Semantic Consistency Evaluation on Semantic-stress diagnostic dataset (evaluation subset)
0.0093Mean ∆lexGemma3-4B
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Gemma3-4BL=27, N=9,0682026.06 | 0.0093 | 0.0092 | 100 | -0.0016 | -0.0016 | 12.8 | |
| Qwen3VL-2BL=24, N=9,6612026.06 | 0.019 | 0.0188 | 100 | -0.0156 | -0.0158 | 5.3 | |
| InternVL3-2BL=23, N=9,7852026.06 | 0.0381 | 0.0377 | 100 | -0.0399 | -0.0403 | 3.2 | |
| Qwen3VL-4BL=24, N=9,2262026.06 | 0.0681 | 0.0678 | 100 | -0.0147 | -0.0149 | 8.6 | |
| Qwen3VL-8BL=32, N=9,6692026.06 | 0.1293 | 0.1284 | 99.9 | 0.0295 | 0.0288 | 85.3 |