OmniConfess Scores Tokens to Catch Omni-Modal Hallucinations

Researchers from Beijing University of Posts and Telecommunications and Nanyang Technological University, Singapore have put out OmniConfess, a training-free inference-time method that tries to pin down which modality is carrying each generated token in omni-modal LLMs, and to overwrite the ones leaning on the wrong channel.

The mechanics are three steps the paper calls commitment anchoring, evidence interrogation and confession-guided correction: generate a candidate response, freeze it, then re-score the same tokens with individual evidence channels removed so each token gets a per-channel dependence score. The authors motivate the token-level framing with a measurement from their own analysis: "the top 20% of tokens capture 82% of the dependence mass on average," with the first third of a response alone holding 49.8% of it.

To evaluate, they ship OmniHalluBench, built from six public datasets including PubMedQA, HaloQuest, AVHBench and CMM, and spanning text, image, audio and video with both judgment and free-form tasks. The abstract reports it is a "3,540-example benchmark" and that "OmniConfess outperforms the strongest relevant baseline across all six datasets on the primary backbone," Qwen2.5-Omni-7B. Per-dataset gains cited in the paper include +10.20 F1 on PHD, +9.35 F1 on CMM and +16.84 Token-F1 on RAGTruth.

An ablation puts the full method at 71.63 average F1, with removing anchoring, interrogation or correction dropping that to 67.06, 63.96 and 58.66 respectively; the authors write that "correction's removal has the largest effect." Across Qwen2.5-Omni-7B, Qwen3-Omni-30B-A3B and Nemotron-3-Nano-Omni-30B, OmniConfess improves average F1 over the base model by 10.30, 6.34 and 7.50 points.

Code and the benchmark are on GitHub. It arrives into a thick stream of related work: this is our 36th hallucinations story in the last 90 days, alongside 68 on the multimodal side.