The paper introduces prediction‑powered smoothing (PP‑S) and its taxonomy‑aware extension (PP‑TS) to improve point and interval estimates of domain‑specific AI performance when only a limited sample of labeled units is available. It also proposes a new design‑based cross‑validation score that is approximately unbiased for selecting between direct and smoothed estimators. Experiments on a curated benchmark and real‑world agent traffic show that the proposed methods outperform direct estimators in both accuracy and coverage, and that the new score matches the performance of an independent validation sample while providing more precise error estimates.
arXiv:2607. 05046v1 Announce Type: new Abstract: Evaluating generative AI models is a routine, but resource-intensive, process that is conducted over and over again during the course of model development.
arXiv:2604. 23099v2 Announce Type: replace-cross Abstract: Evaluating generative AI models is increasingly resource-intensive due to slow inference, expensive raters, and a rapidly growing landscape of models and benchmarks.
arXiv:2607. 16239v1 Announce Type: new Abstract: AI judges offer a scalable, low-cost alternative to human evaluation, but their outputs can be biased relative to human preferences and highly item-dependent, varying across judges, tasks, and domains.
The paper introduces prediction‑powered evaluation, a framework that blends limited human judgments with large‑scale automatic scores to produce unbiased, data‑efficient system comparisons. It offers both parametric and non‑parametric methods, examines the trade‑off between paired and unpaired designs, and validates the approach on six WMT datasets. Additionally, the authors propose the Prediction‑Powered Saving Ratio (PPSR), a meta‑metric that quantifies how much human annotation can be saved by using an automatic metric within this framework, providing more discriminative and stable metric rankings than existing system‑level meta‑metrics.
arXiv:2609.35815v1 Announce Type: cross Abstract: Researchers across academia increasingly base significance claims on LLM judge scores and small-sample AI evaluations. Yet without well-calibrated co...
arXiv:2606. 29784v1 Announce Type: cross Abstract: Reliable generative AI models critically rely on expert human annotations to evaluate output quality, yet these "gold" labels are expensive to collect and limited in quantity.
arXiv:2607. 02032v1 Announce Type: new Abstract: Evaluating LLM agents on benchmarks like SWE-Bench and GAIA can be expensive, time-consuming, and requires complex infrastructure.
The paper proposes a two‑stage framework that first fine‑tunes a large language model (LLM) and then rectifies its outputs, allocating limited labeled data optimally between the stages. It argues that the usual mean‑squared‑error objective for fine‑tuning misaligns with the downstream rectification, and instead suggests minimizing prediction‑error variance for mean estimation or a scalarized variance metric for general M‑estimation. Empirical results confirm that this variance‑based fine‑tuning, combined with optimal data allocation, yields significant efficiency gains over using either fine‑tuning or rectification alone, or using the conventional objective.
arXiv:2607. 08347v1 Announce Type: cross Abstract: Active testing provides a label--efficient approach to risk estimation by adaptively selecting which test points should be labelled.
arXiv:2605. 31278v2 Announce Type: replace-cross Abstract: Reliable evaluation of agentic systems requires unbiased estimates with valid uncertainty, but standard practice navigates between costly human annotation and biased LLM-as-judge proxies.
The paper presents a layered framework for evaluating conversational AI by aligning offline proxy signals with online A/B experiment outcomes. It introduces a three‑step alignment chain—behavioral label to product outcome, classifier to candidate behavior, and offline signal to experiment effect—alongside an audit protocol that compares confidence intervals and rankings. In a real‑world deployment, the composite proxy achieved 81.1% F1 versus 34.3% for the raw classifier, correctly predicting direction on all 113 contrasts and enabling efficient prioritization of candidate models before costly online testing.
arXiv:2602. 15327v2 Announce Type: replace-cross Abstract: Machine learning model performance improvements tend to arise from competition and application.