Question Answering on MMLU (test)
80.48Accuracy5-shot
Evaluation results
| Method | Links | |
|---|---|---|
| 5-shotBackbone=Qwen32026.01 | 80.48 | |
| SFTBackbone=Qwen32026.01 | 79.08 | |
| kNN-MoEBackbone=Qwen32026.01 | 78.86 | |
| Zero-shotBackbone=Qwen32026.01 | 78.59 | |
| SFT (Router Only)Backbone=Qwen32026.01 | 78.36 | |
| MLP ProbeEvaluation Protocol=Probe2026.02 | 73.5 | |
| Teacher 5-shotEvaluation Protocol=5-shot2026.02 | 71.4 | |
| SFTBackbone=GPT-OSS2026.01 | 70.28 | |
| kNN-MoEBackbone=GPT-OSS2026.01 | 70.28 | |
| Zero-shotBackbone=GPT-OSS2026.01 | 70.2 | |
| SFT (Router Only)Backbone=GPT-OSS2026.01 | 69.18 | |
| 5-shotBackbone=GPT-OSS2026.01 | 63.72 | |
| kNN-MoEBackbone=OLMoE2026.01 | 47.81 | |
| SFTBackbone=OLMoE2026.01 | 46.89 | |
| Zero-shotBackbone=OLMoE2026.01 | 46.77 | |
| SFT (Router Only)Backbone=OLMoE2026.01 | 45.27 | |
| Label SmoothingBackbone=DeBERTa-v3-base, Parameters=86M2026.02 | 33.8 | |
| 5-shotBackbone=OLMoE2026.01 | 33.06 | |
| SupervisedBackbone=DeBERTa-v3-base, Parameters=86M2026.02 | 31.8 | |
| PROBE-KD (MLP)Backbone=DeBERTa-v3-base, Parameters=86M, Probe Type=MLP2026.02 | 30.7 | |
| Feature-KDBackbone=DeBERTa-v3-base, Parameters=86M, Distillation Strategy=Feature-KD2026.02 | 29.6 | |
| PROBE-KD (Logistic)Backbone=DeBERTa-v3-base, Parameters=86M, Probe Type=Logistic2026.02 | 27.1 | |
| PROBE-KD (CCS)Backbone=DeBERTa-v3-base, Parameters=86M, Probe Type=CCS2026.02 | 26.8 | |
| Patient-KDBackbone=DeBERTa-v3-base, Parameters=86M, Distillation Strategy=Patient-KD2026.02 | 25.7 | |
| Logit-KDBackbone=DeBERTa-v3-base, Parameters=86M, Distillation Strategy=Logit-KD2026.02 | 24.5 |