List your company

Mathematical Reasoning on Snapask (Accuracy, Delta Metrics)

40.24AccuracyCKA
0.92811.13421.3431.546Mar 17, 2026

Evaluation results

MethodLinks
CKAPruning Budget=10%, Fine-tuning Protocol=without fine-tuning2026.0340.24-10.51-20.7
RandomPruning Budget=10%, Fine-tuning Protocol=without fine-tuning2026.0337.09-13.66-26.9
Ours-MathPruning Budget=10%, Fine-tuning Protocol=without fine-tuning2026.0335.83-14.92-29.4
Ours-NonMathPruning Budget=10%, Fine-tuning Protocol=without fine-tuning2026.0333.54-17.21-33.9
Ours-MixedPruning Budget=10%, Fine-tuning Protocol=without fine-tuning2026.0333.23-17.52-34.5
InterlacePruning Budget=10%, Fine-tuning Protocol=without fine-tuning2026.0328.82-21.93-43.2
Ours-MathPruning Budget=10%, Backbone=Qwen3-VL-2B2026.0328.15-12.32-30.4
Ours-MixedPruning Budget=10%, Backbone=Qwen3-VL-2B2026.0324.88-15.59-38.5
Ours-NonMathPruning Budget=10%, Backbone=Qwen3-VL-2B2026.0324.69-15.78-39
InterlacePruning Budget=10%, Backbone=Qwen3-VL-2B2026.0320.75-19.72-48.7
Ours-NonMathPruning Budget=25%, Fine-tuning Protocol=without fine-tuning2026.0318.19-32.56-64.2
RandomPruning Budget=10%, Backbone=Qwen3-VL-2B2026.0317.09-23.38-57.8
Ours-MixedPruning Budget=25%, Fine-tuning Protocol=without fine-tuning2026.0316.26-34.49-68
CKAPruning Budget=10%, Backbone=Qwen3-VL-2B2026.0315.98-24.49-60.5
Ours-MathPruning Budget=25%, Fine-tuning Protocol=without fine-tuning2026.0315.98-34.77-68.5
CKAPruning Budget=25%, Fine-tuning Protocol=without fine-tuning2026.0312.44-38.31-75.5
InterlacePruning Budget=25%, Fine-tuning Protocol=without fine-tuning2026.0312.09-38.66-76.2
RandomPruning Budget=25%, Fine-tuning Protocol=without fine-tuning2026.0311.3-39.45-77.7
Ours-NonMathPruning Budget=25%, Backbone=Qwen3-VL-2B2026.0311.06-29.41-72.7
InterlacePruning Budget=40%, Fine-tuning Protocol=without fine-tuning2026.0310.79-39.96-78.7
Ours-MathPruning Budget=40%, Fine-tuning Protocol=without fine-tuning2026.039.76-40.99-80.8
Ours-MixedPruning Budget=25%, Backbone=Qwen3-VL-2B2026.039.61-30.86-76.3
Ours-MathPruning Budget=25%, Backbone=Qwen3-VL-2B2026.039.21-31.26-77.2
Ours-NonMathPruning Budget=40%, Fine-tuning Protocol=without fine-tuning2026.039.06-41.69-82.1
Ours-MixedPruning Budget=40%, Fine-tuning Protocol=without fine-tuning2026.037.68-43.07-84.9
RandomPruning Budget=40%, Fine-tuning Protocol=without fine-tuning2026.035.51-45.24-89.1
InterlacePruning Budget=25%, Backbone=Qwen3-VL-2B2026.035.17-35.3-87.2
CKAPruning Budget=25%, Backbone=Qwen3-VL-2B2026.034.8-35.67-88.1
RandomPruning Budget=25%, Backbone=Qwen3-VL-2B2026.034.72-35.75-88.3
Ours-MixedPruning Budget=40%, Backbone=Qwen3-VL-2B2026.034.65-35.82-88.5
InterlacePruning Budget=40%, Backbone=Qwen3-VL-2B2026.034.57-35.9-88.7
Ours-NonMathPruning Budget=40%, Backbone=Qwen3-VL-2B2026.034.06-36.41-90
CKAPruning Budget=40%, Fine-tuning Protocol=without fine-tuning2026.033.7-47.05-92.7
Ours-MathPruning Budget=40%, Backbone=Qwen3-VL-2B2026.033.46-37.01-91.5
CKAPruning Budget=40%, Backbone=Qwen3-VL-2B2026.032.48-37.99-93.9
RandomPruning Budget=40%, Backbone=Qwen3-VL-2B2026.032.44-38.03-94