List your company
ResearchFormerly WizWand

Visual Grounding on Out-of-Distribution Large Table (test)

91.4AccuracyMulti-SFT -> GRPO
-1.36822.71646.870.884Jul 2, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Multi-SFT -> GRPOEvaluation Setting=RL, Model Backbone=Qwen2.5-VL-7B-Instruct2026.0791.4—
Single-SFT -> GRPOEvaluation Setting=RL, Model Backbone=Qwen2.5-VL-7B-Instruct2026.0789.6—
VRRL (Ours)Evaluation Setting=RL, Model Backbone=Qwen2.5-VL-7B-Instruct2026.0789.6—
VRRL (Ours)Evaluation Setting=RL, Model Backbone=Qwen2.5-VL-3B-Instruct2026.0788.6—
Multi-SFT -> GRPOEvaluation Setting=RL, Model Backbone=Qwen2.5-VL-3B-Instruct2026.0778.6—
Reflection TuningEvaluation Setting=SFT, Model Backbone=Qwen2.5-VL-7B-Instruct2026.0775.1—
Multi-SFTEvaluation Setting=SFT, Model Backbone=Qwen2.5-VL-7B-Instruct2026.0766.2—
Single-SFTEvaluation Setting=SFT, Model Backbone=Qwen2.5-VL-7B-Instruct2026.0762.8—
Single-SFT -> GRPOEvaluation Setting=RL, Model Backbone=Qwen2.5-VL-3B-Instruct2026.0753.3—
Reflection TuningEvaluation Setting=SFT, Model Backbone=Qwen2.5-VL-3B-Instruct2026.0752.5—
Multi-SFTEvaluation Setting=SFT, Model Backbone=Qwen2.5-VL-3B-Instruct2026.0750.4—
Single-SFTEvaluation Setting=SFT, Model Backbone=Qwen2.5-VL-3B-Instruct2026.0746.1—
VL-Rethinker-32BEvaluation Setting=Zero-shot, Backbone=32B2026.0718.7—
Qwen2.5-VL-7BSingleEvaluation Setting=Zero-shot, Backbone=7B, Strategy=Single-turn2026.079.7—
Qwen2.5-VL-7B*MultiEvaluation Setting=Zero-shot, Backbone=7B, Strategy=Multi-turn reflection2026.078.949.3
VL-Rethinker-7BEvaluation Setting=Zero-shot, Backbone=7B2026.077.9—
Qwen2.5-VL-3BSingleEvaluation Setting=Zero-shot, Backbone=3B, Strategy=Single-turn2026.074.8—
Qwen2.5-VL-3B*MultiEvaluation Setting=Zero-shot, Backbone=3B, Strategy=Multi-turn reflection2026.072.269.2