List your company
ResearchFormerly WizWand

Mathematical Reasoning on HMMT Nov 2025

94.4AccuracyDeepSeek-V3.2
-2.8483222.3988447.64672.89316Dec 2, 2025Jan 7, 2026Feb 12, 2026Mar 21, 2026Apr 26, 2026Jun 1, 2026Jul 8, 2026
Updated 3mo ago

Evaluation results

MethodLinks
DeepSeek-V3.2variant=Speciale, protocol=Pass@12025.1294.425——
GLM-4.72026.0793.5———
Gemini-3.0variant=Pro, protocol=Pass@12025.1293.315——
DeepSeek-V3.2variant=Thinking, protocol=Pass@12025.1290.218——
GPT-5variant=High, protocol=Pass@12025.1289.220——
Kimi-K2variant=Thinking, protocol=Pass@12025.1289.229——
GPT-5 High2026.0789.2———
Qwen3-30B-A3B SAOBackbone=Qwen3-30B-A3B, Training/Optimization method=SAO2026.0788.3———
Qwen3-30B-A3B - SAO (w/ DIS only)Backbone=Qwen3-30B-A3B, Training/Optimization method=SAO, DIS mechanism=w/ DIS only2026.0786.7———
Qwen3-30B-A3B - GRPO (+ DIS)Backbone=Qwen3-30B-A3B, Training/Optimization method=GRPO, DIS mechanism=+ DIS2026.0784———
Claude-Sonnet-4.52026.0781.7———
Qwen3-30B-A3B w/o pythonBackbone=Qwen3-30B-A3B, python usage=w/o python2026.0776.7———
Qwen3-30B-A3B GRPO (w/ python)Backbone=Qwen3-30B-A3B, Training/Optimization method=GRPO, python usage=w/ python2026.0776———
Qwen3-30B-A3B SFT (w/ python)Backbone=Qwen3-30B-A3B, Training/Optimization method=SFT, python usage=w/ python2026.0775.2———
ExOPDDistillation Setting=Single-Teacher Distillation2026.0239.3———
ExOPDDistillation Setting=Multi-Teacher Distillation2026.0239.2———
TeacherDistillation Setting=Baseline2026.0238.9———
OPDDistillation Setting=Multi-Teacher Distillation2026.0238.3———
OPDDistillation Setting=Single-Teacher Distillation2026.0237.9———
ExPODistillation Setting=Single-Teacher Distillation2026.0237———
ExPODistillation Setting=Multi-Teacher Distillation2026.0236.3———
SFTDistillation Setting=Multi-Teacher Distillation2026.0234.8———
DelTABackbone=Qwen3-14B-Base, Fine-tuning (FT)=false2026.0532.92———
FIPOBackbone=Qwen3-14B-Base, Fine-tuning (FT)=false2026.0532.29———
DAPOBackbone=Qwen3-14B-Base, Fine-tuning (FT)=true2026.0531.67———
DAPOBackbone=Qwen3-14B-Base, Fine-tuning (FT)=false2026.0530———
SAPOBackbone=Qwen3-14B-Base, Fine-tuning (FT)=false2026.0528.33———
DelTABackbone=Qwen3-8B-Base, Fine-tuning (FT)=false2026.0518.54———
Qwen3-30B-A3B w/ pythonBackbone=Qwen3-30B-A3B, python usage=w/ python2026.0717.3———
Qwen3-30B-A3B SFT (w/o python)Backbone=Qwen3-30B-A3B, Training/Optimization method=SFT, python usage=w/o python2026.0717.3———
SAPOBackbone=Qwen3-8B-Base, Fine-tuning (FT)=false2026.0516.04———
DAPOBackbone=Qwen3-8B-Base, Fine-tuning (FT)=true2026.0515.42———
FIPOBackbone=Qwen3-8B-Base, Fine-tuning (FT)=false2026.0512.92———
DAPOBackbone=Qwen3-8B-Base, Fine-tuning (FT)=false2026.0512.08———
StudentDistillation Setting=Baseline2026.028———
