Vision-Language Perception on VL Perception Benchmarks
85.5TextVQABase
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| BaseMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 85.5 | 86.5 | 82.9 | 76.7 | |
| BaseMerged Models=Qwen2.5-VL-7B-Instruct + OpenThinker3-7B2026.01 | 85.5 | 86.4 | 82.9 | 77 | |
| FRISMMerged Models=Qwen2.5-VL-7B-Instruct + OpenThinker3-7B2026.01 | 85.5 | 86.5 | 83 | 77 | |
| DAREMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 85.4 | 86.7 | 82.9 | 76.7 | |
| FRISMMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 85.4 | 86.6 | 82.9 | 76.8 | |
| DAREMerged Models=Qwen2.5-VL-7B-Instruct + OpenThinker3-7B2026.01 | 85.4 | 86.4 | 82.9 | 76.9 | |
| TAMerged Models=Qwen2.5-VL-7B-Instruct + OpenThinker3-7B, lambda=0.12026.01 | 85 | 86.5 | 82.8 | 76.8 | |
| IP-MergingMerged Models=Qwen2.5-VL-7B-Instruct + OpenThinker3-7B2026.01 | 84.8 | 86.1 | 82.6 | 77 | |
| TAMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B, lambda=0.12026.01 | 84.4 | 86.6 | 82.5 | 76.6 | |
| TAMerged Models=Qwen2.5-VL-7B-Instruct + OpenThinker3-7B, lambda=0.152026.01 | 84.1 | 86.4 | 82.3 | 76.4 | |
| Ties-MergingMerged Models=Qwen2.5-VL-7B-Instruct + OpenThinker3-7B2026.01 | 83.8 | 86.4 | 82.4 | 76.9 | |
| IP-MergingMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B, T=0.42026.01 | 83.2 | 87.2 | 82.3 | 76.7 | |
| TAMerged Models=Qwen2.5-VL-7B-Instruct + OpenThinker3-7B, lambda=0.22026.01 | 82.9 | 86.4 | 81.5 | 75.1 | |
| IP-MergingMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B, T=0.22026.01 | 82.6 | 86 | 81.7 | 76.7 | |
| IP-MergingMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B, T=0.32026.01 | 82.5 | 86.8 | 82 | 76.8 | |
| BaseMerged Models=InternVL3-8B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 82.2 | 90.3 | 83.2 | 77.2 | |
| FRISMMerged Models=InternVL3-8B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 82.2 | 90.3 | 83.3 | 77.4 | |
| IP-MergingMerged Models=InternVL3-8B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 82.1 | 90.2 | 83.2 | 77.2 | |
| DAREMerged Models=InternVL3-8B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 82 | 90.7 | 83.4 | 77.4 | |
| TAMerged Models=InternVL3-8B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 81.8 | 90 | 83.1 | 77.6 | |
| Ties-MergingMerged Models=InternVL3-8B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 80.3 | 88.7 | 82.1 | 77.2 | |
| Ties-MergingMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B2026.01 | 74.3 | 84.3 | 74.7 | 65.4 | |
| TAMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B, lambda=0.152026.01 | 68.6 | 72.6 | 67.7 | 61.9 | |
| TAMerged Models=ThinkLite-VL-7B + DeepSeek-R1-distilled-Qwen-7B, lambda=0.22026.01 | 62.8 | 62.6 | 65.4 | 70.7 | |
| FRISMMerging Setup=LLAVA-Next-8B with DeepThought-8B2026.01 | 61.6 | 86.3 | 74 | — | |
| IP-MergingMerging Setup=LLAVA-Next-8B with DeepThought-8B2026.01 | 61.3 | 86.2 | 73.8 | — | |
| BaseBase Model=LLAVA-Next-8B2026.01 | 61.2 | 87.1 | 74.2 | — | |
| TAMerging Setup=LLAVA-Next-8B with DeepThought-8B, Method Type=Task Arithmetic2026.01 | 61.2 | 86.7 | 74 | — |