List your company
ResearchFormerly WizWand

Mathematical Reasoning on Math500 (Accuracy (%))

93Accuracy (%)SPIRAL
11.25632.47853.774.922Jun 30, 2025Aug 25, 2025Oct 21, 2025Dec 17, 2025Feb 12, 2026Apr 10, 2026Jun 6, 2026
Updated 4mo ago

Evaluation results

MethodLinks
SPIRALBackbone=DeepSeek-Distill-Qwen-7B, Training Data=Multi-game2025.0693
SafeXArchitecture=Qwen3-30B-A3B2026.0592.2
SFTBackbone=DeepSeek-Distill-Qwen-7B, Training Data=Multi-game2025.0691.8
RTT-AONBackbone=Qwen3-4B-Instruct2026.0491.8
MESAArchitecture=Qwen3-30B-A3B2026.0591
DeepSeek-Distill-Qwen-7BBackbone=DeepSeek-Distill-Qwen-7B2025.0690.8
Base(instruct)Architecture=Qwen3-30B-A3B2026.0590.6
GRPOArchitecture=Qwen3-30B-A3B2026.0590.2
Qwen3-4B-InstructBackbone=Qwen3-4B-Instruct2026.0487.2
SPIRALBackbone=Qwen3-8B, Training Data=Multi-game2025.0686.6
ConSteer-RLBackbone=Qwen3-8B-Base2026.0686.6
GRPOBackbone=Qwen3-8B-Base2026.0686
ConSteer-RLBackbone=Qwen3-4B-Base2026.0685.8
Seq-ALPtemperature=1.0, sampling=average@32, multi-turn=true2026.0384.29
Stair-SFTArchitecture=Qwen3-30B-A3B2026.0583.8
GRPOBackbone=Qwen3-4B-Base2026.0683.6
Token-ALPtemperature=1.0, sampling=average@32, multi-turn=true2026.0383.48
Token-MIStemperature=1.0, sampling=average@32, multi-turn=true2026.0383.29
Stair-DPOArchitecture=Qwen3-30B-A3B2026.0583.2
SFTBackbone=Qwen3-8B, Training Data=Multi-game2025.0682.8
Seq-Bypasstemperature=1.0, sampling=average@32, multi-turn=true2026.0381.74
GRPOtemperature=1.0, sampling=average@32, multi-turn=true2026.0380.9
Seq-MIStemperature=1.0, sampling=average@32, multi-turn=true2026.0380.61
ConSteer-RLBackbone=Qwen2.5-Math-7B2026.0678.4
SPIRALBackbone=Qwen3-4B, Training Data=Multi-game2025.0678.2
GRPOBackbone=Qwen2.5-Math-7B2026.0677.6
Qwen3-8B-BaseBackbone=Qwen3-8B2025.0677
BaselineBackbone=Qwen3-8B-Base2026.0676.6
SPIRALBackbone=Qwen3-4B, Training Data=Kuhn Poker2025.0676.4
Qwen2.5-7B-InstructBackbone=Qwen2.5-7B-Instruct2026.0474.8
SFTBackbone=Qwen3-4B, Training Data=Multi-game2025.0674.2
SFTBackbone=Qwen3-4B, Training Data=Kuhn Poker2025.0674
RTT-AONBackbone=Qwen2.5-7B-Instruct2026.0473.8
Qwen3-4B-BaseBackbone=Qwen3-4B2025.0673.4
BaselineBackbone=Qwen3-4B-Base2026.0671.4
SFTArchitecture=Qwen3-30B-A3B2026.0569.2
SPIRALBackbone=Octothinker-8B, Training Data=Multi-game2025.0668.6
SFTBackbone=Octothinker-8B, Training Data=Multi-game2025.0666
Octothinker-8B-BaseBackbone=Octothinker-8B2025.0665.6
BaselineBackbone=Qwen2.5-Math-7B2026.0665
SFTBackbone=Llama-3.1-8B-Instruct, Training Data=Multi-game2025.0651.8
SPIRALBackbone=Llama-3.1-8B-Instruct, Training Data=Multi-game2025.0649.8
Llama-3.1-8B-InstructBackbone=Llama-3.1-8B-Instruct2025.0646.4
GRPO2026.0543.67
GRPO + WeMask (TF)Setting=Training-free (TF)2026.0543.47
GRPO + WeMask (TA)Setting=Training-aware (TA)2026.0543.33
Qwen3-4B2026.0543
RTT-AONBackbone=Llama3.2-3B-Instruct2026.0440
Llama3.2-3B-InstructBackbone=Llama3.2-3B-Instruct2026.0436.6
MESAArchitecture=DeepSeek-v2-Lite2026.0528.8
GRPOArchitecture=DeepSeek-v2-Lite2026.0528.2
Base(chat)Architecture=DeepSeek-v2-Lite2026.0524.8
SafeXArchitecture=DeepSeek-v2-Lite2026.0524.2
Stair-SFTArchitecture=DeepSeek-v2-Lite2026.0516.4
SFTArchitecture=DeepSeek-v2-Lite2026.0515
Stair-DPOArchitecture=DeepSeek-v2-Lite2026.0514.4