Mathematical Reasoning on GSM8K (Accuracy & Token Count)
98AccuracyMARS
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MARSBackbone=GPT-4o-mini2025.09 | 98 | 3,389 | — | — | — | — | — | — | — | — | — | — | |
| MADBackbone=GPT-4o-mini2025.09 | 97.67 | 7,440 | — | — | — | — | — | — | — | — | — | — | |
| MARSBackbone=Llama3-70b2025.09 | 97.67 | 3,313 | — | — | — | — | — | — | — | — | — | — | |
| Self-reflectionBackbone=Llama3-70b2025.09 | 97.33 | 777 | — | — | — | — | — | — | — | — | — | — | |
| Self-consistencyBackbone=Llama3-70b2025.09 | 97.33 | 1,410 | — | — | — | — | — | — | — | — | — | — | |
| MADBackbone=Llama3-70b2025.09 | 97.33 | 6,307 | — | — | — | — | — | — | — | — | — | — | |
| DMADBackbone=Llama3-70b2025.09 | 97.33 | 7,659 | — | — | — | — | — | — | — | — | — | — | |
| SCModel=GEMMA-3-27B, k=162026.01 | 97.04 | — | — | 166.93 | 0.58 | — | — | — | — | — | — | — | |
| ESCModel=GEMMA-3-27B, w=42026.01 | 97.04 | — | — | 47.73 | 2.03 | — | — | — | — | — | — | — | |
| ASCModel=GEMMA-3-27B2026.01 | 97.04 | — | — | 47.71 | 2.03 | — | — | — | — | — | — | — | |
| ReASC (online)Model=GEMMA-3-27B2026.01 | 97.04 | — | — | 35.32 | 2.75 | — | — | — | — | — | — | — | |
| CoTBackbone=Llama3-70b2025.09 | 97 | 335 | — | — | — | — | — | — | — | — | — | — | |
| Self-consistencyBackbone=GPT-4o-mini2025.09 | 97 | 1,548 | — | — | — | — | — | — | — | — | — | — | |
| ReASC (offline)Model=GEMMA-3-27B2026.01 | 96.89 | — | — | 29.36 | 3.3 | — | — | — | — | — | — | — | |
| DMADBackbone=GPT-4o-mini2025.09 | 96.33 | 8,831 | — | — | — | — | — | — | — | — | — | — | |
| CoTBackbone=GPT-4o-mini2025.09 | 96 | 384 | — | — | — | — | — | — | — | — | — | — | |
| PIRBackbone=QwQ-32B2026.02 | 95.8 | 738.24 | — | — | — | — | — | — | — | — | — | — | |
| pass@1Model=GEMMA-3-27B2026.01 | 95.6 | — | — | — | — | — | — | — | — | — | — | — | |
| PIRBackbone=DeepSeek-R1-32B2026.02 | 95.6 | 197.03 | — | — | — | — | — | — | — | — | — | — | |
| BaselineBackbone=QwQ-32B, Temperature=0.62025.12 | 95 | 1,450.68 | — | — | — | — | — | — | — | — | — | — | |
| Zero-shotModel=Gemini-1.5-Pro, Model Role=Teacher Model2026.02 | 94.5 | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-8BPruning strategy=None2025.08 | 94.46 | 3,053.67 | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-8Bstrategy=Full COT2025.08 | 94.46 | 3,053.67 | — | — | — | — | — | — | — | — | — | — | |
| DAAOLLM=LLM Pool2025.09 | 94.4 | — | — | — | — | — | — | — | 83.26 | — | — | — | |
| DEERBackbone=QwQ-32B, Temperature=0.62025.12 | 94.39 | 983.77 | — | — | — | -0.64 | -32.2 | — | — | — | — | — | |
