List your company
ResearchFormerly WizWand

Multimodal Hallucination Evaluation on HallusionBench

70.7Hallucination ScoreThinkLite-VL
32.06442.094552.12562.1555Sep 30, 2025Oct 30, 2025Nov 29, 2025Dec 30, 2025Jan 29, 2026Feb 28, 2026Mar 31, 2026
Updated 4mo ago

Evaluation results

MethodLinks
ThinkLite-VLParameters=7B2025.1270.7——
Revisual-R1Parameters=7B2025.1269.2——
PG-CoTParameters=7B2025.1268.7——
Qwen2.5-VLParameters=7B2025.1265——
VAPOModel Scale=7B2025.0957.4——
VLAAModel Scale=7B2025.0954.7——
R1-OVModel Scale=7B2025.0952.5——
Qwen-VLModel Scale=7B2025.0950——
V-R1Model Scale=7B2025.0949.5——
Xuanwu VL-2BModel Size=2B2026.0347.32——
RecursiveVLMTraining Dataset=Data2, Train Steps=2, Eval Step=22026.0247.22——
InternVL 3.5 2BModel Size=2B2026.0346.78——
Standard TransformerTraining Dataset=Data2, Train Steps=1, Eval Step=12026.0246.65——
RecursiveVLMTraining Dataset=Data2, Train Steps=2, Eval Step=12026.0244.9——
InternVL 3.0 2BModel Size=2B2026.0343.32——
RecursiveVLMTraining Dataset=Data1, Train Steps=2, Eval Step=22026.0238.69——
Standard TransformerTraining Dataset=Data1, Train Steps=1, Eval Step=12026.0237.56——
Vanilla recursion w/ RMSNormTraining Dataset=Data1, Train Steps=2, Eval Step=12026.0236.54——
RecursiveVLMTraining Dataset=Data1, Train Steps=2, Eval Step=12026.0236.41——
Vanilla recursion w/ RMSNormTraining Dataset=Data1, Train Steps=2, Eval Step=22026.0235.15——
Vanilla recursionTraining Dataset=Data1, Train Steps=2, Eval Step=22026.0234.61——
Vanilla recursionTraining Dataset=Data1, Train Steps=2, Eval Step=12026.0233.55——
DAMOModel=Qwen2.5-VL-7B, Decoding=DAMO2025.05—71.2946.82
DAMOModel=Qwen3-VL-8B, Decoding=DAMO2025.05—73.0850
DAMOModel=InternVL3-8B, Decoding=DAMO2025.05—51.9443.64
DCLAModel=Qwen2.5-VL-7B, Decoding=DCLA2025.05—72.1348.55
DCLAModel=Qwen3-VL-8B, Decoding=DCLA2025.05—72.3449.42
DCLAModel=InternVL3-8B, Decoding=DCLA2025.05—52.0343.35
GreedyModel=Qwen2.5-VL-7B, Decoding=Greedy2025.05—71.8247.69
GreedyModel=Qwen3-VL-8B, Decoding=Greedy2025.05—71.548.84
GreedyModel=InternVL3-8B, Decoding=Greedy2025.05—50.3241.91
VCDModel=Qwen2.5-VL-7B, Decoding=VCD2025.05—71.6148.55
VCDModel=Qwen3-VL-8B, Decoding=VCD2025.05—69.9348.55
VCDModel=InternVL3-8B, Decoding=VCD2025.05—47.538.73