induce hallucinations in Large Language Models (LLMs) and apply our alignment algorithm to trace these hallucinations back to their possible origins in training data. By creating a database of relevant web documents using keyword filtering on Common Crawl data, our approach demonstrates the effectiveness of identifying candidate origins of LLM hallucinations.
Paper Type: Short
Research Area: NLP Applications
Research Area Keywords: Text Alignment, LLM Hallucinations, LLM Memorization
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 5218
Loading