The benchmark procedure is stable under stochastic decoding, with a maximum accuracy standard deviation of 0.43% over 20 repeated runs¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Running the 899 CDPK questions 20 times on a subset of 8 models yielded a maximum standard deviation of accuracy of 0.43%, indicating the test procedure and scores are robust and repeatable. → Lelièvre 2025

Evidence¶

Lelièvre 2025¶

Lelièvre, M., Waldock, A., Liu, M., Valdes Aspillaga, N., Mackintosh, A., Ogando Portela, M. J., Lee, J., Atherton, P., Ince, R. A. A., & Garrod, O. G. B. (2025). Benchmarking the Pedagogical Knowledge of Large Language Models. https://arxiv.org/abs/2506.18710 doi:10.48550/arxiv.2506.18710

q2 · i? · design · r2

Stability experiment in which the 899 CDPK questions were run 20 times on a subset of 8 models sampled across the obtained accuracy range, to evaluate benchmark robustness given stochastic LLM decoding. The article reports the maximum standard deviation of accuracy as 0.43%.

"Over these 8 models the maximum standard deviation of accuracy (over the 20 runs) was 0.43%. This shows that our test procedure and result scores are robust and repeatable."

Discussion¶

Related Claims¶

• LLM accuracy on the CDPK pedagogy benchmark ranges from 28% to 89%, with closed-weight reasoning models dominating the top of the leaderboard — related
• SEND and CDPK benchmark results correlate highly (r = 0.94), but higher-performing models do relatively better on general pedagogy than SEND — related
• CDPK performance scales with model size, with a sharp Pareto-frontier drop-off below around 8B parameters — related
• Accuracy-cost value frontier: at $0.10 per million tokens, CDPK performance rose from 50% (April 2024) to 82% (June 2025) — related
• Estimated human baseline on the CDPK benchmark is approximately 50%, based on aggregate Chilean teacher exam results — related
• AI grading of the exam is highly stable across five independent runs at the total-score level (ICC(A,1) = 0.967), with lower but still strong cell-level stability (ICC(A,1) = 0.836) — related
• Smaller local LLMs show a minor position bias toward option ordering and partially guessing-like error variation, but accuracy remains relatively stable across permutations and repeated sampling runs — related