Data Contamination Detection on AIME 2024
76F1 ScoreSelf-Critique
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Self-CritiqueTarget Model=DeepSeek-Math-7B-Instruct, Probing Mechanism=Self-Critique Probing, Core Metric=Entropy, Designed for=RL2025.10 | 76 | 67 | |
| Min-K%++Target Model=Qwen2.5-7B-Instruct, Probing Mechanism=Intrinsic Property, Core Metric=Log Probability, Designed for=Pre-training / SFT2025.10 | 73 | 58 | |
| Entropy-TempTarget Model=Qwen2.5-7B-Instruct, Probing Mechanism=Stochastic Sampling, Core Metric=Entropy, Designed for=RL2025.10 | 73 | 64 | |
| Entropy-NoiseTarget Model=Qwen2.5-7B-Instruct, Probing Mechanism=Non-member prefix, Core Metric=Entropy, Designed for=RL2025.10 | 70 | 57 | |
| Entropy-NoiseTarget Model=DeepSeek-Math-7B-Instruct, Probing Mechanism=Non-member prefix, Core Metric=Entropy, Designed for=RL2025.10 | 70 | 56 | |
| Self-CritiqueTarget Model=Qwen2.5-7B-Instruct, Probing Mechanism=Self-Critique Probing, Core Metric=Entropy, Designed for=RL2025.10 | 69 | 72 | |
| Min-K%++Target Model=DeepSeek-Math-7B-Instruct, Probing Mechanism=Intrinsic Property, Core Metric=Log Probability, Designed for=Pre-training / SFT2025.10 | 67 | 53 | |
| RecallTarget Model=Qwen2.5-7B-Instruct, Probing Mechanism=Non-member prefix, Core Metric=Log Probability, Designed for=Pre-training / SFT2025.10 | 62 | 61 | |
| Entropy-TempTarget Model=DeepSeek-Math-7B-Instruct, Probing Mechanism=Stochastic Sampling, Core Metric=Entropy, Designed for=RL2025.10 | 60 | 48 | |
| Min-K%Target Model=Qwen2.5-7B-Instruct, Probing Mechanism=Intrinsic Property, Core Metric=Log Probability, Designed for=Pre-training / SFT2025.10 | 59 | 49 | |
| Min-K%Target Model=DeepSeek-Math-7B-Instruct, Probing Mechanism=Intrinsic Property, Core Metric=Log Probability, Designed for=Pre-training / SFT2025.10 | 55 | 47 | |
| RecallTarget Model=DeepSeek-Math-7B-Instruct, Probing Mechanism=Non-member prefix, Core Metric=Log Probability, Designed for=Pre-training / SFT2025.10 | 54 | 46 | |
| CDDTarget Model=Qwen2.5-7B-Instruct, Probing Mechanism=Stochastic Sampling, Core Metric=Edit Distance, Designed for=Pre-training / SFT2025.10 | 50 | 57 | |
| PPLTarget Model=DeepSeek-Math-7B-Instruct, Probing Mechanism=Intrinsic Property, Core Metric=Log Probability, Designed for=Pre-training / SFT2025.10 | 42 | 53 | |
| PPLTarget Model=Qwen2.5-7B-Instruct, Probing Mechanism=Intrinsic Property, Core Metric=Log Probability, Designed for=Pre-training / SFT2025.10 | 33 | 51 | |
| CDDTarget Model=DeepSeek-Math-7B-Instruct, Probing Mechanism=Stochastic Sampling, Core Metric=Edit Distance, Designed for=Pre-training / SFT2025.10 | 30 | 49 |