Multimodal Mathematical Reasoning on MathVerse 1.0 (test)
73.01ScoreVanilla (Full Tokens)
Evaluation Results
| Method | Links | |
|---|---|---|
| Vanilla (Full Tokens)Model=Qwen3-VL-4B-Thinking, Model Category=Thinking-based MLLMs, Token Retention Ratio=100%2026.04 | 73.01 | |
| Vanilla (Full Tokens)Model=Qwen3-VL-8B, Model Category=Large-scale MLLMs, Token Retention Ratio=100%2026.04 | 62.81 | |
| DSTPModel=Qwen3-VL-4B-Thinking, Model Category=Thinking-based MLLMs, Token Retention Ratio=33.3%2026.04 | 58.49 | |
| DSTPModel=Qwen3-VL-8B, Model Category=Large-scale MLLMs, Token Retention Ratio=33.3%2026.04 | 51.16 | |
| Vanilla (Full Tokens)Model=InternVL3.5-14B, Model Category=Large-scale MLLMs, Token Retention Ratio=100%2026.04 | 47.89 | |
| Vanilla (Full Tokens)Model=InternVL3.5-8B-Thinking, Model Category=Thinking-based MLLMs, Token Retention Ratio=100%2026.04 | 47.46 | |
| DSTPModel=InternVL3.5-14B, Model Category=Large-scale MLLMs, Token Retention Ratio=33.3%2026.04 | 43.95 | |
| FastV (33.3%)Model=Qwen3-VL-4B-Thinking, Model Category=Thinking-based MLLMs, Token Retention Ratio=33.3%2026.04 | 40.28 | |
| VOLDImages in FT=false2025.10 | 37.9 | |
| FastV (33.3%)Model=Qwen3-VL-8B, Model Category=Large-scale MLLMs, Token Retention Ratio=33.3%2026.04 | 37.64 | |
| DSTPModel=InternVL3.5-8B-Thinking, Model Category=Thinking-based MLLMs, Token Retention Ratio=33.3%2026.04 | 37.25 | |
| VLAA-Thinker 3BImages in FT=true2025.10 | 36.4 | |
| XReasoner-3B (repl.)Images in FT=false2025.10 | 35.7 | |
| VLM-R1 3B-MathImages in FT=true2025.10 | 32.2 | |
| FastV (33.3%)Model=InternVL3.5-14B, Model Category=Large-scale MLLMs, Token Retention Ratio=33.3%2026.04 | 31.86 | |
| Qwen2.5-VL-3BImages in FT=-2025.10 | 31.2 | |
| FastV (33.3%)Model=InternVL3.5-8B-Thinking, Model Category=Thinking-based MLLMs, Token Retention Ratio=33.3%2026.04 | 18.48 |