List your company
ResearchFormerly WizWand

Mathematical Reasoning on AQUA (test)

72.44AccuracySoftCoT
20.4433.9447.4460.94Feb 17, 2025May 2, 2025Jul 15, 2025Sep 28, 2025Dec 11, 2025Feb 23, 2026May 9, 2026
Updated 5mo ago

Evaluation results

MethodLinks
SoftCoTBackbone=Qwen2.5-7B-Instruct, Training Objective=Language Modeling Objective, Evaluation Protocol=Projection Module Trained2025.0272.4475.06
Zero-Shot Assist-CoTBackbone=Qwen2.5-7B-Instruct, Training Objective=None, Evaluation Protocol=Zero-Shot2025.0264.9671.64
Zero-Shot CoT-UnkBackbone=Qwen2.5-7B-Instruct, Training Objective=None, Evaluation Protocol=Zero-Shot2025.0264.7270.6
Zero-Shot CoTBackbone=Qwen2.5-7B-Instruct, Training Objective=None, Evaluation Protocol=Zero-Shot2025.0264.5370.29
CoconutBackbone=Qwen2.5-7B-Instruct, Training Objective=Language Modeling Objective, Evaluation Protocol=Finetuned2025.0263.39—
LoRA Fine-TuningBackbone=Qwen2.5-7B-Instruct, Training Objective=Language Modeling Objective, Evaluation Protocol=Fine-Tuning2025.0262.6—
RKUSparsity=30%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0528.35—
AGFSparsity=30%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0527.95—
RKUSparsity=40%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0527.95—
Taylor-FOSparsity=30%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0526.77—
Wanda-StructSparsity=30%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0525.98—
AGFSparsity=50%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0525.98—
Taylor-FOSparsity=40%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0525.59—
AGFSparsity=40%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0525.2—
Wanda-StructSparsity=40%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0524.02—
RKUSparsity=50%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0524.02—
Taylor-FOSparsity=50%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0523.23—
Wanda-StructSparsity=50%, Backbone=Qwen-2.5-7B, Fine-tuning steps=150, PEFT Method=LoRA2026.0522.44—