List your company

Object Hallucination Probing on OKVQA POPE Popular

85AccuracyScalpel
59.7866.327572.87579.4225Mar 27, 2024Jul 19, 2024Nov 10, 2024Mar 4, 2025Jun 26, 2025Oct 18, 2025Feb 10, 2026

Evaluation results

MethodLinks
ScalpelBase Model=LLaVA-1.5-7B2026.0285———————————85.18
ICTBase Model=LLaVA-1.5-7B2026.0283.4———————————84.45
OPERABase Model=LLaVA-1.5-7B2026.0283.22———————————84.67
VCDBase Model=LLaVA-1.5-7B2026.0281.85———————————82.82
Instruction Contrastive Decoding (ICD)Backbone=InstructBLIP2024.0381.64—————————78.588.7783.32
VanillaBase Model=LLaVA-1.5-7B2026.0279.9———————————79.59
Visual Contrastive Decoding (VCD)Backbone=InstructBLIP2024.0379.78—————————7687.0581.15
Standard DecodingBackbone=InstructBLIP2024.0376.19—————————72.1685.2878.17
Instruction Contrastive Decoding (ICD)Backbone=miniGPT42024.0362.81—————————61.6275.7867.97
Visual Contrastive Decoding (VCD)Backbone=miniGPT42024.0362.22—————————62.2368.5565.24
Standard DecodingBackbone=miniGPT42024.0360.75—————————60.6768.8464.5
BaselineBackbone=LLaVA-1.52025.06—86.9774.563.2769.477.0367.73——70.39———
BaselineBackbone=Qwen-VL-Chat2025.06—88.785.667.975.43——58.8775.2772.61———
CLAIMBackbone=LLaVA-1.52025.06—8583.481.8784.2777.97———82.5———
CLAIMBackbone=Qwen-VL-Chat2025.06—88.185.783.968.5——86.03—82.45———
VCDBackbone=LLaVA-1.52025.06——77.5767.168.477.0368.2——71.66———
VCDBackbone=Qwen-VL-Chat2025.06——84.1771.5773.3——52.475.6371.41———