General Reasoning on OOD Performance Suite (ARC-c, GPQA*, MMLU-Pro) (test)
91.4ARC-c ScoreOn-Policy
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| On-PolicyBackbone=Qwen2.5-7B Instruct2025.10 | 91.4 | 19.2 | 57.6 | 56 | |
| ExGRPOBackbone=Qwen2.5-7B Instruct2025.10 | 91.1 | 36.9 | 58.6 | 62.2 | |
| ExGRPOBackbone=Llama3.1-8B Instruct2025.10 | 88.9 | 15.2 | 52.1 | 52 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.1, Optimization=w/ OLR2026.04 | 88.9 | 37.9 | 59.5 | 62.1 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.32026.04 | 88.1 | 29.8 | 56.2 | 58 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.3, Optimization=w/ OLR2026.04 | 88.1 | 36.4 | 57.6 | 60.7 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.7, Optimization=w/ OLR2026.04 | 87.4 | 29.8 | 57.3 | 58.2 | |
| GRPO w/ Confidence PenaltyBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Inactive Noisy Label2026.04 | 87.4 | 31.3 | 56.6 | 58.4 | |
| TTRLBackbone=Qwen3-4B-Base, Noise Ratio=50%, Learning Strategy=Unsupervised2026.04 | 86.9 | 30.3 | 56 | 57.7 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.5, Optimization=w/ OLR2026.04 | 86.8 | 38.4 | 57.8 | 61 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.1, Optimization=w/ OLR2026.04 | 86.8 | 35.4 | 59.1 | 60.4 | |
| GRPO w/ OLRBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Inactive Noisy Label2026.04 | 86.8 | 38.4 | 57.8 | 61 | |
| On-PolicyBackbone=Llama3.1-8B Instruct2025.10 | 86.7 | 0 | 48.1 | 44.9 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.52026.04 | 86.7 | 28.8 | 56 | 57.2 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Inactive Noisy Label2026.04 | 86.7 | 28.8 | 56 | 57.2 | |
| GRPO w/ Label SmoothingBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Inactive Noisy Label2026.04 | 86.5 | 29.8 | 57.4 | 57.9 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.12026.04 | 86.3 | 37.4 | 59 | 60.9 | |
| Co-rewardBackbone=Qwen3-4B-Base, Noise Ratio=50%, Learning Strategy=Unsupervised2026.04 | 86.1 | 29.3 | 55.8 | 57.1 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.9, Optimization=w/ OLR2026.04 | 85.9 | 30.3 | 55.3 | 57.2 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.3, Optimization=w/ OLR2026.04 | 85.4 | 33.3 | 56.7 | 58.5 | |
| DAPO + EchoRLTraining Category=On-Policy RLVR with EchoRL, EchoRL enhancement=true, Backbone Model=Qwen2.5-Math-7B2026.05 | 85.2 | 42.9 | 54.7 | 60.9 | |
| InstructBackbone=Qwen2.5-7B Instruct2025.10 | 84.7 | 24.7 | 54.7 | 54.7 | |
| GRPO + EchoRLTraining Category=On-Policy RLVR with EchoRL, EchoRL enhancement=true, Backbone Model=Qwen2.5-Math-7B2026.05 | 84.7 | 37.4 | 53 | 58.4 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.12026.04 | 84.6 | 30.3 | 55.5 | 56.8 | |
| LUFFY + EchoRLTraining Category=Expert-Supervised RLVR with EchoRL, EchoRL enhancement=true, Backbone Model=Qwen2.5-Math-7B2026.05 | 83.6 | 45.3 | 54.1 | 61 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.72026.04 | 83.4 | 26.3 | 54.5 | 54.7 | |
| GRPO w/ Confidence PenaltyBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Active Noisy Label2026.04 | 82.9 | 32.3 | 53.3 | 56.2 | |
| UFT + EchoRLTraining Category=Expert-Supervised RLVR with EchoRL, EchoRL enhancement=true, Backbone Model=Qwen2.5-Math-7B2026.05 | 82.7 | 43.4 | 53.5 | 59.9 | |
| UFTTraining Category=Expert-Supervised RLVR with EchoRL, EchoRL enhancement=false, Backbone Model=Qwen2.5-Math-7B2026.05 | 82.2 | 38.9 | 49.6 | 56.9 | |
| DAPOTraining Category=On-Policy RLVR with EchoRL, EchoRL enhancement=false, Backbone Model=Qwen2.5-Math-7B2026.05 | 82.1 | 38.9 | 51.8 | 57.6 | |
| GRPOTraining Category=On-Policy RLVR with EchoRL, EchoRL enhancement=false, Backbone Model=Qwen2.5-Math-7B2026.05 | 81.8 | 39.9 | 45.2 | 55.6 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.32026.04 | 80.8 | 25.3 | 51.7 | 52.6 | |
| LUFFYTraining Category=Expert-Supervised RLVR with EchoRL, EchoRL enhancement=false, Backbone Model=Qwen2.5-Math-7B2026.05 | 80.5 | 39.9 | 53 | 57.8 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.7, Optimization=w/ OLR2026.04 | 80.1 | 31.8 | 52.4 | 54.8 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.5, Optimization=w/ OLR2026.04 | 78.5 | 28.3 | 54.8 | 53.9 | |
