Saved in:
Bibliographic Details
Main Authors: Li, Zhuohang, Li, Xiaowei, Huang, Chengyu, Li, Guowang, Goshvadi, Katayoon, Dai, Bo, Schuurmans, Dale, Zhou, Paul, Palangi, Hamid, Song, Yiwen, Goyal, Palash, Kantarcioglu, Murat, Malin, Bradley A., Xue, Yuan
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.00263
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911187494502400
author Li, Zhuohang
Li, Xiaowei
Huang, Chengyu
Li, Guowang
Goshvadi, Katayoon
Dai, Bo
Schuurmans, Dale
Zhou, Paul
Palangi, Hamid
Song, Yiwen
Goyal, Palash
Kantarcioglu, Murat
Malin, Bradley A.
Xue, Yuan
author_facet Li, Zhuohang
Li, Xiaowei
Huang, Chengyu
Li, Guowang
Goshvadi, Katayoon
Dai, Bo
Schuurmans, Dale
Zhou, Paul
Palangi, Hamid
Song, Yiwen
Goyal, Palash
Kantarcioglu, Murat
Malin, Bradley A.
Xue, Yuan
contents The alignment of large language models (LLMs) with human values increasingly relies on using other LLMs as automated judges, or ``autoraters''. However, their reliability is limited by a foundational issue: they are trained on discrete preference labels, forcing a single ground truth onto tasks that are often subjective, ambiguous, or nuanced. We argue that a reliable autorater must learn to model the full distribution of preferences defined by a target population. In this paper, we propose a general framework for calibrating probabilistic autoraters to any given preference distribution. We formalize the problem and present two learning methods tailored to different data conditions: 1) a direct supervised fine-tuning for dense, probabilistic labels, and 2) a reinforcement learning approach for sparse, binary labels. Our empirical results show that finetuning autoraters with a distribution-matching objective leads to verbalized probability predictions that are better aligned with the target preference distribution, with improved calibration and significantly lower positional bias, all while preserving performance on objective tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2510_00263
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Judging with Confidence: Calibrating Autoraters to Preference Distributions
Li, Zhuohang
Li, Xiaowei
Huang, Chengyu
Li, Guowang
Goshvadi, Katayoon
Dai, Bo
Schuurmans, Dale
Zhou, Paul
Palangi, Hamid
Song, Yiwen
Goyal, Palash
Kantarcioglu, Murat
Malin, Bradley A.
Xue, Yuan
Computation and Language
The alignment of large language models (LLMs) with human values increasingly relies on using other LLMs as automated judges, or ``autoraters''. However, their reliability is limited by a foundational issue: they are trained on discrete preference labels, forcing a single ground truth onto tasks that are often subjective, ambiguous, or nuanced. We argue that a reliable autorater must learn to model the full distribution of preferences defined by a target population. In this paper, we propose a general framework for calibrating probabilistic autoraters to any given preference distribution. We formalize the problem and present two learning methods tailored to different data conditions: 1) a direct supervised fine-tuning for dense, probabilistic labels, and 2) a reinforcement learning approach for sparse, binary labels. Our empirical results show that finetuning autoraters with a distribution-matching objective leads to verbalized probability predictions that are better aligned with the target preference distribution, with improved calibration and significantly lower positional bias, all while preserving performance on objective tasks.
title Judging with Confidence: Calibrating Autoraters to Preference Distributions
topic Computation and Language
url https://arxiv.org/abs/2510.00263