LLM Performance Estimation on MMLU (test)
0.842MAE (%)SparseEval
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| SparseEvalNumber of Anchors=1002026.02 | 0.842 | 0.908 | — | — | — | — | — | — | |
| SparseEvalNumber of Anchors=802026.02 | 0.962 | 0.896 | — | — | — | — | — | — | |
| SparseEvalNumber of Anchors=602026.02 | 0.997 | 0.89 | — | — | — | — | — | — | |
| SparseEvalNumber of Anchors=402026.02 | 1.282 | 0.871 | — | — | — | — | — | — | |
| SparseEvalNumber of Anchors=202026.02 | 1.718 | 0.832 | — | — | — | — | — | — | |
| Anchor PointsNumber of Anchors=802026.02 | 1.964 | 0.856 | — | — | — | — | — | — | |
| TailoredBenchNumber of Anchors=1002026.02 | 2.019 | 0.862 | — | — | — | — | — | — | |
| gp-IRTNumber of Anchors=1002026.02 | 2.202 | 0.829 | — | — | — | — | — | — | |
| TailoredBenchNumber of Anchors=802026.02 | 2.216 | 0.876 | — | — | — | — | — | — | |
| Anchor PointsNumber of Anchors=602026.02 | 2.331 | 0.85 | — | — | — | — | — | — | |
| TailoredBenchNumber of Anchors=602026.02 | 2.421 | 0.857 | — | — | — | — | — | — | |
| gp-IRTNumber of Anchors=802026.02 | 2.537 | 0.798 | — | — | — | — | — | — | |
| TailoredBenchNumber of Anchors=402026.02 | 2.677 | 0.845 | — | — | — | — | — | — | |
| Anchor PointsNumber of Anchors=402026.02 | 2.83 | 0.801 | — | — | — | — | — | — | |
| gp-IRTNumber of Anchors=602026.02 | 3.19 | 0.71 | — | — | — | — | — | — | |
| gp-IRTNumber of Anchors=402026.02 | 3.802 | 0.692 | — | — | — | — | — | — | |
| TailoredBenchNumber of Anchors=202026.02 | 4.046 | 0.755 | — | — | — | — | — | — | |
| Anchor PointsNumber of Anchors=202026.02 | 4.898 | 0.727 | — | — | — | — | — | — | |
| gp-IRTNumber of Anchors=202026.02 | 5.94 | 0.569 | — | — | — | — | — | — | |
| Anchor PointsNumber of Anchors=1002026.02 | 7.89 | 0.764 | — | — | — | — | — | — | |
| Domain-Proxy BaselineBase Model=Llama-3-8B-Instruct2026.06 | — | — | 10.01 | 0.66 | 0.81 | 2.1 | 7.4 | 16.1 | |
| Domain-Proxy BaselineBase Model=Qwen2-0.5B2026.06 | — | — | 9.37 | 0.73 | 0.85 | 10 | 16.4 | 25.3 | |
| Domain-Proxy BaselineBase Model=Qwen2.5-7B-Instruct2026.06 | — | — | 6.58 | 0.85 | 0.92 | 8.6 | 22 | 32.8 | |
| Early-Dynamics BaselineBase Model=Llama-3-8B-Instruct2026.06 | — | — | 7.33 | 0.78 | 0.88 | 21.5 | 37.7 | 48.6 | |
| Early-Dynamics BaselineBase Model=Qwen2-0.5B2026.06 | — | — | 7.47 | 0.81 | 0.9 | 19.3 | 28.8 | 39.7 | |
| Early-Dynamics BaselineBase Model=Qwen2.5-7B-Instruct2026.06 | — | — | 3.33 | 0.96 | 0.98 | 29.9 | 50 | 67.5 | |
| Early-Stop ExtrapolationBase Model=Llama-3-8B-Instruct2026.06 | — | — | 10.46 | 0.66 | 0.81 | 3.7 | 8.8 | 17.8 | |
| Early-Stop ExtrapolationBase Model=Qwen2-0.5B2026.06 | — | — | 10.41 | 0.69 | 0.83 | 11.7 | — | 18.7 | |
| Early-Stop ExtrapolationBase Model=Qwen2.5-7B-Instruct2026.06 | — | — | 7.43 | 0.81 | 0.9 | 11.2 | 23.9 | 32.8 | |
| ProxyLMBase Model=Llama-3-8B-Instruct2026.06 | — | — | 6.76 | 0.86 | 0.93 | 30.9 | 45.8 | 62.5 | |
| ProxyLMBase Model=Qwen2-0.5B2026.06 | — | — | 4 | 0.9 | 0.95 | 35.6 | 53.7 | 72.1 | |
| ProxyLMBase Model=Qwen2.5-7B-Instruct2026.06 | — | — | 2.11 | 0.98 | 0.99 | 40.7 | 67.9 | 85.8 | |
| TUNEAHEAD (Full)Base Model=Llama-3-8B-Instruct2026.06 | — | — | 5.02 | 0.86 | 0.93 | 36 | 55.8 | 73.3 | |
| TUNEAHEAD (Full)Base Model=Qwen2-0.5B2026.06 | — | — | 3.75 | 0.91 | 0.95 | 39.2 | 58.4 | 74.6 | |
| TUNEAHEAD (Full)Base Model=Qwen2.5-7B-Instruct2026.06 | — | — | 1.47 | 0.99 | 0.99 | 50 | 82.5 | 95.1 | |
| TUNEAHEAD-Dynamic-OnlyBase Model=Qwen2.5-7B-Instruct2026.06 | — | — | 3.38 | 0.96 | 0.99 | 19.8 | 35.8 | 55.6 | |
| TUNEAHEAD-Static-OnlyBase Model=Qwen2.5-7B-Instruct2026.06 | — | — | 3.5 | 0.96 | 0.99 | 14.9 | 32.8 | 49.3 |