General Reasoning on MMBench
89.2AccuracyGazeVLM (Ours)
Evaluation results
| Method | Links | |
|---|---|---|
| GazeVLM (Ours)Gaze Bias=Enabled, Base Model=Qwen3-VL-4B2026.05 | 89.2 | |
| GazeVLM (w/o gaze bias)Gaze Bias=None, Base Model=Qwen3-VL-4B2026.05 | 88.7 | |
| Qwen3-VL-4BTraining Paradigm=Vanilla Open-source, Parameters=4B2026.05 | 87.7 | |
| HEEDPipeline Stage=C4+, Post-training=SFT+DPO2026.05 | 84.6 | |
| TeacherMethod ID=C0, Distillation Strategy=Teacher, Training Stage=Before SFT+DPO2026.05 | 84 | |
| TeacherPipeline Stage=C0, Post-training=SFT+DPO2026.05 | 84 | |
| Qwen2.5-VL-7BTraining Paradigm=Vanilla Open-source, Parameters=7B2026.05 | 83.8 | |
| Qwen3.5-4B-InstTraining Paradigm=Vanilla Open-source, Parameters=4B2026.05 | 83.8 | |
| HEEDMethod ID=C4, Distillation Strategy=HEED, Training Stage=Before SFT+DPO2026.05 | 83.7 | |
| RSAPipeline Stage=C3+, Post-training=SFT+DPO2026.05 | 83.4 | |
| RSAMethod ID=C3, Distillation Strategy=RSA, Training Stage=Before SFT+DPO2026.05 | 83.3 | |
| HSAMethod ID=C2, Distillation Strategy=HSA, Training Stage=Before SFT+DPO2026.05 | 82.9 | |
| KDMethod ID=C1, Distillation Strategy=KD, Training Stage=Before SFT+DPO2026.05 | 82.1 | |
| HSAPipeline Stage=C2+, Post-training=SFT+DPO2026.05 | 82.1 | |
| KDPipeline Stage=C1+, Post-training=SFT+DPO2026.05 | 81.1 |