Gemini-3.0version=Pro2026.020.94515,000——
Kimi K2.52026.020.91124,000——
DeepSeek-V3.2mode=Thinking2026.020.90218,000——
Kimi K2mode=Thinking2026.020.89232,000——
BaseBackbone=Qwen3-4B-Instruct-2507, Number of rollouts=128, Average train reward (D_train)=5.532026.01—21.1741.61—
BaseModel Family=Qwen3-2507, Model Size=4B2026.05——26.6730
BaseModel Family=Qwen3-2507, Model Size=8B2026.05——3033.33
BaseModel Family=Qwen3.5, Model Size=4B2026.05——36.6753.33
BaseModel Family=Qwen3.5, Model Size=9B2026.05——4060
DeepSeek-R1-0528-Qwen3-8B# Params=8B, Scale Category=Small to Medium Scale (≤ 70B)2026.03——57.7—
DeepSeek-R1-Distill-Llama-70B# Params=70B, Scale Category=Small to Medium Scale (≤ 70B)2026.03——40.2—
DS-V3.2 DistillModel Family=Qwen3-2507, Model Size=4B2026.05——23.3330
DS-V3.2 DistillModel Family=Qwen3-2507, Model Size=8B2026.05——3043.33
DS-V3.2 DistillModel Family=Qwen3.5, Model Size=4B2026.05——43.3353.33
DS-V3.2 DistillModel Family=Qwen3.5, Model Size=9B2026.05——43.3360
Hint-guided RLBackbone=Qwen3-4B-Instruct-2507, Number of rollouts=128, Average train reward (D_train)=23.062026.01—28.5647.27—
InT + RLBackbone=Qwen3-4B-Instruct-2507, Number of rollouts=128, Average train reward (D_train)=28.832026.01—33.7249.77—
MegaScienceModel Family=Qwen3-2507, Model Size=4B2026.05——26.6733.33
MegaScienceModel Family=Qwen3-2507, Model Size=8B2026.05——46.6753.33
MegaScienceModel Family=Qwen3.5, Model Size=4B2026.05——43.3363.33
MegaScienceModel Family=Qwen3.5, Model Size=9B2026.05——6066.67
MINDLOOMModel Family=Qwen3-2507, Model Size=4B2026.05——33.3340
MINDLOOMModel Family=Qwen3-2507, Model Size=8B2026.05——46.6750
MINDLOOMModel Family=Qwen3.5, Model Size=4B2026.05——53.3360
MINDLOOMModel Family=Qwen3.5, Model Size=9B2026.05——7076.67
OpenThoughtModel Family=Qwen3-2507, Model Size=4B2026.05——3040
OpenThoughtModel Family=Qwen3-2507, Model Size=8B2026.05——43.3350
OpenThoughtModel Family=Qwen3.5, Model Size=4B2026.05——5066.67
OpenThoughtModel Family=Qwen3.5, Model Size=9B2026.05——7080
Qwen3-32B# Params=32B, Scale Category=Small to Medium Scale (≤ 70B)2026.03——50—
Qwen3-4B-Thinking-2507# Params=4B, Scale Category=Small to Medium Scale (≤ 70B)2026.03——57.3—
Qwen3-4B-Thinking-2507 + CHIMERA# Params=4B, Scale Category=Small to Medium Scale (≤ 70B)2026.03——67—
Qwen3-4B-Thinking-2507 + OpenScience# Params=4B, Scale Category=Small to Medium Scale (≤ 70B)2026.03——36.9—
RLBackbone=Qwen3-4B-Instruct-2507, Number of rollouts=128, Average train reward (D_train)=13.472026.01—28.2646.46—
SFT on ref. solutions + RLBackbone=Qwen3-4B-Instruct-2507, Number of rollouts=128, Average train reward (D_train)=19.072026.01—20.7627.45—
SFT on self-reflections + RLBackbone=Qwen3-4B-Instruct-2507, Number of rollouts=128, Average train reward (D_train)=23.192026.01—27.638.65—