Distortion Type Classification on PANDABENCH Hard set
30AccuracyLinear Probe
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Linear ProbeType=Baseline2026.04 | 30 | 11 | 9 | 6 | |
| PANDA2026.04 | 27 | 22 | 18 | 19 | |
| DepictQA†Access=Open-source, Training=Trained on PANDASET2026.04 | 22 | 12 | 9 | 9 | |
| Q-InsightAccess=Open-source2026.04 | 21 | 5 | 6 | 4 | |
| GPT-5 NanoAccess=Closed-source2026.04 | 19 | 9 | 7 | 6 | |
| GPT-5 MiniAccess=Closed-source2026.04 | 17 | 10 | 10 | 9 | |
| Q-SiTAccess=Open-source2026.04 | 16 | 7 | 5 | 4 | |
| GPT-4oAccess=Closed-source2026.04 | 15 | 9 | 7 | 8 | |
| SeagullAccess=Open-source2026.04 | 13 | 10 | 13 | 8 | |
| Gemma-3 27BAccess=Open-source2026.04 | 13 | 9 | 9 | 7 | |
| Gemini 2.5 ProAccess=Closed-source2026.04 | 11 | 10 | 7 | 8 | |
| DepictQAAccess=Open-source2026.04 | 10 | 6 | 8 | 5 | |
| Attentive ProbeType=Baseline2026.04 | 10 | 8 | 8 | 7 | |
| RandomType=Baseline2026.04 | 7 | 7 | 7 | 6 |