AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

SHIFT: Smoothing Hallucinations by Information Flow Tuning for Multimodal Large Language Models

Sudong Wang, Yunjian Zhang, Yao Zhu, Enci Liu, Jianing Li, Yanwei Liu, Xiangyang Ji

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Decoding-time hallucination mitigation · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply SHIFT mutation-layer smoothing under greedy decodingLLaVA-1.5 · CHAIRCHAIR_S: 49.6 -> 43.8Improved
apply SHIFT under greedy decodingmPLUG-Owl2 · CHAIRCHAIR_S: 59.8 -> 50.2Improved
apply SHIFT under greedy decodingInstructBlip · CHAIRCHAIR_I: 23.2 -> 19.9Improved
use SHIFT instead of OPERA with beam searchLLaVA-1.5 · CHAIRCHAIR_S: 45.4 -> 36.7Improved
use VCD contrastive decoding with nucleus samplingLLaVA-1.5 · CHAIRCHAIR_S: 48.4 -> 56.4Hurt
use SHIFT instead of ICD with nucleus samplingmPLUG-Owl2 · CHAIRCHAIR_S: 62.4 -> 47.2Improved
apply SHIFT under greedy decodingLLaVA-1.5 · POPE MSCOCO adversarialaccuracy: 68.40 -> 74.40Improved
apply SHIFT under greedy decodingInstructBlip · POPE MSCOCO randomaccuracy: 90.23 -> 90.40No clear effect
use OPERA with beam searchLLaVA-1.5 · POPE MSCOCO randomaccuracy: 89.63 -> 89.30No clear effect
apply SHIFT to nucleus samplingLLaVA-1.5 · POPE MSCOCO adversarialaccuracy: 67.25 -> 74.13Improved
apply SHIFT versus vanilla decodingLLaVA-1.5 · CHAIR average output lengthtokens: 101.1 -> 99.8No clear effect
use SHIFT instead of OPERALLaVA-1.5 · CHAIR inference timetime (ms): 16657 -> 2519Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy