Saved in:
Bibliographic Details
Main Authors: Cheng, Wei, Wang, Tianlu, Ji, Yanmin, Yang, Fan, Tan, Keren, Zheng, Yiyu
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2410.02210
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912056646565888
author Cheng, Wei
Wang, Tianlu
Ji, Yanmin
Yang, Fan
Tan, Keren
Zheng, Yiyu
author_facet Cheng, Wei
Wang, Tianlu
Ji, Yanmin
Yang, Fan
Tan, Keren
Zheng, Yiyu
contents While in-context learning with large language models (LLMs) has shown impressive performance, we have discovered a unique miscalibration behavior where both correct and incorrect predictions are assigned the same level of confidence. We refer to this phenomenon as indiscriminate miscalibration. We found that traditional calibration metrics, such as Expected Calibrated Errors (ECEs), are unable to capture this behavior effectively. To address this issue, we propose new metrics to measure the severity of indiscriminate miscalibration. Additionally, we develop a novel in-context comparative inference method to alleviate miscalibrations and improve classification performance. Through extensive experiments on five datasets, we demonstrate that our proposed method can achieve more accurate and calibrated predictions compared to regular zero-shot and few-shot prompting.
format Preprint
id arxiv_https___arxiv_org_abs_2410_02210
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Calibrate to Discriminate: Improve In-Context Learning with Label-Free Comparative Inference
Cheng, Wei
Wang, Tianlu
Ji, Yanmin
Yang, Fan
Tan, Keren
Zheng, Yiyu
Computation and Language
Machine Learning
While in-context learning with large language models (LLMs) has shown impressive performance, we have discovered a unique miscalibration behavior where both correct and incorrect predictions are assigned the same level of confidence. We refer to this phenomenon as indiscriminate miscalibration. We found that traditional calibration metrics, such as Expected Calibrated Errors (ECEs), are unable to capture this behavior effectively. To address this issue, we propose new metrics to measure the severity of indiscriminate miscalibration. Additionally, we develop a novel in-context comparative inference method to alleviate miscalibrations and improve classification performance. Through extensive experiments on five datasets, we demonstrate that our proposed method can achieve more accurate and calibrated predictions compared to regular zero-shot and few-shot prompting.
title Calibrate to Discriminate: Improve In-Context Learning with Label-Free Comparative Inference
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2410.02210