Gabriel Sargent, Will Wei Sun, Zhengwu Zhang, Yufeng Liu
原文摘要
Ranking large language models (LLMs) through human evaluations is essential for downstream decision-making. Because human feedback is expensive, evaluation platforms often collect it adaptively, prioritizing informative model comparisons using previously observed feedback. To further improve precision with limited human feedback, evaluators can draw on abundant unrated responses and relatively inexpensive LLM judges; however, judge bias and adaptive dependence challenge standard i.i.d. inference frameworks. We propose Prediction-Powered Ranking (PP-Rank), a framework that combines LLM-judge predictions with human feedback to improve precision while providing valid inference under adaptive sampling. Data-driven prediction weights minimize the asymptotic covariance trace within the proposed class. In theory, we establish asymptotic normality for general evaluation targets and simultaneous ranking inference. A joint martingale analysis handles dependence between adaptive human feedback and auxiliary data. We further derive a rule for reducing exploration that prioritizes lower-cost comparisons while matching a conservative variance bound, and characterize allocation limits for regular $\sqrt{T}$-rate estimation of pairwise means. Semi-synthetic experiments on UltraFeedback-Extended show improvements over human-only baselines in score estimation, confidence interval width, and ranking certification. Synthetic experiments demonstrate robustness to judge quality and adaptive allocation, while showing that PP-Rank can substantially reduce evaluation costs without sacrificing precision.
Aligning large language models (LLMs) with human preferences is critical to recent advances in generative artificial intelligence. Reinforcement learning from human feedback (RLHF) is widely applied to achieve this objective. A key step in RLHF is to learn the reward function from human feedback. However, human feedback is costly and time-consuming, making it essential to collect high-quality conversation data for human teachers to label. Additionally, different human teachers have different levels of expertise. It is thus critical to query the most appropriate teacher for their opinions. In this paper, we use reinforcement learning (RL) to formulate the alignment problem. Motivated by the idea of $D$-optimal design, we first propose a dual active reward learning algorithm for the simultaneous selection of conversations and teachers. Next, we apply pessimistic RL to solve the alignment problem, based on the learned reward estimator. Theoretically, we show that the reward estimator obtained through our proposed adaptive selection strategy achieves minimal generalized variance asymptotically, and prove that the sub-optimality of our pessimistic policy scales as $O(1/\sqrt{T})$ with a given sample budget $T$. Through simulations and experiments on LLMs, including annotator-level preference data and downstream Best-of-$N$ response-selection evaluations, we demonstrate the effectiveness of our algorithm and its superiority over existing baselines.
Organizations adopting AI face the challenge of deciding which outputs can be released without human review while meeting the quality standards of human experts. We develop a statistical framework for delegating tasks between experts and black-box generate-verify workflows, in which AI-generated answers are verified and revised over multiple rounds. Before deployment, human experts evaluate selected outputs, and their judgments are used to calibrate the verifier threshold and number of attempts. Our method selects the policy with the highest estimated automation rate among statistically approved candidates, without modifying the generator or verifier. We establish finite-sample control of the error rate among released outputs, characterize the expert evaluations sufficient to attain a target automation rate, and derive a minimax lower bound establishing rate optimality in the quality margin and required confidence. In a MultiHiertt financial question-answering case study, average held-out automation increases from 2.50% to 24.17% as expert evaluations accumulate, while released-output error rates remain around 1.20%. In contrast, heuristics that release answers after consistently high verifier scores either violate the 5% error tolerance or automate at most 1.30% of tasks. Synthetic experiments further examine these theoretical predictions under controlled settings.