List your company
ResearchFormerly WizWand

Mathematical Reasoning on HMMT25 (Avg@16 accuracy (%))

33.2Avg@16 AccuracyTRD
6.88813.71920.5527.381May 21, 2026May 23, 2026May 26, 2026May 29, 2026Jun 1, 2026Jun 4, 2026Jun 7, 2026
Updated 4mo ago

Evaluation results

MethodLinks
TRDModel=Qwen3-4B-Instruct-2507, Traj.=yr2026.0633.2
BaseModel=Qwen3-4B-Instruct-2507, Traj.=–2026.0631.3
Reverse KLModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.0631.3
Reverse KL w/ Top-KModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.0630.6
Forward KLModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.0630.4
Forward KL w/ ClipModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.0629.8
DASDBackbone=Qwen3-4B2026.0525
DASDBackbone=Qwen3-8B2026.0525
TRDModel=Qwen3-1.7B (w/ thinking), Traj.=yr2026.0624.4
BaseModel=Qwen3-1.7B (w/ thinking), Traj.=–2026.0624.2
Reverse KL w/ Top-KModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.0623.8
Forward KLModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.0623.3
Forward KL w/ ClipModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.0622.9
Reverse KLModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.0622.9
GRPOBackbone=Qwen3-4B2026.0522.1
HEPOBackbone=Qwen3-4B2026.0521.2
GRPOBackbone=Qwen3-8B2026.0520.8
RLSDBackbone=Qwen3-8B2026.0520.8
RLSDBackbone=Qwen3-4B2026.0520.4
HEPOBackbone=Qwen3-8B2026.0520.4
DASDBackbone=Qwen3-1.7B2026.0517.7
SRPOBackbone=Qwen3-4B2026.0517.7
SRPOBackbone=Qwen3-8B2026.0517.5
HEPOBackbone=Qwen3-1.7B2026.0515.6
RLSDBackbone=Qwen3-1.7B2026.0515.2
Qwen3-4BBackbone=Qwen3-4B2026.0515.2
OPSDBackbone=Qwen3-8B2026.0515
SDPOBackbone=Qwen3-8B2026.0514.4
SDPOBackbone=Qwen3-4B2026.0513.3
Qwen3-8BBackbone=Qwen3-8B2026.0513.1
OPSDBackbone=Qwen3-4B2026.0512.9
GRPOBackbone=Qwen3-1.7B2026.0511.5
SRPOBackbone=Qwen3-1.7B2026.0510.2
OPSDBackbone=Qwen3-1.7B2026.058.5
SDPOBackbone=Qwen3-1.7B2026.058.1
Qwen3-1.7BBackbone=Qwen3-1.7B2026.057.9