List your company

Vision-Language Multi-task Performance on MergeBench

32.6MMSI-BenchTask-specific FT
27.60828.90430.231.496Feb 5, 2026

Evaluation results

MethodLinks
Task-specific FTBackbone=Qwen2.5-VL-7B-Instruct2026.0232.670.5855.1766.818572.563.78
TIES + OrthoMerge-CBackbone=Qwen2.5-VL-7B-Instruct2026.0232.571.5958.6268.583.169.964.04
TIESBackbone=Qwen2.5-VL-7B-Instruct2026.0232.471.5757.9367.6182.669.163.54
TIES + OrthoMerge-GBackbone=Qwen2.5-VL-7B-Instruct2026.0232.471.7657.9367.6784.369.763.96
TSV-M + OrthoMerge-GBackbone=Qwen2.5-VL-7B-Instruct2026.0232.372.5355.8668.2983.769.163.63
TSV-MBackbone=Qwen2.5-VL-7B-Instruct2026.0231.471.9554.8668.283.26963.1
TSV-M + OrthoMerge-CBackbone=Qwen2.5-VL-7B-Instruct2026.0231.172.3154.868.7483.369.963.36
TA + OrthoMerge-GBackbone=Qwen2.5-VL-7B-Instruct2026.0229.971.2957.2468.2683.568.763.15
TA + OrthoMerge-CBackbone=Qwen2.5-VL-7B-Instruct2026.022971.2654.4867.9484.669.162.73
TABackbone=Qwen2.5-VL-7B-Instruct2026.0228.771.2552.4167.7683.768.462.04
Qwen2.5-VL-7B-InstructBackbone=Qwen2.5-VL-7B-Instruct2026.0227.869.9753.166.384.767.261.51