Adaptive tests speed up language model answer selection with error guarantees
Sequential Anytime Valid Tests for Efficient Inference-Time Alignment
Machine Learning
Summary
Choosing the best answer from multiple options generated by a language model can be slow because traditional methods spend the same amount of time on every question. The authors developed a new testing method that decides when to stop looking for better answers based on a formal error tolerance, ensuring the most likely correct answer is selected without wasting effort. Their method uses a sophisticated mathematical approach called a betting supermartingale to provide strong guarantees on the correctness of the chosen answer while saving computational resources. Experiments show this new approach is more efficient and accurate than previous stopping strategies.
What this means in practice
- •For language ai engineers: Reduce the computation needed to select high-quality answers from large language models by using statistically sound adaptive stopping rules during inference.
- •For real-time chatbot operators: Improve response speed while maintaining answer quality by halting answer generation as soon as the system is confident about the best choice.
Authors
Vishesh Kumar Singh, Sayak Ray Chowdhury
Abstract
Inference-time alignment methods, such as Best-of-N (BoN), improve language model outputs by generating multiple candidates per prompt and selecting among them using a reward model. Much of this cost is wasted: a fixed budget is applied to every prompt, although the computation required by prompts varies widely. Adaptive methods halt once sampled answers appear settled, but their thresholds are tuned heuristics with no statistical control over early-stopping errors. We instead seek an anytime-valid stopping rule that takes an error tolerance $α$ as input and halts only once an answer is certified to hold a majority ($θ_a > 1/2$) under the generator's distribution. We aggregate candidates by final answer with a reward-weighted self-consistency (RWSC) score and treat stopping as a sequential test over the resulting clusters. Applying Wald's sequential probability ratio test to these clusters yields SPRT-RWSC, which performs well empirically, but its guarantee fails because the leading answer shifts during sampling, and the run is continuously monitored. Our main contribution, e-RWSC, replaces Wald's test with a betting supermartingale (an e-process) for every answer the policy produces, bounding the probability of ever certifying a non-dominant answer by $α$, uniformly over time and across all answers. The RWSC scores drive the bets, so they affect only how quickly a dominant answer $a^\star$ is certified, not validity; certification takes $O(\log(1/α)/d_{\mathrm{KL}}(θ_{a^\star}\|1/2))$ samples. Experiments across several language models, datasets, and reward models confirm the predicted $1/d_{\mathrm{KL}}$ scaling of cost and show that accuracy improves over baselines at substantially lower token cost.