List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 24 (accuracy, delta)

14AccuracyBerr. Latent
-0.563.22710.78Feb 20, 2026Mar 2, 2026Mar 12, 2026Mar 22, 2026Apr 1, 2026Apr 11, 2026Apr 22, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Berr. LatentCollaboration space=Latent-space, Description=Best latent-space DDLM→ARM result, Projector training=trained Linear-GELU-Linear projector on 35K domain-specific samples2026.0214—
Qwen2.5-Math-7B w/ CADFTBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=CADFT, Evaluation Protocol=Average@162026.0410.41—
Qwen2.5-Math-1.5B w/ CADFTBackbone=Qwen2.5-Math-1.5B, Fine-tuning Protocol=CADFT, Evaluation Protocol=Average@162026.048.76—
Qwen2.5-Math-7B w/ DFTBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=DFT, Evaluation Protocol=Average@162026.048.56—
Qwen2.5-Math-1.5B w/ DFTBackbone=Qwen2.5-Math-1.5B, Fine-tuning Protocol=DFT, Evaluation Protocol=Average@162026.046.87—
Qwen2.5-Math-7BBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=None, Evaluation Protocol=Average@162026.046.68—
Qwen2.5-Math-1.5BBackbone=Qwen2.5-Math-1.5B, Fine-tuning Protocol=None, Evaluation Protocol=Average@162026.044.16—
LLaDA + Sonnet PlanPlan conditioning=Sonnet Plan2026.023.33.3
Qwen2.5-Math-7B w/ SFTBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=SFT, Evaluation Protocol=Average@162026.042.48—
Qwen2.5-Math-1.5B w/ SFTBackbone=Qwen2.5-Math-1.5B, Fine-tuning Protocol=SFT, Evaluation Protocol=Average@162026.041.87—
DeepSeekMath-7B w/ CADFTBackbone=DeepSeekMath-7B, Fine-tuning Protocol=CADFT, Evaluation Protocol=Average@162026.041.65—
Berr. BLDescription=LLaDA-only baseline by Berrayana et al.2026.021.5—
Berr. TextCollaboration space=Text-space, Description=Best text-space collaboration result2026.021.5—
DeepSeekMath-7B w/ DFTBackbone=DeepSeekMath-7B, Fine-tuning Protocol=DFT, Evaluation Protocol=Average@162026.041.24—
LLaMA-3.2-3B w/ CADFTBackbone=LLaMA-3.2-3B, Fine-tuning Protocol=CADFT, Evaluation Protocol=Average@162026.041.05—
LLaMA-3.2-3B w/ DFTBackbone=LLaMA-3.2-3B, Fine-tuning Protocol=DFT, Evaluation Protocol=Average@162026.040.83—
LLaMA-3.1-8B w/ CADFTBackbone=LLaMA-3.1-8B, Fine-tuning Protocol=CADFT, Evaluation Protocol=Average@162026.040.65—
LLaMA-3.2-3BBackbone=LLaMA-3.2-3B, Fine-tuning Protocol=None, Evaluation Protocol=Average@162026.040.41—
LLaMA-3.1-8B w/ DFTBackbone=LLaMA-3.1-8B, Fine-tuning Protocol=DFT, Evaluation Protocol=Average@162026.040.41—
DeepSeekMath-7B w/ SFTBackbone=DeepSeekMath-7B, Fine-tuning Protocol=SFT, Evaluation Protocol=Average@162026.040.41—
LLaMA-3.1-8BBackbone=LLaMA-3.1-8B, Fine-tuning Protocol=None, Evaluation Protocol=Average@162026.040.21—
DeepSeekMath-7BBackbone=DeepSeekMath-7B, Fine-tuning Protocol=None, Evaluation Protocol=Average@162026.040.21—
LLaDADescription=LLaDA-only baseline2026.020—
LLaMA-3.2-3B w/ SFTBackbone=LLaMA-3.2-3B, Fine-tuning Protocol=SFT, Evaluation Protocol=Average@162026.040—
LLaMA-3.1-8B w/ SFTBackbone=LLaMA-3.1-8B, Fine-tuning Protocol=SFT, Evaluation Protocol=Average@162026.040—