| Qwen3-8BPruning strategy=80% lowest-entropy steps2025.08 | 94.39 | 2,557.47 | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-8Bstrategy=step-entropy pruning, pruning_ratio=80%2025.08 | 94.39 | 2,557.47 | — | — | — | — | — | — | — | — | — | — | |
| SCModel=QWEN-2.5-7B, k=162026.01 | 94.31 | — | — | 41.59 | 2.27 | — | — | — | — | — | — | — | |
| ESCModel=QWEN-2.5-7B, w=42026.01 | 94.24 | — | — | 13.87 | 6.8 | — | — | — | — | — | — | — | |
| ASCModel=QWEN-2.5-7B2026.01 | 94.24 | — | — | 13.4 | 7.04 | — | — | — | — | — | — | — | |
| ReASC (online)Model=QWEN-2.5-7B2026.01 | 94.24 | — | — | 12.4 | 7.6 | — | — | — | — | — | — | — | |
| DeepSeek-R1-32BBackbone=DeepSeek-R1-32B2026.02 | 94.2 | 263.04 | — | — | — | — | — | — | — | — | — | — | |
| ReASC (offline)Model=QWEN-2.5-7B2026.01 | 94.09 | — | — | 10.43 | 9.02 | — | — | — | — | — | — | — | |
| QwQ-32Bstrategy=Full COT2025.08 | 94.09 | 1,978.91 | — | — | — | — | — | — | — | — | — | — | |
| GRPOBackbone=Qwen3-14B, Sampling Strategy=mean@82026.03 | 94 | 736 | — | — | — | — | — | — | — | — | — | — | |
| CODABackbone=Qwen3-14B, Sampling Strategy=mean@82026.03 | 94 | 266 | — | — | — | — | — | — | — | — | — | — | |
| VLPBackbone=Qwen3-14B, Sampling Strategy=mean@82026.03 | 93.7 | 327 | — | — | — | — | — | — | — | — | — | — | |
| ASRRBackbone=Qwen3-14B, Sampling Strategy=mean@82026.03 | 93.6 | 265 | — | — | — | — | — | — | — | — | — | — | |
| LYNXBackbone=QwQ-32B, Conf=0.97, Temperature=0.62025.12 | 93.56 | 1,095.1 | — | — | — | -1.52 | -24.5 | — | — | — | — | — | |
| QwQ-32Bstrategy=step-entropy pruning, pruning_ratio=80%2025.08 | 93.56 | 1,629.03 | — | — | — | — | — | — | — | — | — | — | |
| Zero-shot-EIModel=Gemini-1.5-Pro, Model Role=Teacher Model2026.02 | 93.51 | — | — | — | — | — | — | — | — | — | — | — | |
| C3oTBackbone=Qwen2.5-7B2026.02 | 93.5 | 291.22 | — | — | — | — | — | — | — | — | — | — | |
| StandardpBackbone=Qwen2.5-7B2026.02 | 93.33 | 289.82 | — | — | — | — | — | — | — | — | — | — | |
| StandardpBackbone=DeepSeek-R1-7B2026.02 | 93.33 | 1,710.27 | — | — | — | — | — | — | — | — | — | — | |
| S3-CoTBackbone=Qwen2.5-7B2026.02 | 93.17 | 182.8 | — | — | — | — | — | — | — | — | — | — | |
| LASERDEBackbone=DeepSeek-R1-7B2026.02 | 93 | 974.57 | — | — | — | — | — | — | — | — | — | — | |
| AutoTHINKBackbone=DeepSeek-R1-7B2026.02 | 92.83 | 1,121.24 | — | — | — | — | — | — | — | — | — | — | |
| SPOT-stage2Backbone=DeepSeek-R1-Distill-Qwen-7B2026.03 | 92.72 | 630 | — | — | — | — | — | — | — | — | — | — | |
| PIRBackbone=DeepSeek-R1-7B2026.02 | 92.6 | 362.5 | — | — | — | — | — | — | — | — | — | — | |
| S3-CoTscBackbone=Qwen2.5-7B2026.02 | 92.5 | 184.1 | — | — | — | — | — | — | — | — | — | — | |
