AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Taming the Phantom: Token-Asymmetric Filtering for Hallucination Mitigation in Large Vision-Language Models

Shuyi Ouyang, Hongyi Wang, Gongfan Fang, Xinyin Ma, Lanfen Lin, Xinchao Wang

AAAI 2026 · 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 summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Token-level analysis of LVLM hallucination motivates a token-asymmetric filtering method to mitigate hallucinations (details unavailable beyond the abstract).

Problem

Hallucination in large vision-language models undermines reliability; prior work studied causes at the modality level, leaving token-level interaction patterns underexplored.

Approach

Conducts a token-level analysis of LVLM interactions beyond modality level and uses the identified asymmetry between tokens to filter them for hallucination mitigation.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply TAF (phantom isolation + anchor emphasis) training-freeLLaVA-1.5-7B · POPE AdversarialF1: 76.84 -> 86.21Improved
apply TAF vs best baseline AGLALLaVA-1.5-7B · CHAIR MSCOCOCHAIRs: 47.6 -> 42.5Improved
apply TAF to mPLUG-Owl2 captioningmPLUG-Owl2 · CHAIR MSCOCOCHAIRs: 53.7 -> 52.7Improved
apply TAF to MiniGPT-4MiniGPT-4-7B · CHAIR MSCOCORecall: 55.8 -> 56.3Mixed
increase anchor emphasizing coefficient alpha (ablation)TAF on LLaVA-1.5-7B · POPEF1Mixed
raise phantom isolation ratio from 0 to 100% (ablation)TAF on LLaVA-1.5-7B · POPEF1Improved

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