AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

DEGAP: Dynamic Entropy-Guided Attention Perturbation for Contrastive Decoding in Large Vision-Language Models

Hyein Seo, Yuna Jeong, Mingyu Kang, Junhyeong Park, Yong Suk Choi

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Bias-aware multi-image robustness

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.

Proposes DEGAP, a dynamic entropy-guided attention perturbation method for contrastive decoding that reduces hallucinations in large vision-language models.

Problem

LVLMs hallucinate by relying on language priors without visual grounding; existing contrastive decoding methods build contrast branches to mitigate this.

Approach

Uses entropy-guided, dynamically determined attention perturbation to construct the contrast branch for contrastive decoding, rather than fixed perturbations or input distortions (inferred from title; abstract truncated).

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use entropy-guided dynamic masking layer instead of static layer 0 (ablation)DEGAP on LLaVA-1.5 · CHAIRCHAIRs: 25.6 -> 19.1Improved
use entropy-guided dynamic masking layer instead of static layer 5 (ablation)DEGAP on LLaVA-1.5 · CHAIRCHAIRs: 28.6 -> 19.1Improved
use entropy-guided dynamic masking layer instead of static layer 20 (best static) (ablation)DEGAP on LLaVA-1.5 · CHAIRCHAIRs: 21.0 -> 19.1Improved
use dynamic layer instead of static layer 5 (ablation)DEGAP on InstructBLIP · CHAIRCHAIRs: 40.6 -> 22.1Improved
use dynamic layer instead of static layer 20 (best static) (ablation)DEGAP on mPLUG-Owl2 · CHAIRCHAIRs: 22.2 -> 20.2Improved
increase visual token masking ratio from 20% to 100% (ablation)DEGAP on LLaVA-1.5 · CHAIRCHAIRs: 26.0 -> 19.1Improved
increase visual token masking ratio from 80% to 100% (ablation)DEGAP on LLaVA-1.5 · CHAIRCHAIRs: 22.4 -> 19.1Improved
apply DEGAP contrastive decoding over regular decodingLLaVA-1.5 · CHAIRCHAIRs: 27.2 -> 19.1Improved
apply DEGAP over regular decodingInstructBLIP · POPE MS-COCO Randomaccuracy: 82.60 -> 89.69Improved
apply DEGAP over regular decodingLLaVA-1.5 · MME total scoreMME score: 573.33 -> 631.11Improved
use DEGAP instead of ICDLLaVA-1.5 · POPE Adversarial/PopularaccuracyMixed
reuse hidden state and run only later layers for the contrast branch (DEGAP vs DeGF)DEGAP on LLaVA-1.5 · latencylatency (s): 11.86 -> 5.09Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Builds on: contrastive decoding

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