response-level-hallucination-detection-retrieval-augmented-settings-having-model-first-convert CLM-0245For response-level hallucination detection in retrieval-augmented settings, having the model first convert the reference documents into a relational SQL database and then check the response against SELECT query results raises precision and macro F1 over zero-shot direct prediction, over atomic-claim decomposition, and over flat JSON key-value extraction, without any task-specific fine-tuning.
Capability: Stating false facts confidently
Observed on
RAGTruth (QA, data-to-text, summarisation) and the task-oriented portion of DiaHalu; gains largest on backbones with decent text-to-SQL ability, and absent on a small low-text-to-S.
Sources
- supportsLeveraging Low-Level Symbolic Competences for Unsupervised Grounding in Hallucination DetectionValidated by: Macro precision/recall/F1 on RAGTruth response-level detection (Gemini-2.5-Flash: 61.7/88.9/71.3 vs direct prediction 48.2/95.4/61.6, atomic fact checking 65.7 F1, JSON key-value 64.9 F1; below fine-tuned RAG-HAT at 78.0), plus DiaHalu TOD F1 across Gemini-1.5 Pro, GPT-4 and Gemini-2.5-Flash, an ablation on strict ID-intersection prediction, and a manual audit of SQL parsing fidelity (85.33% complete).. Measured on two benchmarks with 900 pairs per RAGTruth task and 210 DiaHalu dialogues, against both unsupervised and fine-tuned baselines. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For response-level hallucination detection against reference documents, having the model first convert the reference into a relational SQL database and then check response claims by executing SELECT queries raises F1 over zero-shot direct prediction mainly by cutting false positives, and the gain shrinks or vanishes on backbones with weak text-to-SQL ability.Stating false facts confidently · reviewed by AI
- When retrieval for a summarization task pulls definitions by surface term match from a general encyclopedia, ambiguous terms fetch the wrong sense and the model incorporates that unsupported background into its output, lowering claim-level factuality below the no-retrieval baseline even when the prompt tells it to ignore unrelated retrieved content.Stating false facts confidently · reviewed by AI
- For instruction tuning an 8B open model on text-to-SQL, a curated set of under 900 examples selected to cover every table in the target schemas and score well on question-SQL alignment reaches higher execution accuracy than pipelines trained on tens of thousands to millions of samples, and dropping the schema-coverage term from the selection objective costs more accuracy than dropping any single quality term.Writing a reusable locator: regex, SQL, XPath · reviewed by AI
- An abstention rule built from self-consistency, where the model samples several answers and itself scores how many agree with a candidate, can be thresholded by conformal prediction to bound the hallucination rate at a chosen level on closed-book QA; against log-probability confidence it abstains much less often on long-form answers (Temporal Sequences) for the same error bound and matches it on short answers (TriviaQA).Saying it does not know · reviewed by AI
- On natural-language-to-SQL synthesis over multi-table databases, a model's accuracy overstates its ability whenever the same database or the same query appears in both training and test: splitting by database rather than by example cut the best system's exact-match accuracy from 33.0% to 12.4%, and accuracy falls further as the target database has more foreign keys.Writing a reusable locator: regex, SQL, XPath · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A replication where SQL-grounded verification fails to beat atomic-claim decomposition or direct prediction on RAGTruth macro F1 under matched backbones, or where the gain disappears once the multi-step baselines are given comparable token budgets.