Multimodal Chain-of-Thought Reasoning on M3CoT
79.7AccuracyMMS-PRM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MMS-PRMSize=8B2026.06 | 79.7 | 73 | |
| InternVL2.5-MPO + SFTSize=8B2026.06 | 75.7 | 71 | |
| InternVL2.5-MPOSize=8B2026.06 | 67.5 | 69.4 | |
| LLaVA-NeXT-LLaMA3 + SFTSize=8B2026.06 | 67.4 | 60.8 | |
| InternVL-2.5Size=8B2026.06 | 62.4 | 68 | |
| Insight-VSize=7B2026.06 | 61.5 | 65.7 | |
| Qwen2-VLSize=7B2026.06 | 57.8 | 65.1 | |
| MiniCPM-V-2-6Size=8B2026.06 | 56 | 64.2 | |
| LLaVA-OneVisionSize=7B2026.06 | 52.3 | 64.6 | |
| Ours (PSP)Model=LaViDa, Generation length L / step T=256/1282026.04 | 50.5 | — | |
| Ours (PSP & VRG)Model=LaViDa, Generation length L / step T=256/1282026.04 | 50.3 | — | |
| MarginModel=LaViDa, Generation length L / step T=256/1282026.04 | 49.2 | — | |
| Low-confModel=LaViDa, Generation length L / step T=256/1282026.04 | 49 | — | |
| Ours (PSP & VRG)Model=LaViDa, Generation length L / step T=128/642026.04 | 48.6 | — | |
| Ours (PSP & VRG)Model=LaViDa, Generation length L / step T=64/322026.04 | 48.4 | — | |
| CCoTModel=LaViDa, Generation length L / step T=256/1282026.04 | 47.7 | — | |
| Ours (PSP)Model=LaViDa, Generation length L / step T=64/322026.04 | 47.6 | — | |
| Ours (PSP)Model=LaViDa, Generation length L / step T=128/642026.04 | 47.3 | — | |
| EntropyModel=LaViDa, Generation length L / step T=128/642026.04 | 46.9 | — | |
| EntropyModel=LaViDa, Generation length L / step T=256/1282026.04 | 46.8 | — | |
| DDCoTModel=LaViDa, Generation length L / step T=128/642026.04 | 46.7 | — | |
| DDCoTModel=LaViDa, Generation length L / step T=256/1282026.04 | 46.7 | — | |
| CCoTModel=LaViDa, Generation length L / step T=128/642026.04 | 46.5 | — | |
| MarginModel=LaViDa, Generation length L / step T=128/642026.04 | 46.5 | — | |
| EntropyModel=LaViDa, Generation length L / step T=64/322026.04 | 46.4 | — | |
| MarginModel=LaViDa, Generation length L / step T=64/322026.04 | 46.3 | — | |
| Low-confModel=LaViDa, Generation length L / step T=128/642026.04 | 46.2 | — | |
| Low-confModel=LaViDa, Generation length L / step T=64/322026.04 | 45.8 | — | |
| DDCoTModel=LaViDa, Generation length L / step T=64/322026.04 | 45.7 | — | |
| LLaVA-NeXT-LLaMA3Size=8B2026.06 | 45.6 | 52.1 | |
| CCoTModel=LaViDa, Generation length L / step T=64/322026.04 | 45.2 | — | |
| LLaVA-1.5Size=13B2026.06 | 39.5 | — | |
| Ours (PSP & VRG)Model=MMaDa, Generation length L / step T=128/642026.04 | 36.6 | — | |
| Ours (PSP & VRG)Model=MMaDa, Generation length L / step T=256/1282026.04 | 36.4 | — | |
| Ours (PSP & VRG)Model=MMaDa, Generation length L / step T=64/322026.04 | 36.3 | — | |
| Ours (PSP)Model=MMaDa, Generation length L / step T=256/1282026.04 | 35.8 | — | |
| Ours (PSP)Model=MMaDa, Generation length L / step T=64/322026.04 | 35.6 | — | |
| Ours (PSP)Model=MMaDa, Generation length L / step T=128/642026.04 | 35.2 | — | |
| CCoTModel=MMaDa, Generation length L / step T=64/322026.04 | 34.7 | — | |
| Low-confModel=MMaDa, Generation length L / step T=256/1282026.04 | 34.6 | — | |
| MarginModel=MMaDa, Generation length L / step T=64/322026.04 | 34.5 | — | |
| EntropyModel=MMaDa, Generation length L / step T=256/1282026.04 | 34.3 | — | |
| DDCoTModel=MMaDa, Generation length L / step T=64/322026.04 | 34.1 | — | |
| DDCoTModel=MMaDa, Generation length L / step T=256/1282026.04 | 34.1 | — | |
| EntropyModel=MMaDa, Generation length L / step T=64/322026.04 | 34.1 | — | |
| DDCoTModel=MMaDa, Generation length L / step T=128/642026.04 | 34 | — | |
| MarginModel=MMaDa, Generation length L / step T=128/642026.04 | 33.8 | — | |
| MarginModel=MMaDa, Generation length L / step T=256/1282026.04 | 33.8 | — | |
| Low-confModel=MMaDa, Generation length L / step T=128/642026.04 | 33.8 | — | |
| Low-confModel=MMaDa, Generation length L / step T=64/322026.04 | 33.7 | — | |
| EntropyModel=MMaDa, Generation length L / step T=128/642026.04 | 33.6 | — | |
| CCoTModel=MMaDa, Generation length L / step T=256/1282026.04 | 33.3 | — | |
| CCoTModel=MMaDa, Generation length L / step T=128/642026.04 | 31.8 | — |