OmniConfess mitigates omni-modal hallucinations via token-level evidence scoring
October 5, 2026
OmniConfess is a training-free method that identifies hallucination sources in omni-modal LLMs by re-scoring candidate responses through channel-wise evidence interventions. It uses structured token confessions to reveal whether a model is relying on contradictory or irrelevant text, audio, or video data to sustain a commitment.
HOW THIS AFFECTS YOU
●
builderYou can use this to improve reliability in multimodal applications without retraining models.
●
researcherThis provides a new framework for evaluating evidential dependence in omni-modal architectures.