Retrieval-Augmented Generation

Cross-Task Calibration and Failure Modes in RAG Hallucination Detectors

Effect to explainPartly addressed

Possible candidate · 3/5 runs8 papers report this88% from 2025+

Generated automatically from the limitations stated in 8 papers (ICLR, SIGIR, TheWebConference), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.

The problem

Current hallucination detection and mitigation mechanisms in retrieval-augmented generation fail to transfer across task formats: span-level detectors achieve near-zero recall on subtle conflicts, agreement modules collapse into high-abstention trivial classifiers, and score distributions require manual per-task calibration. Consequently, deploying RAG systems in heterogeneous pipelines either results in pervasive false alarms that discard valid outputs or unflagged hallucinations in grounded summarization and multi-hop QA. Without understanding the operational boundaries and calibration dynamics across varying answer structures, hallucination filters cannot reliably act as automated triggers or guardrails.

Why it matters

Hallucination detectors can serve as reliable, plug-and-play triggers for secondary retrieval and automated abstention across mixed RAG pipelines without requiring task-specific ground-truth calibration data.

Ways to approach it

Prior-work checks are free with an account. Results someone already ran are shown to everyone.

1
Benchmark transferability and score calibration of existing open-source hallucination detectors (e.g., LettuceDetect, RAGTruth baselines, agreement-based checkers) across distinct task formats (QA, long-context extraction, grounded summarization) measuring calibration error (ECE), span-level F1, and precision-recall curves under uniform versus task-specific thresholds.
2
Evaluate abstention-tuned models (e.g., 'I don't know' fine-tuned Llama/Mistral) across PopQA, HotpotQA, and multi-document summarization to measure the exact tradeoff between true hallucination reduction and unnecessary abstention on previously correct responses.
3
Formulate a density-aware, task-conditioned calibration framework that adjusts detection thresholds based on output grounding density (e.g., sentence-level context overlap) rather than static global cutoffs, measured by the reduction in false-negative rates on summarization without degrading precision on factual QA.

Have a different approach?

Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.

Why it might fail

If upstream foundation models become sufficiently reliable through native long-context reasoning that post-hoc hallucination detection in RAG becomes obsolete, or if subtle hallucination detection remains fundamentally intractable without task-specific labeled data.

Evidence

Each paper's own statement of the limitation, verbatim.

SeaRAG: Reducing Hallucination in Retrieval-Augmented Generation via Statement-Entity Adaptive RankingTheWebConference 2026
Absolute accuracies remain low (e.g., 35.96% on PopQA, 36.30% on HotpotQA), indicating the method only partially mitigates hallucination on hard knowledge-intensive tasks
LUMINA: Detecting Hallucinations in RAG System with Context–Knowledge SignalsICLR 2026
64% of false negatives come from the summarization task, where mostly-grounded responses dilute the hallucination signal; task-dependent score distributions require per-task calibration
Fast and Faithful: Real-Time Verification for Long-Document Retrieval-Augmented Generation SystemsSIGIR 2026
Long-context benchmark hallucinations are injected via controlled prompting into LLM-generated responses with automatic annotation, and their 32K model actually trails LettuceDetect-large on the standard RAGTruth benchmark
Beyond RAG vs. Long-Context: Learning Distraction-Aware Retrieval for Efficient Knowledge GroundingICLR 2026
Requires task-specific training data with ground-truth answers and LLM-in-the-loop evaluation signals for training; performance on the Hallucination task degrades into degenerate strategies when trained on it, so it is only evaluated zero-shot there
Sufficient Context: A New Lens on Retrieval Augmented Generation SystemsICLR 2025
Fine-tuning open-source models (Llama 3.1 8B, Mistral 3 7B) with 'I don't know' targets does not reliably reduce hallucination—it can raise abstention at the cost of fewer correct answers
RARR Unraveled: Component-Level Insights into Hallucination Detection and MitigationSIGIR 2025
The agreement module with the 8B model flags 69.6% of inputs as hallucinatory, behaving nearly like a trivial all-hallucination classifier (which achieves recall 1.0 at precision 0.155/0.618), so high recall figures partly reflect over-flagging driven by absence of evidence rather than genuine detection
Bridging External and Parametric Knowledge: Mitigating Hallucination of LLMs with Shared-Private Semantic Synergy in Dual-Stream KnowledgeEMNLP 2025
Hallucination detection can miss cases where the model produces stable but incorrect answers, or coincidentally consistent responses to varied prompts, causing missed or false retrieval triggers
RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language ModelsACL 2024
Span-level detection remains weak even for the fine-tuned model (52.7% F1), and recall on subtle hallucination types is low (e.g., 2.5% span-level recall for subtle conflict with GPT-3.5 prompting)
Hallucination suppression requires multiple candidate responses per instance; with only two candidates, 122/450 instances had no response meeting the strict no-hallucination criterion