List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 2025 (held-out) (Pass@1)

59.07Pass@1ExpRL-Outcome
20.101230.218140.33550.4519Jun 14, 2026Jun 15, 2026
Updated 4mo ago

Evaluation results

MethodLinks
ExpRL-OutcomePriming=Dense terminal reward to full rollouts, Downstream RL=Stage-II RL setup, Evaluation=Sampling 128 responses per problem2026.0659.07
ExpRL-ProcessPriming=Dense rewards to partial rollouts and prefixes, Downstream RL=Stage-II RL setup, Evaluation=Sampling 128 responses per problem2026.0658.08
GRPOPriming=Verifiable sparse reward RL, Downstream RL=Stage-II RL setup, Evaluation=Sampling 128 responses per problem2026.0655.99
Self-DistillationPriming=Distillation-based, Downstream RL=Stage-II RL setup, Evaluation=Sampling 128 responses per problem2026.0655.59
Qwen3-4B-InstructModel type=Original base model, Evaluation=Sampling 128 responses per problem2026.0646.46
HSDBase Model=Qwen3-32B-Base2026.0641.7
GRPO + OPSDBase Model=Qwen3-32B-Base2026.0638.6
RLSDBase Model=Qwen3-32B-Base2026.0638.2
SDPOBase Model=Qwen3-32B-Base2026.0637.8
OPSDBase Model=Qwen3-32B-Base2026.0637.1
DAPOBase Model=Qwen3-32B-Base2026.0636.7
GSPOBase Model=Qwen3-32B-Base2026.0636
Dr. GRPOBase Model=Qwen3-32B-Base2026.0635.4
GRPOBase Model=Qwen3-32B-Base2026.0634.5
HSDBase Model=Qwen3-8B-Base2026.0628.1
SFTPriming=Supervised fine-tuning, Downstream RL=Stage-II RL setup, Evaluation=Sampling 128 responses per problem2026.0626.62
GRPO + OPSDBase Model=Qwen3-8B-Base2026.0625.7
RLSDBase Model=Qwen3-8B-Base2026.0625.3
SDPOBase Model=Qwen3-8B-Base2026.0625
OPSDBase Model=Qwen3-8B-Base2026.0624.4
DAPOBase Model=Qwen3-8B-Base2026.0623.7
GSPOBase Model=Qwen3-8B-Base2026.0623.2
Dr. GRPOBase Model=Qwen3-8B-Base2026.0622.5
GRPOBase Model=Qwen3-8B-Base2026.0621.6