SPOT-E: Test-Time Entropy Shaping with Visual Spotlights for Frozen VLMs

Benchmark Model Rank Results
chart-question-answering-on-chartqaSPOT-E (Qwen3-VL-8B)#151:1 Accuracy: 89.0
multimodal-reasoning-on-mmmuSPOT-E (Gemini-2.5-Flash)#25Accuracy: 71.5
multimodal-reasoning-on-mathvistaSPOT-E (InternVL3-8B)#51Accuracy: 74.4