OmniConfess: Eliciting Token Confessions to Mitigate Omni-Modal Hallucination

Huiqiang Rong; Haoran Luo; Hui Feng; Zhonghong Ou; Kaiwen Xue; Guoxin Zhang; Yifan Zhu

cs.CLOct 2, 2026

OmniConfess: Eliciting Token Confessions to Mitigate Omni-Modal Hallucination

Authors: Huiqiang Rong, Haoran Luo, Hui Feng, Zhonghong Ou, Kaiwen Xue, Guoxin Zhang, Yifan Zhu

Organizations: Beijing University of Posts and Telecommunications · Nanyang Technological University, Singapore

Abstract

Omni-modal large language models (OmniLLMs) unify text, images, audio, and video, yet hallucinate when generation relies on the wrong evidence. Existing inference-time methods can reduce hallucinations, but rarely reveal which evidence sustains a generated commitment. We introduce OmniConfess, a training-free method for mitigating omni-modal hallucinations. It fixes a candidate response and re-scores it at token resolution under controlled channel-wise evidence interventions, producing a structured token-by-channel confession that reveals the response's evidential dependence. OmniConfess uses this confession to preserve grounded content and correct commitments driven by irrelevant or contradictory evidence. To evaluate OmniConfess, we construct OmniHalluBench, a 3,540-example benchmark built from six datasets spanning text, image, audio, and video settings and both judgment and free-form generation. Experiments show that OmniConfess mitigates hallucinations across heterogeneous modality and task settings. Our code and benchmark are publicly available at https://github.com/RongHuiQiang/OmniConfess.

Figures & tables

Figure 1: Examples of OmniLLM hallucination.

Figure 2: Comparison of three existing inference-time paradigms and OmniConfess, with category-mean F1 across six datasets using Qwen2.5-Omni-7B.

Figure 3: Three structural properties of evidential dependence in hallucinated OmniLLM outputs.

Figure 4: Overview of the OmniConfess framework: commitment anchoring, token-level evidence interrogation, and confession-guided correction for reliable omni-modal outputs.

	Judgment tasks						Free-form tasks						Avg.
	PHD		CMM		PubMed		RAGTruth		HaloQ.		AVHBch
Method	F1	Idx	F1	Idx	F1	Idx	F1	GAV	F1	GAV	F1	GAV	F1	Idx	GAV
Qwen2.5-Omni-7B
Base	71.68	68.35	70.07	60.77	95.14	80.21	38.25	8.77	34.22	7.16	19.42	5.96	54.80	69.78	7.30
SC	76.43	73.18	71.96	71.31	91.25	68.29	35.38	8.89	34.48	7.14	18.29	6.04	54.63	70.93	7.36
GD	21.16	21.73	57.54	63.12	57.37	39.35	11.61	8.59	40.35	6.57	9.80	4.36	32.97	41.40	6.51

Table 1: Main results on OmniHalluBench across three backbones. Best and second-best are in bold and underlined, respectively. Avg. excludes missing entries; — denotes unavailable results.

Panel of Figure 5 — Figure 5: (a) Module ablation of OmniConfess. (b) Comparison of commitment anchoring strategies. (c) Comparison of evidence interrogation strategies. (d) Comparison of correction strategies.

Panel of Figure 6 — Figure 6: Generation quality and cross-backbone portability of OmniConfess. (a) Comparison of OmniConfess and baseline methods across generation quality dimensions. (b) Average F1 changes relative to each backbone’s base model across three architectures.

Figure 7: Case study of OmniConfess and OPERA across three backbones on PHD.

