Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UCSC at SemEval-2025 Task 3: Context, Models and Prompt Optimization for Automated Hallucination Detection in LLM Output

Sicong Huang, Jincheng He, Shiyuan Huang, Karthik Raja Anandan, Arkajyoti Chakraborty, Ian Lane

TL;DR

The paper addresses the challenge of locating hallucinated content in LLM outputs across 14 languages. It proposes a three-stage pipeline—context retrieval, hallucination detection, and span mapping—augmented with MiPROv2-based prompt optimization and optional multi-system ensembling. Across languages, the approach achieves top-tier IoU and Corr scores, demonstrating that grounding responses in retrieved context and refining prompts substantially improve hallucination localization. The findings highlight the importance of context, the practicality of prompt optimization, and the variability in human annotations, with the authors releasing code and results to support further research.

Abstract

Hallucinations pose a significant challenge for large language models when answering knowledge-intensive queries. As LLMs become more widely adopted, it is crucial not only to detect if hallucinations occur but also to pinpoint exactly where in the LLM output they occur. SemEval 2025 Task 3, Mu-SHROOM: Multilingual Shared-task on Hallucinations and Related Observable Overgeneration Mistakes, is a recent effort in this direction. This paper describes the UCSC system submission to the shared Mu-SHROOM task. We introduce a framework that first retrieves relevant context, next identifies false content from the answer, and finally maps them back to spans in the LLM output. The process is further enhanced by automatically optimizing prompts. Our system achieves the highest overall performance, ranking #1 in average position across all languages. We release our code and experiment results.

UCSC at SemEval-2025 Task 3: Context, Models and Prompt Optimization for Automated Hallucination Detection in LLM Output

TL;DR

The paper addresses the challenge of locating hallucinated content in LLM outputs across 14 languages. It proposes a three-stage pipeline—context retrieval, hallucination detection, and span mapping—augmented with MiPROv2-based prompt optimization and optional multi-system ensembling. Across languages, the approach achieves top-tier IoU and Corr scores, demonstrating that grounding responses in retrieved context and refining prompts substantially improve hallucination localization. The findings highlight the importance of context, the practicality of prompt optimization, and the variability in human annotations, with the authors releasing code and results to support further research.

Abstract

Hallucinations pose a significant challenge for large language models when answering knowledge-intensive queries. As LLMs become more widely adopted, it is crucial not only to detect if hallucinations occur but also to pinpoint exactly where in the LLM output they occur. SemEval 2025 Task 3, Mu-SHROOM: Multilingual Shared-task on Hallucinations and Related Observable Overgeneration Mistakes, is a recent effort in this direction. This paper describes the UCSC system submission to the shared Mu-SHROOM task. We introduce a framework that first retrieves relevant context, next identifies false content from the answer, and finally maps them back to spans in the LLM output. The process is further enhanced by automatically optimizing prompts. Our system achieves the highest overall performance, ranking #1 in average position across all languages. We release our code and experiment results.
Paper Structure (37 sections, 1 equation, 1 figure, 8 tables)

This paper contains 37 sections, 1 equation, 1 figure, 8 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Related work
  4. Task Description
  5. System Overview
  6. Context Retrieval
  7. Context from Questions
  8. Context from Claims
  9. Hallucinated Content Detection
  10. Direct Text Extraction
  11. Verification with Knowledge Graph
  12. Minimal Cost Revision
  13. Span Mapping
  14. Substring Match
  15. Fact-to-Span Mapping
  16. ...and 22 more sections

Figures (1)

  • Figure 1: The UCSC hallucination detection framework. We retrieve context from external sources, identify false content in the answer, and then map these errors back to specific spans in the LLM output. In multilingual settings, we explore retrieving context either in the original language or in English by translating the question. In all cases the hallucinated content generated in the second step remains in the original language and is mapped to the answer.