benchgap
Calibration

SciCode → LiveCodeBench v6

LiveCodeBench v6 is estimated from SciCode with a Hill curve fitted on 14 models measured on both: y = 0.6575 + (0.9644 − 0.6575)·x^6.00 / (0.41494^6.00 + x^6.00), R² = 0.88, cross-validated error 4.4 pp. It is used for 9 estimates.

Estimated modelSciCodeLiveCodeBench v6Source
Step 5 Preview58.9%93.1%estimated ± 4.4 pp, high confidence
Hy3 Preview41.2%80.8%estimated ± 4.4 pp, high confidence
Ling 3.0 Flash FP840.4%79.8%estimated ± 4.4 pp, high confidence
Mercury 2.538.0%77.1%estimated ± 4.4 pp, high confidence
K-EXAONE 2.037.4%76.5%estimated ± 4.4 pp, high confidence
LFM2.5-24B-A2B32.7%71.7%estimated ± 4.4 pp, high confidence
Nemotron 3 Nano Omni 30B A3B32.0%71.1%estimated ± 4.4 pp, high confidence
Nemotron 3.5 Lightning 30B A3B NVFP431.4%70.6%estimated ± 4.4 pp, high confidence
Ling 2.6 Flash27.0%67.9%estimated ± 4.4 pp, high confidence

Fitted on

ModelSciCodeLiveCodeBench v6
A.X K241.0%84.0%
Agents-A1-4B29.6%59.6%
Granite 4.2 30B38.8%75.8%
Granite 4.2 3B24.1%69.7%
Granite 4.2 8B36.1%73.2%
Kimi K2.652.2%89.6%
Kimi K2.548.7%85.0%
LFM3-1.7B28.5%73.5%
MiniCPM5-2B26.3%69.1%
Nemotron 3 Ultra44.6%89.0%
Qwen3.7 Max53.5%91.6%
Qwen3.7 Plus51.3%89.6%
Sakana Fugu60.1%92.9%
Sakana Fugu-Ultra58.7%93.2%