OmniHallu: Unified Hallucination Detection for Cross-Modal Comprehension and Generation in Multimodal Large Language Models
Jianjiang Yang, Peihang Li, Shanqing Xu, Mengchen Qian, Lu Zhang, Meng Luo
cs.CL · cs.CV
Quick verdict
Claims unified cross-modal hallucination detection across six tasks; benchmark is solid, but gains hinge on a stronger backbone than baselines use.
Summary
OmniHallu proposes a single pipeline for detecting hallucinations in multimodal LLM outputs across six tasks: image-, video-, and audio-to-text comprehension plus text-to-image, -video, and -audio generation. The paper contributes OmniHallu-Bench, 10,000 samples with claim-level human annotations (full consensus of three annotators required, Fleiss' kappa 0.83-0.89), and a decompose-verify-aggregate multi-agent system: GPT-4.1 splits outputs into atomic claims, modality-specific expert ensembles (e.g., Grounding DINO 1.5 Pro, Qwen2.5-VL-72B, SALMONN) verify each claim, and GPT-5.2 aggregates evidence. A GRPO-trained Qwen2.5-VL-7B verifier abstains on hard claims, cutting expert calls by 66% (filter mode Mac.F1 80.52 vs 83.15 for the aggregator mode in Table 8). Reported Mac.F1 ranges from 82.95 (I2T) to 73.39 (T2A), with gains of +3.4 to +8.1 over the strongest baseline per task and a consistent image > video > audio gradient. Key limitations: with a matched GPT-4.1 backbone the framework scores below UNIHD on both image tasks (78.44 vs 78.88; 76.79 vs 78.45); video/audio baselines are limited to Self-Check; the verifier trains on a split of the benchmark itself; and no code or data release is stated.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single decompose-verify-aggregate pipeline can detect claim-level hallucinations uniformly across I2T, V2T, A2T, T2I, T2V, and T2A.
- OmniHallu-Bench is the first 10,000-sample benchmark with claim-level human annotations and rationales covering all six cross-modal tasks (Table 1).
- The framework improves Mac.F1 by +3.4 to +8.1 over the strongest baseline on every task, with a consistent image > video > audio gradient (Section 6.1).
- A GRPO-trained compact verifier approximates the multi-agent decision boundary, reducing expert calls by 66% (filter mode Mac.F1 80.52 vs 83.15 aggregator mode, Table 8).
- Detection difficulty follows object > attribute > event > relation across modalities, governed by compositional complexity rather than perception alone (Section 6.5).
Strengths
Benchmark scale and annotation rigor
10,000 samples across six tasks with claim-level labels; full-consensus protocol among three annotators, 24.3% disagreement removal, and Fleiss' kappa 0.89/0.86/0.83 reported before consensus filtering (Section 4).
Uniform cross-modal formulation
One formal definition and a four-type taxonomy (object/attribute/relation/event) applied identically to all six tasks (Section 3), enabling the comparisons Table 1 shows no prior benchmark (POPE, AMBER, FactVC, MHaluBench) supports.
Thorough ablation program
Cost-aware deployment analysis
Calibration-aware verifier design
Weaknesses
Matched-backbone negative result on image tasks is not stated
No structured detection baseline for video/audio
Verifier training data drawn from the benchmark itself
Unexplained vision-language verifier on audio tasks
No release or reproducibility plan
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
OmniHallu (arXiv:2609.11244) targets claim-level hallucination detection across six tasks — I2T, V2T, A2T (comprehension) and T2I, T2V, T2A (generation) — under one decompose–verify–aggregate architecture. Three central claims: (1) OmniHallu-Bench is the first 10,000-sample, claim-level, rationale-backed benchmark spanning all six tasks (Table 1); (2) the multi-agent pipeline beats the strongest per-task baseline by +3.4 to +8.1 Mac.F1 (Section 6.1, Table 3); (3) a GRPO-trained verifier cuts expert calls by 66% at minimal loss (Section 5, Table 8). The scope is genuinely broader than prior benchmarks — POPE, AMBER, FactVC, MHaluBench are all single-modality or single-pair per Table 1 — and applying one four-type taxonomy (object/attribute/relation/event) uniformly across modalities is what makes side-by-side comparison meaningful at all.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is OmniHallu?
OmniHallu is a framework plus benchmark (OmniHallu-Bench, 10,000 human-annotated samples) for detecting hallucinations — outputs that contradict or misrepresent the input — in multimodal LLMs across six tasks: image-, video-, and audio-to-text, and text-to-image, -video, and -audio. It decomposes outputs into atomic claims, verifies each claim with modality-specific expert models, and aggregates evidence with a reasoning model.
How well does OmniHallu perform?
Reported macro-F1 runs from 82.95 (image-to-text) down to 73.39 (text-to-audio), beating the strongest per-task baseline by +3.4 to +8.1 points, with a consistent image > video > audio difficulty gradient. Note: under an equal GPT-4.1 backbone, the framework scores below the UNIHD baseline on both image tasks.
How much cheaper is the trained verifier?
The GRPO-trained verifier handles easy claims so that only 34.2% of samples need the full expert ensemble — a 66% reduction in expert calls. Estimated cost falls from about $0.12 to $0.05 per sample and latency from about 15s to 6s (Table 7).
Is the OmniHallu-Bench dataset or code publicly available?
The paper does not state any code or data release. All quantitative results depend on closed APIs (GPT-5.2, GPT-4.1, Grounding DINO 1.5 Pro), so independent verification would currently require rebuilding the annotation pipeline from the public source datasets it lists.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 appear in the extraction only as caption text; the plotted values in Figure 2 (distribution statistics) and Figure 4 (per-type performance) are not recoverable, so figure-internal trends could not be independently checked.
- Per-task sample counts are given only as ratios in prose (Section 4); exact per-task n would live in Figure 2 and could not be verified from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.