List your company

Mathematical Reasoning on Minerva-Math (Accuracy %)

57.4AccuracyNemotron-49B (Teacher)
16.819227.354637.8948.4254Aug 5, 2025Sep 28, 2025Nov 21, 2025Jan 14, 2026Mar 9, 2026May 2, 2026Jun 26, 2026

Evaluation results

MethodLinks
Nemotron-49B (Teacher)Model Role=Teacher, Category=Off-the-shelf Models2026.0657.4
KL Annealing onlySEAD Components=KL Annealing only, Category=Proposed2026.0656.6
Token Zones + KL AnnealingSEAD Components=Token Zones + KL Annealing, Category=Proposed2026.0654.8
Token Zones onlySEAD Components=Token Zones only, Category=Proposed2026.0654
GRPOCategory=Baselines2026.0653.7
OPD (RKL)Category=Baselines2026.0653.3
Nemotron-8B (Student)Model Role=Student, Category=Off-the-shelf Models2026.0652.9
OPSDCategory=Baselines2026.0649.6
VRPOBase Model=Qwen3-8B2025.0831.99
λ-GRPOBase Model=Qwen3-8B2025.0831.62
KTAEBase Model=Qwen3-8B2025.0829.41
BaseBase Model=Qwen3-8B2025.0828.68
VRPOBase Model=Qwen2.5-7B-Cold Start2025.0819.85
KTAEBase Model=Qwen2.5-7B-Cold Start2025.0819.12
BaseBase Model=Qwen2.5-7B-Cold Start2025.0818.75
λ-GRPOBase Model=Qwen2.5-7B-Cold Start2025.0818.38