Science Question Answering Robustness on SQA-IMG
78.8Robustness Score (Noise)Latent Denoising
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Latent DenoisingArchitecture=Qwen-2.5-VL, Severity=5, Method Variant=SQA2026.04 | 78.8 | 80 | 83.6 | 84.2 | |
| BaselineArchitecture=Qwen-2.5-VL, Severity=5, Method Variant=SQA2026.04 | 77.3 | 77.4 | 81.5 | 81.7 | |
| Latent DenoisingBackbone=LLaVA+CLIP, Corruption Severity=52026.04 | 69 | 69.4 | 68.9 | 69.5 | |
| BaselineBackbone=LLaVA+CLIP, Corruption Severity=52026.04 | 68.2 | 67.8 | 67.6 | 68.1 |