OmniConfess: Eliciting Token Confessions to Mitigate Omni-Modal Hallucination

OmniConfess. Eliciting Token Confessions to Mitigate Omni-Modal Hallucination. By Huiqiang Rong and six others, from the Beijing University of Posts and Telecommunications and Nanyang Technological University. Published on the fifth of October. Its code repository is public and real: nineteen source files.

Omni-modal large language models unify text, images, audio and video in one model, and they still hallucinate. The failure the authors point to is a specific one: generation that leans on the wrong evidence. Existing inference-time methods can reduce hallucinations, but they rarely reveal which evidence is actually sustaining a given commitment in the response. A correction that cannot see what a claim rests on is working blind.

OmniConfess is training-free, and its first move is to stop generating. It fixes a candidate response and then interrogates it. The response is re-scored at token resolution while the evidence is intervened on in a controlled way, one channel at a time. What comes out is what the authors call a confession: a structured record, token by channel, of how the response depends on its evidence.

That confession is then put to work in two directions. Content that the confession shows to be grounded is preserved. Commitments driven by irrelevant or contradictory evidence are the ones that get corrected. The unit of judgment throughout is the individual token, and the question asked of each is the same: which channel is holding this up.

To test it, the authors built OmniHalluBench, a benchmark of three thousand five hundred and forty examples drawn from six datasets. It spans text, image, audio and video settings, and covers both judgment tasks and free-form generation. Experiments show OmniConfess mitigating hallucinations across those heterogeneous modality and task settings. The code and the benchmark are both publicly available.

If one idea carries over from this, it is the order of operations. Commit to an answer first. Then ask each channel of evidence what that answer owes it, token by token, and let the answer's own dependencies decide what stays and what changes.