Intervention Target Prediction on CausalVLBench
47.4PrecisionGEMINI-2.0-FLASH (CM)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| GEMINI-2.0-FLASH (CM)Category=Commercial2026.06 | 47.4 | 56.4 | 55.7 | 66.1 | |
| BRIDGEVLM-7BModel Size=7B2026.06 | 36 | 54.4 | 43.6 | 83.7 | |
| PHI-4-MMI-7B (ANSWER)Model Size=7B, Setting=Answer-only fine-tuning2026.06 | 31.2 | 31 | 22.1 | 39.8 | |
| PHI-4-MMI-7B (CAUSAL-TRACE)Model Size=7B, Setting=Causal-trace supervision2026.06 | 29.2 | 33.2 | 26.2 | 44.3 | |
| QWEN2.5-VL-32BModel Size=32B2026.06 | 27.4 | 32.2 | 37.3 | 32 | |
| LLAVA-ONEVISION-7BModel Size=7B2026.06 | 27.1 | 31.9 | 32.7 | 35.9 | |
| PHI-4-MMI-7B (NO-FT, CAUSAL-PROMPT)Model Size=7B, Setting=No-FT, Causal-Prompt2026.06 | 26.4 | 25.4 | 23.7 | 26.3 | |
| DEEPSEEK-VL2-S-16BModel Size=16B2026.06 | 24.4 | 29 | 34.4 | 28.1 |