benchgap
Calibration

Vals LiveCodeBench → NL2Repo

NL2Repo is estimated from Vals LiveCodeBench with a offset logistic curve fitted on 12 models measured on both: y = 0.4596 + (0.6498 − 0.4596) / (1 + exp(−200.00·(x − 0.8755))), R² = 0.31, cross-validated error 8.3 pp. It is used for 4 estimates.

Estimated modelVals LiveCodeBenchNL2RepoSource
GPT-5.1-Codex85.6%46.3%estimated ± 8.3 pp, low confidence
GPT-5.1-Codex-Max83.6%46.0%estimated ± 8.3 pp, low confidence
GLM-4.681.0%46.0%estimated ± 8.3 pp, low confidence
GLM-4.567.4%46.0%estimated ± 8.3 pp, low confidence

Fitted on

ModelVals LiveCodeBenchNL2Repo
DeepSeek V4 Flash 073187.3%54.2%
DeepSeek V4 Pro 081387.5%61.5%
GLM-5.181.4%42.7%
GLM-5.269.5%48.9%
GLM-5.380.5%58.0%
GLM-5.3-Flash80.5%56.3%
MiniMax M2.779.9%39.8%
MiniMax M382.2%42.1%
Qwen3.6 Plus86.0%37.9%
Qwen3.7 Max87.1%47.2%
Qwen3.8-27B84.0%42.3%
Qwen3.8 Max87.9%55.9%