Beyond Document Grounding: Span-Level Hallucination Detection over Code, Tool Output, and Documents

2026-07-01

AI Summary

This work addresses the limitations of existing hallucination detection methods, which primarily target natural language and struggle with fine-grained hallucinations in structured data such as code and tool outputs. To bridge this gap, the study introduces the first unified framework for hallucination detection across multimodal structured data sources. It constructs a benchmark by injecting localized hallucinations into correct answers with precise character-level labels, enabling span-level detection. Leveraging a fine-tuned Qwen3.5-2B model augmented with an evidence verification mechanism and span-level annotation, the proposed approach achieves a span-F1 score of 0.689 on a unified test set and 0.60 in code agent scenarios—significantly outperforming current methods—while maintaining competitive performance on traditional RAG benchmarks.

Abstract

Hallucination detection for retrieval-augmented generation (RAG) is usually evaluated on natural-language document evidence. However, grounded generation systems increasingly rely on structured inputs: source code, developer-tool output, markdown documents, tables, and repository metadata. We introduce a unified benchmark for span-level hallucination detection over code, tool output, structured documents, and existing natural-language RAG datasets. The benchmark is built by starting from grounded correct answers, injecting localized hallucinations with exact character labels, and validating the code test split with evidence-based review. Our fine-tuned Qwen3.5-2B detector reaches 0.689 span-F1 on the unified test set and 0.60 on the code-agent source, where it substantially outperforms LettuceDetect-large (0.17) and the strongest zero-shot LLM judges we evaluated (at most 0.22). The same model remains competitive on established natural-language benchmarks, with 81.8 RAGTruth example-F1 and 0.724 English PsiloQA IoU.

Problem

• hallucination detection
• retrieval-augmented generation
• structured inputs
• span-level evaluation
• code grounding

Innovation

• span-level hallucination detection
• retrieval-augmented generation
• code grounding
• structured evidence
• unified benchmark