When and How Unlabeled Data Provably Improve In-Context Learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Li, Yingcong, Chang, Xiangyu, Kara, Muti, Liu, Xiaofeng, Roy-Chowdhury, Amit, Oymak, Samet
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918303139627008
author Li, Yingcong
Chang, Xiangyu
Kara, Muti
Liu, Xiaofeng
Roy-Chowdhury, Amit
Oymak, Samet
author_facet Li, Yingcong
Chang, Xiangyu
Kara, Muti
Liu, Xiaofeng
Roy-Chowdhury, Amit
Oymak, Samet
contents Recent research shows that in-context learning (ICL) can be effective even when demonstrations have missing or incorrect labels. To shed light on this capability, we examine a canonical setting where the demonstrations are drawn according to a binary Gaussian mixture model (GMM) and a certain fraction of the demonstrations have missing labels. We provide a comprehensive theoretical study to show that: (1) The loss landscape of one-layer linear attention models recover the optimal fully-supervised estimator but completely fail to exploit unlabeled data; (2) In contrast, multilayer or looped transformers can effectively leverage unlabeled data by implicitly constructing estimators of the form $\sum_{i\ge 0} a_i (X^\top X)^iX^\top y$ with $X$ and $y$ denoting features and partially-observed labels (with missing entries set to zero). We characterize the class of polynomials that can be expressed as a function of depth and draw connections to Expectation Maximization, an iterative pseudo-labeling algorithm commonly used in semi-supervised learning. Importantly, the leading polynomial power is exponential in depth, so mild amount of depth/looping suffices. As an application of theory, we propose looping off-the-shelf tabular foundation models to enhance their semi-supervision capabilities. Extensive evaluations on real-world datasets show that our method significantly improves the semisupervised tabular learning performance over the standard single pass inference.
format Preprint
id arxiv_https___arxiv_org_abs_2506_15329
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle When and How Unlabeled Data Provably Improve In-Context Learning
Li, Yingcong
Chang, Xiangyu
Kara, Muti
Liu, Xiaofeng
Roy-Chowdhury, Amit
Oymak, Samet
Machine Learning
Artificial Intelligence
Computation and Language
Optimization and Control
Recent research shows that in-context learning (ICL) can be effective even when demonstrations have missing or incorrect labels. To shed light on this capability, we examine a canonical setting where the demonstrations are drawn according to a binary Gaussian mixture model (GMM) and a certain fraction of the demonstrations have missing labels. We provide a comprehensive theoretical study to show that: (1) The loss landscape of one-layer linear attention models recover the optimal fully-supervised estimator but completely fail to exploit unlabeled data; (2) In contrast, multilayer or looped transformers can effectively leverage unlabeled data by implicitly constructing estimators of the form $\sum_{i\ge 0} a_i (X^\top X)^iX^\top y$ with $X$ and $y$ denoting features and partially-observed labels (with missing entries set to zero). We characterize the class of polynomials that can be expressed as a function of depth and draw connections to Expectation Maximization, an iterative pseudo-labeling algorithm commonly used in semi-supervised learning. Importantly, the leading polynomial power is exponential in depth, so mild amount of depth/looping suffices. As an application of theory, we propose looping off-the-shelf tabular foundation models to enhance their semi-supervision capabilities. Extensive evaluations on real-world datasets show that our method significantly improves the semisupervised tabular learning performance over the standard single pass inference.
title When and How Unlabeled Data Provably Improve In-Context Learning
topic Machine Learning
Artificial Intelligence
Computation and Language
Optimization and Control
url https://arxiv.org/abs/2506.15329