Instance-level Recognition on ILR Benchmark (test)
92.9Object AccuracyIIR-VLM
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| IIR-VLMTraining Stage=Stage 1, Training Objective=trained for recognition, Expert Encoder=true2026.01 | 92.9 | 92.6 | 91.4 | 88.3 | 91.3 | |
| IIR-VLMTraining Stage=Stage 1 + 2, Training Objective=trained for recognition and caption, Expert Encoder=true2026.01 | 91.2 | 84.4 | 86.2 | 92.2 | 88.5 | |
| Gemini 2.5 Pro2026.01 | 89.8 | 75.4 | 72.8 | 79.6 | 79.4 | |
| IIR-VLMTraining Stage=Stage 1, Training Objective=trained for recognition, Expert Encoder=false2026.01 | 89.6 | 90.4 | 91.4 | 86.7 | 89.5 | |
| IIR-VLMTraining Stage=Stage 1 + 2, Training Objective=trained for recognition and caption, Expert Encoder=false2026.01 | 89.4 | 81.6 | 85.8 | 90.9 | 86.9 | |
| Qwen2.5 VL-7BParameters=7B2026.01 | 70.8 | 73.6 | 84.8 | 68.4 | 74.4 | |
| IDA-VLM2026.01 | 59.6 | 42 | 75 | 51.4 | 57 | |
| Qwen2.5 VL-3BParameters=3B2026.01 | 54 | 52.6 | 65.4 | 49.6 | 55.4 |