| C3oTBackbone=DeepSeek-R1-7B2026.02 | 92.5 | 1,475.27 | — | — | — | — | — | — | — | — | — | — | |
| GRPOBackbone=Qwen3-8B, Sampling Strategy=mean@82026.03 | 92.5 | 825 | — | — | — | — | — | — | — | — | — | — | |
| VLPBackbone=Qwen3-8B, Sampling Strategy=mean@82026.03 | 92.5 | 391 | — | — | — | — | — | — | — | — | — | — | |
| ASRRBackbone=Qwen3-8B, Sampling Strategy=mean@82026.03 | 92.5 | 326 | — | — | — | — | — | — | — | — | — | — | |
| CODABackbone=Qwen3-8B, Sampling Strategy=mean@82026.03 | 92.5 | 311 | — | — | — | — | — | — | — | — | — | — | |
| LYNXBackbone=QwQ-32B, Conf=0.95, Temperature=0.62025.12 | 92.49 | 974.08 | — | — | — | -2.64 | -32.9 | — | — | — | — | — | |
| QwQ-32BBackbone=QwQ-32B2026.02 | 92.4 | 1,672.23 | — | — | — | — | — | — | — | — | — | — | |
| Self-reflectionBackbone=GPT-4o-mini2025.09 | 92.33 | 894 | — | — | — | — | — | — | — | — | — | — | |
| MaASLLM=gpt-4o-mini2025.09 | 92.3 | — | — | — | — | — | — | — | 80.43 | — | — | — | |
| ZIP-RC samplingModel=Qwen3-1.7B2025.12 | 92.2 | — | — | — | — | — | — | 1.43 | — | — | — | — | |
| PIRBackbone=DeepSeek-R1-14B2026.02 | 92.2 | 360.17 | — | — | — | — | — | — | — | — | — | — | |
| SPOT-stage1Backbone=DeepSeek-R1-Distill-Qwen-7B2026.03 | 92.19 | 576 | — | — | — | — | — | — | — | — | — | — | |
| SCModel=GEMMA-3-4B, k=162026.01 | 92.12 | — | — | 32.67 | 2.82 | — | — | — | — | — | — | — | |
| ASCModel=GEMMA-3-4B2026.01 | 92.12 | — | — | 12.26 | 7.52 | — | — | — | — | — | — | — | |
| PIAST (E)Time [m]=1598.34±234.542025.12 | 92.12 | — | — | — | — | — | — | — | — | — | — | — | |
| ESCModel=GEMMA-3-4B, w=42026.01 | 92.04 | — | — | 12.93 | 7.12 | — | — | — | — | — | — | — | |
| ReASC (offline)Model=GEMMA-3-4B2026.01 | 92.04 | — | — | 9.45 | 9.74 | — | — | — | — | — | — | — | |
| ReASC (online)Model=GEMMA-3-4B2026.01 | 92.04 | — | — | 10.25 | 8.98 | — | — | — | — | — | — | — | |
| MaASLLM=gemini-1.5-flash2025.09 | 92 | — | — | — | — | — | — | — | 80.18 | — | — | — | |
| MasRouterLLM=LLM Pool2025.09 | 92 | — | — | — | — | — | — | — | 80.66 | — | — | — | |
| ASRRBackbone=Qwen3-4B, Sampling Strategy=mean@82026.03 | 92 | 289 | — | — | — | — | — | — | — | — | — | — | |
| Few-shot-CoT-CP (GPT-4) + SCBase Model=GPT-4, Prompting Strategy=Few-shot-CoT-CP, Self-Consistency=true2024.03 | 91.9 | — | — | — | — | — | — | — | — | — | — | — | |
| ExGRPOModel=Qwen2.5-7B-Instruct, Model Role=Student Model, Adapter=On-Policy Adapt.2026.02 | 91.85 | — | — | — | — | — | — | — | — | — | — | — | |
| EffReaBackbone=DeepSeek-R1-7B2026.02 | 91.67 | 1,082.18 | — | — | — | — | — | — | — | — | — | — | |
| S3-CoTscBackbone=DeepSeek-R1-7B2026.02 | 91.67 | 1,149.86 | — | — | — | — | — | — | — | — | — | — | |
