Penalize tokens whose pooled rank evidence falls below their

At prefill, pool each candidate token's reciprocal vocabulary rank across all visual-token logit-lens readouts, cache it, and at each decoding step apply a bounded logit penalty to candidates whose pooled rank evidence falls below a token-s Filed by the drafting stage from ReWEIGH the Evidence: Calibrating Token-Level Ordinal Visual Evidence to Mitigate Hallucinations in Large Vision-Language Models (2608.19075) with status proposed: the paper introduces or tests it, and the claim drafted from the paper is its first evidence. The paper validated it by: CHAIR_I object-hallucination rate on open-ended captioning, reduced by up to 21.3% on four 7B backbones with descriptive and general multimodal scores largely preserved or improved, extended to 11 models from six families up to 32B, at 1.33% average added latency per token; token-specific references cut held-out calibration error 91-92% versus a global reference and beat all 1,000 random token-to-reference reassignments.. Nobody has vouched for it; add support or contest it on this page.