List your company
ResearchFormerly WizWand

Mathematical Reasoning on MATH level 5

84Success RateGradient Shift
36.6848.96561.2573.535Jan 31, 2026Feb 18, 2026Mar 8, 2026Mar 26, 2026Apr 13, 2026May 1, 2026May 19, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Gradient Shiftvariant=mult2026.018426-0.41—
Gradient Shiftvariant=hybrid2026.018019-0.08—
Weak-to-Strongalpha=2.0, k=402026.0174110.32—
Weak-to-Strongalpha=1.5, k=402026.0168150—
Weak-to-Strongalpha=1.0, k=402026.0166130.1—
LC-ERD (Dual-72B)#Params=144B, VRAM (Avg.)=~112GB, Time (Rel.)=~2.1x, Multi-agent or MARL framework=true2026.0561.4———
LC-ERD (72B)#Params=72B, VRAM (Avg.)=~56GB, Time (Rel.)=~1.8x, Multi-agent or MARL framework=true2026.0558.1———
PRM#Params=72B+, VRAM (Avg.)=~64GB, Time (Rel.)=~1.6x, Multi-agent or MARL framework=true2026.0551.9———
Gradient Shiftvariant=add2026.0151110.02—
MaAS#Params=72B+, VRAM (Avg.)=~60GB, Time (Rel.)=~2.1x, Multi-agent or MARL framework=true2026.0549.2———
InfiGUI-R1#Params=72B, VRAM (Avg.)=~63GB, Time (Rel.)=~1.3x, Multi-agent or MARL framework=true2026.0546.8———
AFlow#Params=72B+, VRAM (Avg.)=~64GB, Time (Rel.)=~3.2x, Multi-agent or MARL framework=true2026.0546.7———
Gen-Verifier#Params=72B, VRAM (Avg.)=~68GB, Time (Rel.)=~2.1x2026.0545.2———
LPO#Params=72B, VRAM (Avg.)=~52GB, Time (Rel.)=~1.5x2026.0544.5———
EndoRM#Params=72B, VRAM (Avg.)=~54GB, Time (Rel.)=~1.4x2026.0544.1———
STaR#Params=72B, VRAM (Avg.)=~58GB, Time (Rel.)=~2.5x, Multi-agent or MARL framework=true2026.0543.8———
SimPO#Params=72B, VRAM (Avg.)=~52GB, Time (Rel.)=~1.1x2026.0541.8———
Iterative-DPO#Params=72B, VRAM (Avg.)=~56GB, Time (Rel.)=~2.1x2026.0541.5———
DPO#Params=72B, VRAM (Avg.)=~52GB, Time (Rel.)=~1.8x2026.0540.2———
Qwen2.5-72B (Base)#Params=72B, VRAM (Avg.)=~52GB, Time (Rel.)=1.0x (SFT)2026.0538.5———
Benchmark LLM (BenchAcc)2026.01—22——
Gemma-3-1B# Total Params=1B, # Trained Tokens=2T2025.11———14.64
Helper LLM2026.01—10——
LFM2-1.2B# Total Params=1.2B, # Trained Tokens=11T2025.11———18.61
LFM2-350M# Total Params=0.35B, # Trained Tokens=11T2025.11———5.79
LFM2-700M# Total Params=0.70B, # Trained Tokens=11T2025.11———11.27
Llama-3.2-1B# Total Params=1.2B, # Trained Tokens=9T2025.11———14.2
Qwen3-0.6B# Total Params=0.6B, # Trained Tokens=36T2025.11———19.43
Qwen3-1.7B# Total Params=1.7B, # Trained Tokens=36T2025.11———36.99