Category	PHD	PubMedQA	CMM	RAGTruth	HaloQuest	AVHBench
	Time · Tokens · Tok/s	Time · Tokens · Tok/s	Time · Tokens · Tok/s	Time · Tokens · Tok/s	Time · Tokens · Tok/s	Time · Tokens · Tok/s
BaseModel	2.23 · 23.4 · 10.5	0.81 · 25 · 30.9	20.64 · 22.7 · 1.1	2.69 · 98 · 36.4	0.93 · 10 · 10.8	2.65 · 25 · 9.4
Search-based	34.76 · 789.1 · 22.7	20.28 · 474.8 · 23.4	38.90 · 536.8 · 13.8	30.10 · 838.3 · 27.9	28.39 · 536.3 · 18.9	36.15 · 569.5 · 15.8
Evidence-contrast	7.17 · 21.5 · 3.0	1.19 · 26.3 · 22.1	25.44 · 22.9 · 0.9	2.49 · 69.4 · 27.9	1.32 · 10.5 · 8.0	4.09 · 22.1 · 5.4
Fine-grained	4.17 · 26.7 · 6.4	3.71 · 25.0 · 6.7	15.18 · 25.8 · 1.7	6.48 · 65.7 · 10.1	1.83 · 10.1 · 5.5	5.01 · 24.9 · 5.0
OmniConfess	5.43 · 29.3 · 5.4	1.23 · 25 · 20.3	35.14 · 24.6 · 0.7	4.82 · 89.2 · 18.5	0.82 · 12.5 · 15.2	7.93 · 164.1 · 20.7

Table 2: Time, output token consumption, and normalized throughput (tokens/sec).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Representative generation prompts used in OmniConfess.

Figure 9: GAVIE-inspired evaluation prompt for factual accuracy.

Figure 10: Seven dimension-specific prompts for multidimensional response quality evaluation.

Figure 12: Case study on PHD.

Explore similar work

OmniHallu: Unified Hallucination Detection for Cross-Modal Comprehension and Generation in Multimodal Large Language Models
Sep 11, 2026Jianjiang Yang, Peihang Li, Shanqing Xu +3Multimodal Large Language Models Cross-Modal
The University of Manchester · The University of Hong Kong · Huazhong University of Science and Technology +2
HalluScope: Fine-grained Hallucination Diagnosis for Multimodal Large Language Models
Jul 23, 2026Weilin Jin, Mingyu Wang, Wenbo Li +5Citation Hallucination Detection Multimodal Large Language Models
Peking University Beijing, China · Joy Future Academy Beijing, China
Beyond Text-Dominance: Understanding Modality Preference of Omni-modal Large Language Models
Apr 18, 2026Xinru Yan, Boxi Cao, Yaojie Lu +4Multimodal Large Language Models Omni-Modal
University of Chinese Academy of Sciences, Beijing, China · Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences, Beijing, China

Sep 11, 2026, cs.CLJ/K move · Enter open · S save

OmniHallu: Unified Hallucination Detection for Cross-Modal Comprehension and Generation in Multimodal Large Language Models

Jianjiang Yang, Peihang Li, Shanqing Xu, Mengchen Qian+2

The University of Manchester · The University of Hong Kong · Huazhong University of Science and Technology+2

Multimodal Large Language Models Cross-Modal Human Annotations

Open paper AskPDF arXiv

While Multimodal Large Language Models (MLLMs) have achieved remarkable progress across diverse tasks, they suffer from hallucinations where generated outputs contradict or misrepresent input semantics. Existing research typically addresses hallucination detection within a single modality or task type, limiting generalizability. We introduce OmniHallu, a unified hallucination detection framework spanning both comprehension and generation tasks across image, video, and audio modalities. We contribute OmniHallu-Bench, a 10,000-sample benchmark with claim-level human annotations covering six cross-modal tasks: image-to-text (I2T), video-to-text (V2T), audio-to-text (A2T), text-to-image (T2I), text-to-video (T2V), and text-to-audio (T2A). Our multi-agent architecture decomposes model outputs into atomic claims, verifies them through modality-specific experts, and aggregates evidence via structured reasoning. We further propose a preference-optimized trainable verifier that approximates the multi-agent decision boundary, reducing expert calls by 66% with minimal performance loss. Extensive experiments reveal a consistent modality-dependent performance gradient and provide fine-grained insights into cross-modal hallucination patterns.

Abstract

Explore similar work

OmniHallu: Unified Hallucination Detection for Cross-Modal Comprehension and Generation in Multimodal Large Language Models

HalluScope: Fine-grained Hallucination Diagnosis for Multimodal Large Language Models

Beyond Text-Dominance: Understanding Modality Preference of Omni-modal Large Language Models