| PIASTTime [m]=80.26±2.952025.12 | 91.65 | — | — | — | — | — | — | — | — | — | — | — | |
| MV length-pruneModel=Qwen3-1.7B2025.12 | 91.6 | — | — | — | — | — | — | 1.46 | — | — | — | — | |
| Weighted BoN Self-evalModel=Qwen3-1.7B2025.12 | 91.6 | — | — | — | — | — | — | 1.4 | — | — | — | — | |
| ExGRPOModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.02 | 91.46 | — | — | — | — | — | — | — | — | — | — | — | |
| Weighted BoN ext. RMModel=Qwen3-1.7B2025.12 | 91.4 | — | — | — | — | — | — | 1.43 | — | — | — | — | |
| VLPBackbone=Qwen3-4B, Sampling Strategy=mean@82026.03 | 91.3 | 405 | — | — | — | — | — | — | — | — | — | — | |
| Majority VotingModel=Qwen3-1.7B2025.12 | 91.2 | — | — | — | — | — | — | 1.4 | — | — | — | — | |
| S3-CoTBackbone=DeepSeek-R1-7B2026.02 | 91.17 | 1,182.04 | — | — | — | — | — | — | — | — | — | — | |
| AFlowLLM=gpt-4o-mini2025.09 | 91.16 | — | — | — | — | — | — | — | 79.73 | — | — | — | |
| DEERBackbone=DeepSeek-R1-Distill-Qwen-7B2026.03 | 91.05 | 985 | — | — | — | — | — | — | — | — | — | — | |
| COMT+CCRLBackbone=QWEN3-4B2026.01 | 90.9 | — | — | — | — | — | — | — | — | — | — | — | |
| pass@1Model=QWEN-2.5-7B2026.01 | 90.9 | — | — | — | — | — | — | — | — | — | — | — | |
| GRPOBackbone=Qwen3-4B, Sampling Strategy=mean@82026.03 | 90.9 | 722 | — | — | — | — | — | — | — | — | — | — | |
| Zero-shotModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.02 | 90.9 | — | — | — | — | — | — | — | — | — | — | — | |
| TokenSkipBackbone=Qwen2.5-7B2026.02 | 90.83 | 260.63 | — | — | — | — | — | — | — | — | — | — | |
| DeepSeek-R1-14BBackbone=DeepSeek-R1-14B2026.02 | 90.8 | 439.25 | — | — | — | — | — | — | — | — | — | — | |
| VanillaBase Model=Qwen3-1.7B2026.02 | 90.67 | 2,094 | — | — | — | — | — | — | — | — | — | — | |
| LightThinkerBackbone=DeepSeek-R1-Distill-Qwen-7B2026.03 | 90.55 | 1,166 | — | — | — | — | — | — | — | — | — | — | |
| CoT-ValveBackbone=Qwen2.5-7B2026.02 | 90.5 | 298.82 | — | — | — | — | — | — | — | — | — | — | |
| CODABackbone=Qwen3-4B, Sampling Strategy=mean@82026.03 | 90.5 | 168 | — | — | — | — | — | — | — | — | — | — | |
| L1-MaxBackbone=DeepSeek-R1-Distill-Qwen-7B2026.03 | 90.48 | 1,448 | — | — | — | — | — | — | — | — | — | — | |
| Step EntropyBackbone=DeepSeek-R1-Distill-Qwen-7B2026.03 | 90.46 | 1,269 | — | — | — | — | — | — | — | — | — | — | |
| ARLCPBase Model=DeepSeek-R1-Distill-Llama-8B2026.02 | 90.45 | 694 | — | — | — | — | — | — | — | — | — | — | |
| AFlowLLM=gemini-1.5-flash2025.09 | 90.43 | — | — | — | — | — | — | — | 77.29 | — | — | — | |
| MARSBackbone=Mixtral 8x22b2025.09 | 90.33 | 3,515 | — | — | — | — | — | — | — | — | — | — |