List your company

Medical Visual Question Answering on Pooled Medical VQA Open-ended

79.3AUROC (Sampling)HAC-Gate
63.59667.67371.7575.827Apr 2, 2026

Evaluation results

MethodLinks
HAC-GateModel=InternVL3-8B, Calibration Method=HAC-Gate2026.0479.373.6—
HAC-PlattModel=InternVL3-8B, Calibration Method=HAC-Platt2026.0478.974.2—
UncalibratedModel=InternVL3-8B, Calibration Method=Uncalibrated2026.0476.461.4—
Platt ScalingModel=InternVL3-8B, Calibration Method=Platt Scaling2026.0476.461.4—
Isotonic Regr.Model=InternVL3-8B, Calibration Method=Isotonic Regr.2026.0476.362—
HAC-PlattModel=Qwen3-VL-8B, Calibration Method=HAC-Platt2026.047577.1—
HAC-GateModel=Qwen3-VL-8B, Calibration Method=HAC-Gate2026.047574.3—
HAC-PlattModel=LLaVA-NeXT-7B, Calibration Method=HAC-Platt2026.0472.570.3—
HAC-GateModel=LLaVA-NeXT-7B, Calibration Method=HAC-Gate2026.0472.170.2—
UncalibratedModel=LLaVA-NeXT-7B, Calibration Method=Uncalibrated2026.0470.660.3—
Platt ScalingModel=LLaVA-NeXT-7B, Calibration Method=Platt Scaling2026.0470.660.3—
Isotonic Regr.Model=LLaVA-NeXT-7B, Calibration Method=Isotonic Regr.2026.0470.359.9—
UncalibratedModel=Qwen3-VL-8B, Calibration Method=Uncalibrated2026.0464.371—
Platt ScalingModel=Qwen3-VL-8B, Calibration Method=Platt Scaling2026.0464.371—
Isotonic Regr.Model=Qwen3-VL-8B, Calibration Method=Isotonic Regr.2026.0464.271—
HAC-GateBackbone=Qwen3-VL-2B, Confidence Source=Sampling2026.04——72.9
HAC-GateBackbone=Qwen3-VL-2B, Confidence Source=Verbalized2026.04——70.4
HAC-GateBackbone=Qwen3-VL-8B, Confidence Source=Sampling2026.04——75
HAC-GateBackbone=Qwen3-VL-8B, Confidence Source=Verbalized2026.04——74.3
HAC-GateBackbone=InternVL3-38B, Confidence Source=Sampling2026.04——75.7
HAC-GateBackbone=InternVL3-38B, Confidence Source=Verbalized2026.04——69.2
HAC-PlattBackbone=Qwen3-VL-2B, Confidence Source=Sampling2026.04——72.9
HAC-PlattBackbone=Qwen3-VL-2B, Confidence Source=Verbalized2026.04——70.3
HAC-PlattBackbone=Qwen3-VL-8B, Confidence Source=Sampling2026.04——75
HAC-PlattBackbone=Qwen3-VL-8B, Confidence Source=Verbalized2026.04——77.1
HAC-PlattBackbone=InternVL3-38B, Confidence Source=Sampling2026.04——75.6
HAC-PlattBackbone=InternVL3-38B, Confidence Source=Verbalized2026.04——70.1