List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 2025 (Acc@16, pass@16)

75Avg@16Dense Rollout Dense Training
-317.2537.557.75Mar 19, 2026Apr 1, 2026Apr 14, 2026Apr 28, 2026May 11, 2026May 24, 2026Jun 7, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Dense Rollout Dense TrainingModel Backbone=Qwen3-14B, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0675—82.69
Sparsity setup (0.86) RolloutModel Backbone=Qwen3-14B, Sparsity Level=0.86, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0675—82.8
Sparsity setup (0.86) RolloutModel Backbone=Qwen3-4B, Sparsity Level=0.86, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0674.16—83.21
Sparsity setup (0.92) RolloutModel Backbone=Qwen3-8B, Sparsity Level=0.92, Generation Length Cutoff=37K, Training Data Size=33K Polaris examples2026.0672.29—80.76
Dense Rollout Dense TrainingModel Backbone=Qwen3-4B, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0671.45—79.33
Sparsity setup (0.86) RolloutModel Backbone=Qwen3-8B, Sparsity Level=0.86, Generation Length Cutoff=37K, Training Data Size=33K Polaris examples2026.0671.45—80.55
Dense Rollout Dense TrainingModel Backbone=Qwen3-8B, Generation Length Cutoff=37K, Training Data Size=33K Polaris examples2026.0670.83—80
Sparsity setup (0.92) RolloutModel Backbone=Qwen3-4B, Sparsity Level=0.92, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0670—79.78
Sparsity setup (0.92) RolloutModel Backbone=Qwen3-1.7B, Sparsity Level=0.92, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0646.25—57.29
Sparse with LoRA Distillation (DistillSparse)Model Backbone=Qwen3-1.7B, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0643.54—55.62
Dense Rollout Dense TrainingModel Backbone=Qwen3-1.7B, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0643.22—54.4
Sparsity setup (0.86) RolloutModel Backbone=Qwen3-1.7B, Sparsity Level=0.86, Generation Length Cutoff=37K, Training Data Size=53K Polaris examples2026.0641.45—54.36
TeacherBackbone=JustRL-1.5B2026.0635.6——
OPRD2026.0634.6——
OPD top-162026.0634——
OPD top-1Variant=sampled-token2026.0633.5——
StudentBackbone=R1-distill-1.5B2026.0621.9——
PowerFlowBackbone=Qwen2.5-Math-7B2026.0314.4——
GRPOBackbone=Qwen2.5-Math-7B2026.0312.9——
EMPOBackbone=Qwen2.5-Math-7B2026.0312.3——
InstructBackbone=Qwen2.5-Math-7B2026.0312.3——
TTRLBackbone=Qwen2.5-Math-7B2026.0311.9——
PowerFlowBackbone=Qwen2.5-Math-1.5B2026.0310——
PowerSamplingBackbone=Qwen2.5-Math-7B2026.0310——
Qwen2.5-32B-InstructBackbone=Qwen2.5-32B-Instruct2026.039.4——
Low-tempBackbone=Qwen2.5-Math-7B2026.037.7——
GRPOBackbone=Qwen2.5-Math-1.5B2026.036.7——
One-shot EMBackbone=Qwen2.5-Math-7B2026.036.2——
InstructBackbone=Qwen2.5-Math-1.5B2026.036——
Format-onlyBackbone=Qwen2.5-Math-1.5B2026.035——
EMPOBackbone=Qwen2.5-Math-1.5B2026.034.6——
BaseBackbone=Qwen2.5-Math-7B2026.034.2——
Low-tempBackbone=Qwen2.5-Math-1.5B2026.034——
BaseBackbone=Qwen2.5-Math-1.5B2026.031.9——
PowerFlowBackbone=Qwen2.5-1.5B2026.031.5——
IntuitorBackbone=Qwen2.5-1.5B2026.030.8——
Low-tempBackbone=Llama-3.2-3B-Instruct2026.030.6——
Low-tempBackbone=Qwen2.5-1.5B2026.030.4——
GRPOBackbone=Qwen2.5-1.5B2026.030.4——
PowerFlowBackbone=Llama-3.2-3B-Instruct2026.030.4——
IntuitorBackbone=Llama-3.2-3B-Instruct2026.030.2——
BaseBackbone=Qwen2.5-1.5B2026.030——
InstructBackbone=Qwen2.5-1.5B2026.030——
BaseBackbone=Llama-3.2-3B-Instruct2026.030——
GRPOBackbone=Llama-3.2-3B-Instruct2026.030——
Det. Trunc.Base Model=Qwen2.5-Math-7B2026.02—2.77.1
Det. Trunc.Base Model=Qwen3-8B2026.02—15.937.4
GRPOBase Model=Qwen2.5-Math-7B2026.02—12.625.6
GRPOBase Model=Qwen3-8B2026.02—20.238.6
Qwen3-1.7BModel Scale=1.7B, Method=Base2026.05——30
Qwen3-1.7B + DASDModel Scale=1.7B, Method=DASD2026.05——63.3
Qwen3-1.7B + GRPOModel Scale=1.7B, Method=GRPO2026.05——56.7
Qwen3-1.7B + HEPOModel Scale=1.7B, Method=HEPO2026.05——60
Qwen3-1.7B + OPSDModel Scale=1.7B, Method=OPSD2026.05——23.3
Qwen3-1.7B + RLSDModel Scale=1.7B, Method=RLSD2026.05——46.7
Qwen3-1.7B + SDPOModel Scale=1.7B, Method=SDPO2026.05——26.7
Qwen3-1.7B + SRPOModel Scale=1.7B, Method=SRPO2026.05——43.3
Qwen3-4BModel Scale=4B, Method=Base2026.05——43.3
Qwen3-4B + DASDModel Scale=4B, Method=DASD2026.05——83.3
Qwen3-4B + GRPOModel Scale=4B, Method=GRPO2026.05——86.7
Qwen3-4B + HEPOModel Scale=4B, Method=HEPO2026.05——76.7
Qwen3-4B + OPSDModel Scale=4B, Method=OPSD2026.05——40
Qwen3-4B + RLSDModel Scale=4B, Method=RLSD2026.05——76.7
Qwen3-4B + SDPOModel Scale=4B, Method=SDPO2026.05——36.7
Qwen3-4B + SRPOModel Scale=4B, Method=SRPO2026.05——60
Qwen3-8BModel Scale=8B, Method=Base2026.05——40
Qwen3-8B + DASDModel Scale=8B, Method=DASD2026.05——83.3
Qwen3-8B + GRPOModel Scale=8B, Method=GRPO2026.05——80
Qwen3-8B + HEPOModel Scale=8B, Method=HEPO2026.05——76.7
Qwen3-8B + OPSDModel Scale=8B, Method=OPSD2026.05——40
Qwen3-8B + RLSDModel Scale=8B, Method=RLSD2026.05——76.7
Qwen3-8B + SDPOModel Scale=8B, Method=SDPO2026.05——43.3
Qwen3-8B + SRPOModel Scale=8B, Method=SRPO2026.05——60
RPCBase Model=Qwen2.5-Math-7B2026.02—12.228.2
RPCBase Model=Qwen3-8B2026.02—20.137
URSBase Model=Qwen2.5-Math-7B2026.02—11.622.9
URSBase Model=Qwen3-8B2026.02—20.738.4