Multimodal Reasoning on COMT IID (Creation, Deletion, Selection, Update, Avg.)
89Creation ScoreZero-shot
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Zero-shotBackbone=Qwen3-VL-8B, Paradigm=Direct Ans., Training Stage=Standard2025.12 | 89 | 28 | 10 | 21 | 37 | |
| Direct-FTBackbone=Qwen3-VL-8B, Paradigm=Direct Ans., Training Stage=Standard2025.12 | 89 | 67 | 49 | 53 | 64.5 | |
| ILVRBackbone=Qwen3-VL-8B, Paradigm=Interleaved, Training Stage=Stage 2: Latent Relaxation2025.12 | 87 | 73 | 60 | 62 | 70.5 | |
| ILVRBackbone=Qwen3-VL-8B, Paradigm=Interleaved, Training Stage=Stage 1: Latent Alignment2025.12 | 84 | 63 | 57 | 55 | 64.8 | |
| MirageBackbone=Qwen3-VL-8B, Paradigm=Single-step, Training Stage=Stage 2: Latent Relaxation2025.12 | 84 | 66 | 54 | 57 | 65.3 | |
| CoT-FTBackbone=Qwen3-VL-8B, Paradigm=Text CoT, Training Stage=Standard2025.12 | 83 | 62 | 49 | 44 | 59.8 | |
| MirageBackbone=Qwen3-VL-8B, Paradigm=Single-step, Training Stage=Stage 1: Latent Alignment2025.12 | 81 | 58 | 43 | 50 | 58 | |
| CoT-FTBackbone=Qwen2.5-VL-7B, Paradigm=Text CoT, Training Stage=Standard2025.12 | 80 | 52 | 45 | 46 | 55.8 | |
| ILVRBackbone=Qwen2.5-VL-7B, Paradigm=Interleaved, Training Stage=Stage 2: Latent Relaxation2025.12 | 71 | 68 | 53 | 51 | 60.8 | |
| ILVRBackbone=Qwen2.5-VL-7B, Paradigm=Interleaved, Training Stage=Stage 1: Latent Alignment2025.12 | 69 | 66 | 46 | 47 | 57 | |
| Zero-shotBackbone=Qwen2.5-VL-7B, Paradigm=Direct Ans., Training Stage=Standard2025.12 | 68 | 38 | 35 | 14 | 38.8 | |
| MirageBackbone=Qwen2.5-VL-7B, Paradigm=Single-step, Training Stage=Stage 2: Latent Relaxation2025.12 | 65 | 62 | 47 | 50 | 56 | |
| MirageBackbone=Qwen2.5-VL-7B, Paradigm=Single-step, Training Stage=Stage 1: Latent Alignment2025.12 | 53 | 54 | 45 | 42 | 48.5 | |
| Direct-FTBackbone=Qwen2.5-VL-7B, Paradigm=Direct Ans., Training Stage=Standard2025.12 | 52 | 60 | 51 | 49 | 53 |