List your company

Mathematical Reasoning on Omni-Math (Accuracy)

78.5AccuracyConfident Decoding
-3.1418.05539.2560.445May 10, 2026May 16, 2026May 23, 2026May 30, 2026Jun 6, 2026Jun 13, 2026Jun 20, 2026

Evaluation results

MethodLinks
Confident DecodingModel=Qwen3.5-122B-A10B, Decoding Method=Confident2026.0678.5
Last Layer DecodingModel=Qwen3.5-122B-A10B, Decoding Method=Last Layer2026.0678.3
Last Layer DecodingModel=Qwen3.5-27B, Decoding Method=Last Layer2026.0676.1
Confident DecodingModel=Qwen3.5-27B, Decoding Method=Confident2026.0676.1
Confident DecodingModel=Qwen3.5-35B-A3B, Decoding Method=Confident2026.0672.7
Last Layer DecodingModel=Qwen3.5-35B-A3B, Decoding Method=Last Layer2026.0672.6
Confident DecodingModel=gpt-oss-120B, Decoding Method=Confident2026.0671
Last Layer DecodingModel=gpt-oss-120B, Decoding Method=Last Layer2026.0670.2
Confident DecodingModel=Gemma-4-31B, Decoding Method=Confident2026.0669
Last Layer DecodingModel=Gemma-4-31B, Decoding Method=Last Layer2026.0668.3
Confident DecodingModel=gpt-oss-20b, Decoding Method=Confident2026.0661.6
Last Layer DecodingModel=gpt-oss-20b, Decoding Method=Last Layer2026.0660.9
Last Layer DecodingModel=Qwen3.5-9B-Base, Decoding Method=Last Layer2026.0654.7
Confident DecodingModel=Qwen3.5-9B-Base, Decoding Method=Confident2026.0653
Last Layer DecodingModel=Qwen3.5-9B, Decoding Method=Last Layer2026.0649.1
Confident DecodingModel=Qwen3.5-9B, Decoding Method=Confident2026.0647.1
Qwen2.5-7B w/ CoT Instruct (teacher)mode=CoT Instruct, role=teacher2026.0536.5
GKD (online distillation)distillation_type=online2026.0532
SFT w/ DISCORDdistillation_type=offline, correction=DISCORD2026.0529.5
Symmetric KL w/ DISCORDdistillation_type=offline, correction=DISCORD2026.0529
Qwen2.5-3B w/ CoT Instructmode=CoT Instruct2026.0528.5
Symmetric KLdistillation_type=offline2026.0528.5
ToDidistillation_type=offline2026.0528
Qwen2.5-7Brole=reference2026.0527
SFTdistillation_type=offline2026.0526.5
KLdistillation_type=offline2026.0526
T-MLPBackbone=LLaMA-3.1-8B-Instruct2026.0518.4
T-AllBackbone=LLaMA-3.1-8B-Instruct2026.0518.1
Qwen2.5-3B (base)mode=base2026.0517.5
BaseBackbone=LLaMA-3.1-8B-Instruct2026.0517.3
Vanilla SFTBackbone=LLaMA-3.1-8B-Instruct2026.0517.1
T-AllBackbone=LLaMA-3.2-3B-Instruct2026.0513.6
T-MLPBackbone=LLaMA-3.2-3B-Instruct2026.0511.9
BaseBackbone=LLaMA-3.2-3B-Instruct2026.0511.6
T-AllBackbone=Mistral-7B-Instruct2026.059.2
T-MLPBackbone=Mistral-7B-Instruct2026.059.1
BaseBackbone=Mistral-7B-Instruct2026.057.3
Vanilla SFTBackbone=Mistral-7B-Instruct2026.056.3
Vanilla SFTBackbone=LLaMA-3.2-3B-Instruct2026.055.2
Last Layer DecodingModel=Qwen3.5-0.8B, Decoding Method=Last Layer2026.060
Confident DecodingModel=Qwen3.5-0.8B, Decoding Method=Confident2026.060