List your company
ResearchFormerly WizWand

Captioning Hallucination on OBJHal

56CRsRLHF
5.35218.50131.6544.799Jul 29, 2025
Updated 6mo ago

Evaluation results

MethodLinks
RLHFBase Model=LLaVA-v1.5-7B, RL Technique=RLHF2025.075615.2
LLaVA-v1.5-7BModel Scale=7B2025.075415.8
LLaVA-v1.5-13BModel Scale=13B2025.075014.5
HALVABase Model=LLaVA-v1.5-7B, RL Technique=HALVA2025.0747.314.6
RLHFBase Model=LLaVA-v1.5-13B, RL Technique=RLHF2025.0746.711.6
HALVABase Model=LLaVA-v1.5-13B, RL Technique=HALVA2025.0742.712.1
Qwen-VL2.5-8BModel Scale=8B2025.0740.78.6
Intern-VL2.5-7BModel Scale=7B2025.07369.1
CHiP-DPOBase Model=LLaVA-v1.5-13B, RL Technique=CHiP-DPO2025.07306.2
GPT-4o2025.0729.36.7
DPOBase Model=LLaVA-v1.5-13B, RL Technique=DPO2025.07197.2
OPA-DPOBase Model=LLaVA-v1.5-13B, RL Technique=OPA-DPO2025.0718.35.1
TARS (Replace)Base Model=LLaVA-v1.5-13B, Perturbation Strategy=Replace2025.0714.93.4
TARS (Mask)Base Model=LLaVA-v1.5-13B, Perturbation Strategy=Mask2025.0714.62.8
DPOBase Model=LLaVA-v1.5-7B, RL Technique=DPO2025.07145
RLAIFBase Model=LLaVA-v1.5-7B, RL Technique=RLAIF2025.0713.74.2
TARS (Replace)Base Model=LLaVA-v1.5-7B, Perturbation Strategy=Replace2025.0713.43.3
OPA-DPOBase Model=LLaVA-v1.5-7B, RL Technique=OPA-DPO2025.0713.34.5
TARS (Mask)Base Model=LLaVA-v1.5-7B, Perturbation Strategy=Mask2025.07123.2
DeepSeek-VL2-27BModel Scale=27B2025.07107
CHiP-DPOBase Model=LLaVA-v1.5-7B, RL Technique=CHiP-DPO2025.077.34.3