Skip to content
CASRAI·Indexed by this site · Originally published· Alignment →· Its place in the scan record →

Hannah Li's study: why user adaptation distorts LLM benchmarks and A/B tests

Why User Adaptation Distorts LLM Benchmarks — CASRAI

AI summary

Columbia Business School researcher Hannah Li released a preprint on November 12, 2026 (arXiv:2611.17239v1) proving that user adaptation makes offline benchmarks and online A/B tests measure different statistics, leading to rank reversals. Users who receive high-quality answers escalate to harder tasks, while those who receive unhelpful answers lower the difficulty or leave, so that the A/B test's estimate of the mean difference is mixed with changes in task difficulty. Two rounds of email-writing experiments showed that M1 was no lower than M0 at every task level—90% usable for both on regular tasks, and 70% versus 30% on hard tasks—but 80% of its users escalated to hard tasks, and its two-round average success rate of 81.10% was lower than M0's 86.65%, so the online A/B test selected the weaker M0. The study also proves that dependence between rounds amplifies estimation variance, and explains reward hacking and novelty effects: a model can raise aggregate metrics without improving core capability by steering users toward simpler prompts. The authors recommend that evaluation teams report joint metrics such as task-stratified accuracy, changes in prompt composition, and users' task-escalation trajectories.

Related papers
Read originalcasrai.org