AI Paper Atlas

Research map / Video / Video-language understanding

MACD: Model-Aware Contrastive Decoding via Counterfactual Data for Video-LLMs

Qixin Xiao, Kun Zhou

ICML 2026 · Video › Video-language understanding · approach: Video LLMs with fine-grained understanding

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
use model-aware optimized counterfactual masks in contrastive decoding instead of random perturbation (VCD)Video-LLM contrastive decoding · EventHallusionAccuracy: 0.72 -> 0.82 (Qwen3-VL-2B, VCD vs MACD)Improved
add gradient-based mask strength training (ablation)MACD · EventHallusion (Qwen2.5-VL-3B)Accuracy: 0.61 -> 0.71Improved
add frame-level mask (ablation)MACD · EventHallusion (Qwen2.5-VL-3B)Accuracy: 0.61 -> 0.71Improved
frame extraction (temporal subsampling) on counterfactual view without frame mask (ablation)MACD without frame mask · EventHallusion (Qwen2.5-VL-3B)Accuracy: 0.61 -> 0.64Improved
use masks rather than learned noise for the counterfactual (ablation)MACD · EventHallusion (Qwen2.5-VL-3B)Accuracy: 0.65 (global noise) / 0.67 (object noise) / 0.66 (frame noise) -> 0.71Improved
discretize mask strengths into three levels instead of continuous (ablation)MACD · appendix Table 9AccuracyImproved
swap YOLOv11 detector for Grounding DINO (ablation)MACD · appendix Table 16AccuracyNo clear effect
apply MACD instead of baseline decoding on object-hallucination (video-POPE) questionsQwen2.5-VL-3B · Video-POPE (EventHallusion subset)Accuracy: 0.72 -> 0.85Improved
apply MACD to reduce false-yes answers on absent-object questionsQwen2.5-VL-3B · Video-POPE (EventHallusion subset)false-yes rate: 40.0% -> 17.0%Improved
use MACD masks vs random masks of equal area (ablation)mask construction · Human evaluation, 50 EventHallusion videosmean score: 2.40 -> 3.10Improved
apply MACD to Qwen2-VL-7B on Video-MMEQwen2-VL-7B · Video-MMEAccuracy: 0.60 (VCD) vs 0.59 (MACD); baseline 0.51Hurt
vary alpha and beta decoding coefficients (ablation)MACD · EventHallusion (Qwen2.5-VL-7B)Accuracy: 0.55 -> 0.61 -> 0.58 across alpha 2.1 to 3.8No clear effect

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