{"ID":5551683,"CreatedAt":"2026-07-02T01:54:51.863792489Z","UpdatedAt":"2026-07-04T12:48:09.865479953Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2607.00895","arxiv_id":"2607.00895","title":"Beyond Document Grounding: Span-Level Hallucination Detection over Code, Tool Output, and Documents","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.","short_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 benchm...","url_abs":"https://arxiv.org/abs/2607.00895","url_pdf":"https://arxiv.org/pdf/2607.00895v1","authors":"[\"Ádám Kovács\",\"Bowei He\",\"Xue Liu\",\"István Boros\",\"Szilveszter Tóth\",\"Gábor Recski\"]","published":"2026-07-01T13:01:42Z","proceeding":"cs.CL","tasks":"[\"cs.CL\"]","methods":"[\"RAG\",\"Large Language Model\"]","has_code":false}