User adaptivity can reverse LLM rankings between offline and live tests
Original title: When Both Sides Adapt: The Challenges of Evaluating Interactive LLM Systems
Offline difference-in-means tracks a shared prompt distribution, while live A/B tests track prompts users form after seeing each model.
The paper argues that interactive LLM users change prompting strategy and later task difficulty after seeing responses, so model outputs reshape the prompt distribution used for later evaluation. It defines a common-distribution effect, comparing models on the same prompts, and an induced-distribution effect, comparing each model on the prompts its own interaction produces. Offline benchmarks and live A/B tests estimate those different quantities and can rank models differently. In a stylized linear interaction model, user adaptivity also raises estimation variance and is linked to reward hacking and novelty effects.
- Standard difference-in-means estimators converge to different effects for the two methods.
- The paper gives a sufficient condition for the two rankings to agree.
- A stylized linear model shows adaptivity can increase estimation variance.
- Reward hacking and novelty effects are treated as shared consequences of adaptivity.
From the original article
1 min read at arxiv.orgIn interactive systems where users engage with LLMs, users are adaptive and, after observing LLM responses, can change their prompting strategy, as well as the type and difficulty of subsequent tasks. Thus model responses can influence the set of prompts on which the model is subsequently evaluated. We study the implications of this user adaptivity and prompt endogeneity for the evaluation of LLMs.
By Hannah Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.