List your company

Visual Question Answering on BLINK

66.1AccuracyCLAIMDIFF-RL actor-only
41.171247.643154.11560.5869Oct 18, 2025Nov 22, 2025Dec 28, 2025Feb 1, 2026Mar 9, 2026Apr 13, 2026May 19, 2026

Evaluation results

MethodLinks
CLAIMDIFF-RL actor-onlyReward=actor-only, Initialization=Qwen2026.0566.1
HOLISTIC-RLReference Setting=w/ ref, Initialization=SFT2026.0565.8
CLAIMDIFF-RL relativeReward=relative, Initialization=SFT2026.0565.28
Qwen3VL-32B-Instruct2026.0564.9
HOLISTIC-RLReference Setting=w/o ref, Initialization=SFT2026.0564.3
CLAIMDIFF-RL actor-onlyReward=actor-only, Initialization=SFT2026.0563.76
SFT2026.0561.6
GDRMModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Beam Search, Decoding Framework=GDRM2026.0448.39
GDRMModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Temp’ Only, Decoding Framework=GDRM2026.0448.29
RotationCategory=SSL4RL-3B2025.1048.18
GDRMModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Top-k, Decoding Framework=GDRM2026.0447.71
GDRMModel=Qwen3-VL-Thinking (4B), Decoding Strategy=η-Sampling, Decoding Framework=GDRM2026.0447.66
Beam SearchModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Beam Search, Decoding Framework=Baseline2026.0447.4
η-SamplingModel=Qwen3-VL-Thinking (4B), Decoding Strategy=η-Sampling, Decoding Framework=Baseline2026.0447.13
ε-SamplingModel=Qwen3-VL-Thinking (4B), Decoding Strategy=ε-Sampling, Decoding Framework=Baseline2026.0446.61
GDRMModel=Qwen3-VL-Thinking (4B), Decoding Strategy=ε-Sampling, Decoding Framework=GDRM2026.0446.45
PositionCategory=SSL4RL-3B2025.1046.39
Temp’ OnlyModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Temp’ Only, Decoding Framework=Baseline2026.0446.08
GDRMModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Top-p (nucleus), Decoding Framework=GDRM2026.0445.98
GreedyModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Greedy, Decoding Framework=Baseline2026.0445.71
GDRMModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Min-p, Decoding Framework=GDRM2026.0445.4
Min-pModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Min-p, Decoding Framework=Baseline2026.0445.29
JigsawCategory=SSL4RL-3B2025.1045.18
Top-p (nucleus)Model=Qwen3-VL-Thinking (4B), Decoding Strategy=Top-p (nucleus), Decoding Framework=Baseline2026.0445.13
ContrastiveCategory=SSL4RL-3B2025.1045.13
Top-kModel=Qwen3-VL-Thinking (4B), Decoding Strategy=Top-k, Decoding Framework=Baseline2026.0444.77
Qwen2.5-VL-3BCategory=Base2025.1042.13