List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 2024 (held-out)

79.4AccuracyAntiSD
22.30437.12751.9566.773May 12, 2026May 17, 2026May 23, 2026May 28, 2026Jun 3, 2026Jun 8, 2026Jun 14, 2026
Updated 4mo ago

Evaluation results

MethodLinks
AntiSDBackbone=Qwen3-30B-A3B, Training Condition=+AntiSD, Speedup=2.9×, Peak-mean step=140, Sampling strategy=avg@322026.0579.4
AntiSDBackbone=Qwen3-8B, Training Condition=+AntiSD, Speedup=5.0×, Peak-mean step=180, Sampling strategy=avg@322026.0578.4
AntiSDBackbone=Olmo3-7B-TK, Training Condition=+AntiSD, Speedup=2.0×, Peak-mean step=40, Sampling strategy=avg@322026.0577.6
SDBackbone=Olmo3-7B-TK, Training Condition=+SD, Peak-mean step=40, Sampling strategy=avg@322026.0577.2
AntiSDBackbone=Qwen3-4B-IT-2507, Training Condition=+AntiSD, Speedup=10.0×, Peak-mean step=100, Sampling strategy=avg@322026.0576.6
GRPOBackbone=Olmo3-7B-TK, Training Condition=+GRPO, Speedup=1.0×, Peak-mean step=80, Sampling strategy=avg@322026.0576.5
Olmo3-7B-TKBackbone=Olmo3-7B-TK, Training Condition=Base, Sampling strategy=avg@322026.0574.6
GRPOBackbone=Qwen3-30B-A3B, Training Condition=+GRPO, Speedup=1.0×, Peak-mean step=150, Sampling strategy=avg@322026.0574.1
GRPOBackbone=Qwen3-8B, Training Condition=+GRPO, Speedup=1.0×, Peak-mean step=200, Sampling strategy=avg@322026.0573.5
GRPOBackbone=Qwen3-4B-IT-2507, Training Condition=+GRPO, Speedup=1.0×, Peak-mean step=200, Sampling strategy=avg@322026.0567.8
AntiSDBackbone=Olmo3-7B-IT, Training Condition=+AntiSD, Speedup=9.5×, Peak-mean step=200, Sampling strategy=avg@322026.0562.4
Qwen3-4B-IT-2507Backbone=Qwen3-4B-IT-2507, Training Condition=Base, Sampling strategy=avg@322026.0562.1
SDBackbone=Qwen3-4B-IT-2507, Training Condition=+SD, Peak-mean step=10, Sampling strategy=avg@322026.0559.8
GRPOBackbone=Olmo3-7B-IT, Training Condition=+GRPO, Speedup=1.0×, Peak-mean step=190, Sampling strategy=avg@322026.0557
SDBackbone=Olmo3-7B-IT, Training Condition=+SD, Peak-mean step=10, Sampling strategy=avg@322026.0554.5
Olmo3-7B-ITBackbone=Olmo3-7B-IT, Training Condition=Base, Sampling strategy=avg@322026.0553.3
HSDBase Model=Qwen3-32B-Base2026.0646
GRPO + OPSDBase Model=Qwen3-32B-Base2026.0643.2
RLSDBase Model=Qwen3-32B-Base2026.0642.8
SDPOBase Model=Qwen3-32B-Base2026.0642.3
OPSDBase Model=Qwen3-32B-Base2026.0641.5
DAPOBase Model=Qwen3-32B-Base2026.0641.1
SDBackbone=Qwen3-30B-A3B, Training Condition=+SD, Peak-mean step=60, Sampling strategy=avg@322026.0540.9
GSPOBase Model=Qwen3-32B-Base2026.0640.5
SDBackbone=Qwen3-8B, Training Condition=+SD, Peak-mean step=200, Sampling strategy=avg@322026.0540.1
Dr. GRPOBase Model=Qwen3-32B-Base2026.0639.8
GRPOBase Model=Qwen3-32B-Base2026.0638.7
HSDBase Model=Qwen3-8B-Base2026.0631.4
GRPO + OPSDBase Model=Qwen3-8B-Base2026.0628.6
RLSDBase Model=Qwen3-8B-Base2026.0628.2
Qwen3-30B-A3BBackbone=Qwen3-30B-A3B, Training Condition=Base, Sampling strategy=avg@322026.0528.1
SDPOBase Model=Qwen3-8B-Base2026.0627.9
OPSDBase Model=Qwen3-8B-Base2026.0627.3
DAPOBase Model=Qwen3-8B-Base2026.0626.8
GSPOBase Model=Qwen3-8B-Base2026.0626.3
Dr. GRPOBase Model=Qwen3-8B-Base2026.0625.6
Qwen3-8BBackbone=Qwen3-8B, Training Condition=Base, Sampling strategy=avg@322026.0525.5
GRPOBase Model=Qwen3-8B-Base2026.0624.5