OmniHalluc-L benchmark shows open-weight omni models at 32-41% strict-pair accuracy on long-form hallucination, raised to 36-51% by Modality-Perturbation Reliability Calibration that fuses audio-negative probe shifts with native confidence.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.MM 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
OmniHalluc-L: Counterfactual Benchmarking and Modality-Perturbation Reliability Calibration for Long-Form Omni Hallucination
OmniHalluc-L benchmark shows open-weight omni models at 32-41% strict-pair accuracy on long-form hallucination, raised to 36-51% by Modality-Perturbation Reliability Calibration that fuses audio-negative probe shifts with native confidence.