| GRPO w/ OLRBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Active Noisy Label2026.04 | 78.5 | 28.3 | 54.8 | 53.9 | |
| GRPO w/ Random SelectBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Inactive Noisy Label2026.04 | 78.1 | 31.8 | 55 | 55 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Inactive Noisy Label, Noise Ratio (rho)=0.92026.04 | 76.5 | 27.3 | 51.5 | 51.8 | |
| SFTTraining Category=Supervised Learning Methods, Backbone Model=Qwen2.5-Math-7B2026.05 | 75.2 | 24.7 | 42.7 | 47.5 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.52026.04 | 73.9 | 24.7 | 48.6 | 49.1 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Active Noisy Label2026.04 | 73.9 | 24.7 | 48.6 | 49.1 | |
| GRPO w/ Samll-loss SelectBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Inactive Noisy Label2026.04 | 73.9 | 30.3 | 50.7 | 51.6 | |
| PRIME-ZeroTraining Category=Previous Zero RLVR Methods, Backbone Model=Qwen2.5-Math-7B2026.05 | 73.3 | 18.2 | 32.7 | 41.4 | |
| Qwen2.5-Math-7B-InstructBackbone Model=Qwen2.5-Math-7B2026.05 | 70.3 | 24.7 | 34.1 | 43 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.72026.04 | 66.2 | 24.7 | 49.2 | 46.7 | |
| OpenReasoner-ZeroTraining Category=Previous Zero RLVR Methods, Backbone Model=Qwen2.5-Math-7B2026.05 | 66.2 | 29.8 | 58.7 | 51.6 | |
| ExGRPOBackbone=Qwen2.5-Math 1.5B2025.10 | 59.3 | 30.8 | 30.4 | 40.2 | |
| ExGRPOBackbone=Llama3.1-8B2025.10 | 59 | 7.1 | 26.2 | 30.8 | |
| On-PolicyBackbone=Qwen2.5-Math 1.5B2025.10 | 57.5 | 28.3 | 30.7 | 38.9 | |
| GRPO w/ Label SmoothingBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Active Noisy Label2026.04 | 55.5 | 24.2 | 47.4 | 42.4 | |
| GRPO w/ Random SelectBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Active Noisy Label2026.04 | 45.6 | 14.6 | 46.3 | 35.5 | |
| InstructBackbone=Llama3.1-8B Instruct2025.10 | 42.4 | 0 | 40.5 | 27.6 | |
| Qwen3-4B-BaseBackbone=Qwen3-4B-Base2026.04 | 35.8 | 14.1 | 33.3 | 27.7 | |
| Qwen3-4B-BaseNoise Ratio=50%2026.04 | 35.8 | 14.1 | 33.3 | 27.7 | |
| GRPOBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.92026.04 | 34.2 | 12.6 | 30.5 | 25.8 | |
| SFT-KLTraining Category=Supervised Learning Methods, Backbone Model=Qwen2.5-Math-7B2026.05 | 34.1 | 23.2 | 31.5 | 29.6 | |
| GRPO w/ Samll-loss SelectBackbone=Qwen3-4B-Base, Noise Ratio=50%, Noise Scenario=Active Noisy Label2026.04 | 32.3 | 9.6 | 26.7 | 22.9 | |
| OLRBackbone=Qwen3-4B-Base, Noise Scenario=Active Noisy Label, Noise Ratio (rho)=0.9, Optimization=w/ OLR2026.04 | 31.1 | 16.2 | 32.1 | 26.5 | |
| SimpleRL-ZeroTraining Category=Previous Zero RLVR Methods, Backbone Model=Qwen2.5-Math-7B2026.05 | 30.2 | 23.2 | 34.5 | 29.3 | |
| PivotTraceBackbone=Deepseek-R1-Distill-Llama-8B, multiple stochastic inferences=false2026.06 | 26.9 | 23.4 | 46.3 | 32.2 | |
| RandomBackbone=Deepseek-R1-Distill-Llama-8B, multiple stochastic inferences=false2026.06 | 25.5 | 22.2 | 43.2 | 30.3 | |
| Self-certaintyBackbone=Qwen3-4B-Base, Noise Ratio=50%, Learning Strategy=Unsupervised2026.04 | 24.3 | 12.1 | 30.6 | 22.3 | |
| ConsistencyBackbone=Deepseek-R1-Distill-Llama-8B, multiple stochastic inferences=true2026.06 | 23.3 | 20.3 | 40.7 | 28.1 | |
| Seq-entropyBackbone=Qwen3-4B-Base, Noise Ratio=50%, Learning Strategy=Unsupervised2026.04 | 18.6 | 6.6 | 28.4 | 17.9 | |
| Qwen2.5-Math-7BBackbone Model=Qwen2.5-Math-7B2026.05 | 18.2 | 11.1 | 16.9 | 15.4 | |
| Token-entropyBackbone=Qwen3-4B-Base, Noise Ratio=50%, Learning Strategy=Unsupervised2026.04 | 17 | 7.6 | 28.9 | 17.8 | |
| EntropyBackbone=Deepseek-R1-Distill-Llama-8B, multiple stochastic inferences=false2026.06 | 16.5 | 23.1 | 47.5 | 29 | |
| CoT-KineticsBackbone=Deepseek-R1-Distill-Llama-8B, multiple stochastic inferences=false2026.06 | 15.6 | 18.9 | 42.1 | 25.5 | |
| Self-CertaintyBackbone=Deepseek-R1-Distill-Llama-8B, multiple stochastic inferences=false2026.06 | 15.6 | 20.2 | 42.9 | 26.2 | |
| CoEBackbone=Deepseek-R1-Distill-Llama-8B, multiple stochastic inferences=false2026.06 | 8.7 | 18.4 | 46.8 | 24.6 | |
| BaseBackbone=Qwen2.5-Math 1.5B2025.10 | 3.1 | 3 | 2.3 | 2.8 | |
| On-PolicyBackbone=Llama3.1-8B2025.10 | 0.6 | 0 | 3 | 1.2 | |
| BaseBackbone=Llama3.1-8B2025.10 | 0.2 | 0 | 3.6 | 1.3 |