List your company
ResearchFormerly WizWand

Mathematical Reasoning on HMMT 25 (Accuracy)

45.42Accuracy (HMMT 25)p1
0.159211.909623.6635.4104Mar 17, 2026Mar 28, 2026Apr 9, 2026Apr 21, 2026May 2, 2026May 14, 2026May 26, 2026
Updated 5mo ago

Evaluation results

MethodLinks
p1Backbone=Qwen3-4B-Instruct-2507, S=[1, 23], M=16, Training Reward Imp.=0.212026.0445.42
General Teacher2026.0545
p1Backbone=Qwen3-4B-Instruct-2507, S=[17, 27], M=16, Training Reward Imp.=0.092026.0442.34
p1Backbone=Qwen3-4B-Instruct-2507, S=[4, 5, 17, 20], M=8, Training Reward Imp.=0.042026.0442.08
p1Backbone=Qwen3-4B-Instruct-2507, S=[23], M=32, Training Reward Imp.=0.072026.0441.25
baseBackbone=Qwen3-4B-Instruct-2507, S=/, M=/, Training Reward Imp.=/2026.0440.68
GEPABackbone=Qwen3-4B-Instruct-2507, S=15/15, M=/, Training Reward Imp.=/2026.0440.57
RLBackbone=Qwen3-4B-Instruct-2507, S=all 30, M=1, Training Reward Imp.=02026.0440.26
GEPABackbone=Qwen3-4B-Instruct-2507, S=28/[1,23], M=/, Training Reward Imp.=/2026.0440.21
GEPABackbone=Qwen3-4B-Instruct-2507, S=[1,23]/28, M=/, Training Reward Imp.=/2026.0440.1
Uni-OPDDistillation Scenario=Single-Teacher Distillation, Model Scale=4B2026.0539.8
Uni-OPDDistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.0539.6
ExOPDDistillation Scenario=Single-Teacher Distillation, Model Scale=4B2026.0539.3
ExOPDDistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.0539.2
p1Backbone=Qwen3-4B-Instruct-2507, S=[25], M=32, Training Reward Imp.=0.072026.0438.59
TeacherTraining Strategy=RL2026.0538.5
OPDDistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.0538.3
OPDDistillation Scenario=Single-Teacher Distillation, Model Scale=4B2026.0537.8
ExPODistillation Scenario=Single-Teacher Distillation, Model Scale=4B2026.0537
ExPODistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.0536.3
SFTDistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.0534.8
CaMOPD2026.0534.58
Vanilla MOPD2026.0533.96
SelecTKD2026.0533.54
Relaxed OPD2026.0532.92
Medical Teacher2026.0528.33
p1Backbone=Qwen3-1.7B, S=[5, 26], M=16, Training Reward Imp.=0.132026.0427.08
p1Backbone=Qwen3-1.7B, S=[10, 11], M=16, Training Reward Imp.=0.142026.0426.93
GEPABackbone=Qwen3-1.7B, S=28/[5,26], M=/, Training Reward Imp.=/2026.0425.52
baseBackbone=Qwen3-1.7B, S=/, M=/, Training Reward Imp.=/2026.0425.36
GEPABackbone=Qwen3-1.7B, S=15/15, M=/, Training Reward Imp.=/2026.0425.16
p1Backbone=Qwen3-1.7B, S=[5], M=32, Training Reward Imp.=0.252026.0425.05
GEPABackbone=Qwen3-1.7B, S=[5,26]/28, M=/, Training Reward Imp.=/2026.0425
RLBackbone=Qwen3-1.7B, S=all 30, M=1, Training Reward Imp.=02026.0424.9
p1Backbone=Qwen3-1.7B, S=[5, 14, 20, 26], M=8, Training Reward Imp.=0.052026.0424.79
p1Backbone=Qwen3-1.7B, S=[26], M=32, Training Reward Imp.=0.292026.0424.38
DyJR (M = 16)Backbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@256, M=162026.0313.6
DyJRBackbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@256, alpha_JS=0.05, M=82026.0312.7
DyJR (alpha_JS = 0.01)Backbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@256, alpha_JS=0.012026.0312.5
DyJR (alpha_JS = 0.2)Backbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@256, alpha_JS=0.22026.0310.9
RLEPBackbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@2562026.0310.8
DPH-RLBackbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@2562026.0310.5
DyJR (Forward KL)Backbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@256, Divergence Type=Forward KL2026.0310.5
Ex-GRPOBackbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@2562026.0310.4
DyJR (M = 32)Backbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@256, M=322026.0310.3
GRPOBackbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@2562026.039.8
DyJR (M = 64)Backbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@256, M=642026.039.7
DAPOBackbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@2562026.039.2
StudentModel Scale=4B2026.059.2
Base ModelBackbone=Qwen3-4B-Base, Evaluation Protocol (Mean@k)=Mean@2562026.031.9