List your company
ResearchFormerly WizWand

Mathematical Reasoning on MATH500 (Pass@1)

96.4Pass@1No Defense
10.0832.4954.977.31May 16, 2025Jun 28, 2025Aug 11, 2025Sep 23, 2025Nov 6, 2025Dec 19, 2025Feb 1, 2026
Updated 5mo ago

Evaluation results

MethodLinks
No DefenseBackbone=R1-Qwen-14B2025.0896.4
No DefenseBackbone=R1-Qwen-32B2025.0895.5
SGPOBackbone=QwQ-32B, Judge Model=DeepSeek-V3-03242025.0595
BaselineBackbone=QwQ-32B2025.0594.6
GRPOBackbone=QwQ-32B2025.0594.6
ReasoningGuardBackbone=R1-Qwen-14B2025.0894.5
SAFEPATH-ZSBackbone=R1-Qwen-32B2025.0894.5
ReasoningGuardBackbone=R1-Qwen-32B2025.0893.6
ThinkingIBackbone=R1-Qwen-32B2025.0893.5
SafeKeyBackbone=R1-Qwen-14B2025.0892.7
ThinkingIBackbone=R1-Qwen-14B2025.0892.7
SGPOBackbone=DeepSeek-R1-Distill-Qwen-7B, Judge Model=Qwen3-235B-A22B2025.0592
SGPOBackbone=DeepSeek-R1-Distill-Qwen-7B, Judge Model=QwQ-32B2025.0592
ParaphraseBackbone=R1-Qwen-14B2025.0891.8
SGPOBackbone=DeepSeek-R1-Distill-Qwen-7B, Judge Model=DeepSeek-V3-03242025.0591
ParaphraseBackbone=R1-Qwen-32B2025.0890.9
SGPOBackbone=DeepSeek-R1-Distill-Qwen-7B, Judge Model=o4-mini-04162025.0590.8
SGPOBackbone=DeepSeek-R1-Distill-Qwen-7B, Judge Model=QwQ-32B, Stability Parameters=w/o {beta,gamma}2025.0590.2
GRPOBackbone=DeepSeek-R1-Distill-Qwen-7B2025.0589.4
SAFEPATH-ZSBackbone=R1-Qwen-14B2025.0889.1
MinPROModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=22026.0188.1
M2POModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=22026.0187.7
CISPOModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=22026.0187.4
BaselineBackbone=DeepSeek-R1-Distill-Qwen-7B2025.0587
MinPROModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=22026.0186.5
M2POModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=22026.0185.8
GSPOModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=22026.0185.4
CISPOModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=22026.0184.5
GRPOBackbone=DeepSeek-R1-Distill-Llama-8B2025.0584
SGPOBackbone=DeepSeek-R1-Distill-Llama-8B, Judge Model=Claude-3.72025.0583.6
BaselineBackbone=Qwen2.5-32B-Instruct2025.0583.6
GRPOBackbone=Qwen2.5-32B-Instruct2025.0583.2
SGPOBackbone=Qwen2.5-32B-Instruct, Judge Model=o4-mini-04162025.0583
SGPOBackbone=Qwen2.5-14B-Instruct, Judge Model=o4-mini-04162025.0581
SetPO+GRPOBase Model=Qwen2.5-Math-7B, Optimization Algorithm=GRPO2026.0280.8
BaselineBackbone=Qwen2.5-14B-Instruct2025.0580.4
GRPOBackbone=Qwen2.5-14B-Instruct2025.0580.2
GRPOModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=22026.0180
SCR (Ours)Backbone=Qwen2.5-7B-Instruct2026.0180
SetPO+DAPOBase Model=Qwen2.5-Math-7B, Optimization Algorithm=DAPO2026.0279.2
RealSafe-R1Backbone=R1-Qwen-32B2025.0879.1
SFT + GRPOBackbone=Qwen2.5-7B-Instruct2026.0178.8
GSPOModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=22026.0178.4
DAPOBase Model=Qwen2.5-Math-7B2026.0277.6
GRPOBackbone=Qwen2.5-7B-Instruct2026.0177.4
BaselineBackbone=DeepSeek-R1-Distill-Llama-8B2025.0577.4
SCR-Stage IBackbone=Qwen2.5-7B-Instruct2026.0177
R1-zero-DivBase Model=Qwen2.5-Math-7B2026.0276.9
RealSafe-R1Backbone=R1-Qwen-14B2025.0876.4
SetPO+GSPOBase Model=Qwen2.5-Math-7B, Optimization Algorithm=GSPO2026.0275.7
BaseBackbone=Qwen2.5-7B-Instruct2026.0175.6
SCR-SFTBackbone=Qwen2.5-7B-Instruct2026.0174.2
GRPOBase Model=Qwen2.5-Math-7B2026.0273.5
GSPOBase Model=Qwen2.5-Math-7B2026.0273
SCR (Ours)Backbone=Qwen2.5-3B-Instruct2026.0170.4
GRPOModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=22026.0170.2
GRPOBackbone=Qwen2.5-3B-Instruct2026.0170.2
SFT + GRPOBackbone=Qwen2.5-3B-Instruct2026.0170
Self-ReminderBackbone=R1-Qwen-14B2025.0869.1
BaseModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=22026.0168.8
Self-RefineBackbone=Qwen2.5-7B-Instruct2026.0168.4
BaseBackbone=Qwen2.5-3B-Instruct2026.0166.6
Self-RefineBackbone=Qwen2.5-3B-Instruct2026.0165.6
SCR-Stage IBackbone=Qwen2.5-3B-Instruct2026.0165.2
SCR-SFTBackbone=Qwen2.5-3B-Instruct2026.0163.8
SCR (Ours)Backbone=Llama3.1-8B-Instruct2026.0160.8
SFT + GRPOBackbone=Llama3.1-8B-Instruct2026.0157.6
SCR-Stage IBackbone=Llama3.1-8B-Instruct2026.0155.8
SmoothLLMBackbone=R1-Qwen-32B2025.0853.8
SCR-SFTBackbone=Llama3.1-8B-Instruct2026.0153.4
GRPOBackbone=Llama3.1-8B-Instruct2026.0152.2
Self-ReminderBackbone=R1-Qwen-32B2025.0851.8
SmoothLLMBackbone=R1-Qwen-14B2025.0849.1
Qwen2.5-Math-7BModel Type=Base Model2026.0248.1
BaseModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=22026.0144.5
BaseBackbone=Llama3.1-8B-Instruct2026.0142.8
Self-RefineBackbone=Llama3.1-8B-Instruct2026.0113.4