List your company

Mathematical Reasoning on AMC (test)

100Accuracy (Pass@1)S-GRPO
22.31242.48162.6582.819Sep 27, 2025Nov 6, 2025Dec 16, 2025Jan 26, 2026Mar 7, 2026Apr 16, 2026May 27, 2026

Evaluation results

MethodLinks
S-GRPOBackbone=Qwen3-14B2026.04100—4,538—
EDSBackbone=Qwen3-8B2026.0298.756,893——
BaseBackbone=QWQ-32B2026.0297.57,542——
VanillaBackbone=Qwen3-14B2026.0497.5—6,562—
RL + Length PenaltyBackbone=Qwen3-14B2026.0497.5—5,059—
DTSRBackbone=Qwen3-14B2026.0497.5—3,975—
Full-suppressBackbone=Qwen3-8B2026.0296.256,564——
BaseBackbone=Qwen3-8B2026.0295.628,091——
EDSBackbone=QWQ-32B2026.02956,719——
Full-suppressBackbone=QWQ-32B2026.0293.756,526——
BaseBackbone=DeepSeek-7B2026.0291.255,401——
EDSBackbone=DeepSeek-7B2026.0290.635,145——
Full-suppressBackbone=DeepSeek-7B2026.02905,067——
Laser-DEBackbone=DeepSeek-Distill-Qwen-7B2025.0982.93,357——
Laser-DBackbone=DeepSeek-Distill-Qwen-7B2025.0982.83,505——
DAPOBackbone=DeepSeek-Distill-Qwen-7B2025.0981.74,903——
JETBackbone=DeepSeek-Distill-Qwen-7B2025.09814,301——
BaseBackbone=DeepSeek-Distill-Qwen-7B2025.0978.96,013——
SFTBackbone=DeepSeek-Distill-Qwen-7B2025.0978.65,836——
DPOBackbone=DeepSeek-Distill-Qwen-7B2025.0977.25,797——
AdaThinkBackbone=DeepSeek-Distill-Qwen-7B2025.0977.23,871——
LCR1Backbone=DeepSeek-Distill-Qwen-7B2025.0976.53,173——
GAC + Token-φVariant=Token-φ2026.0567.2———
DAPOBackbone=DeepSeek-Distill-Qwen-1.5B2025.0966.35,666——
JETBackbone=DeepSeek-Distill-Qwen-1.5B2025.0966.13,872——
Laser-DBackbone=DeepSeek-Distill-Qwen-1.5B2025.0965.84,355——
GAC w/o φVariant=w/o φ2026.0565.8———
Laser-DEBackbone=DeepSeek-Distill-Qwen-1.5B2025.0965.25,018——
HPTCategory=Recent hybrid SFT–RL2026.0563.4———
BaseBackbone=DeepSeek-Distill-Qwen-1.5B2025.0963.37,644——
LUFFYCategory=Recent hybrid SFT–RL2026.0563.1———
KL-ctrlCategory=Rule-based controllers2026.0562.8———
CHORDCategory=SFT–RL mixing baselines2026.0562.5———
GradNorm-ctrlCategory=Rule-based controllers2026.0562.1———
Nash-MTLCategory=Multi-objective solvers2026.0561.9———
SRFTCategory=Recent hybrid SFT–RL2026.0561.8———
CAGradCategory=Multi-objective solvers2026.0561.7———
DPOBackbone=DeepSeek-Distill-Qwen-1.5B2025.09595,482——
AdaThinkBackbone=DeepSeek-Distill-Qwen-1.5B2025.0958.73,983——
SFT-best + RLCategory=SFT–RL mixing baselines2026.0558.4———
DPOCategory=RL-free alignment2026.0557.3———
SFTBackbone=DeepSeek-Distill-Qwen-1.5B2025.0957.25,394——
IPOCategory=RL-free alignment2026.0556.8———
LCR1Backbone=DeepSeek-Distill-Qwen-1.5B2025.0956.43,897——
SFT-best2026.0555.9———
GRPO (pure RL)Category=SFT–RL mixing baselines2026.0552.1———
Qwen2.5-7B-Inst.2026.0543.8———
ClusterTrain Size=4%2026.0540.96———
SHIFTTrain Size=2%2026.0538.55———
CoT SimilarityTrain Size=2%2026.0537.35———
SHIFTTrain Size=4%2026.0536.14———
Qwen2.5-Math-1.5BTrain Size=100%2026.0533.73———
CoT SimilarityTrain Size=4%2026.0533.73———
A-PPLTrain Size=4%2026.0532.53———
SC-EntropyTrain Size=2%2026.0531.33———
A-PPLTrain Size=2%2026.0531.33———
SC-EntropyTrain Size=4%2026.0531.33———
Q-PPLTrain Size=4%2026.0531.33———
Random (avg)Train Size=4%2026.0531.08———
Q-PPLTrain Size=2%2026.0530.12———
Qwen2.5-Math-1.5BTrain Size=N/A2026.0527.71———
Random (avg)Train Size=2%2026.0525.78———
ClusterTrain Size=2%2026.0525.3———
CoreSetTrain Size=2%2026.0525.3———
CoreSetTrain Size=4%2026.0525.3———
Ent-Adv2026.05———0.188
Ent-Reg2026.05———0.155
FGRPO2026.05———0.192
GRPO2026.05———0.197
KL-Cov2026.05———0.176
Pass@K2026.05———0.19
UCPO2026.05———0.217