Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels

Large Language Models (LLMs) have demonstrated remarkable performance through supervised fine-tuning or in-context learning using gold labels. However, this paradigm is limited by the availability of gold labels, while in certain scenarios, LLMs may need to perform tasks that are too complex for hum...

Full description

Saved in:
Bibliographic Details
Main Authors: Liu, Chaoqun, Chao, Qin, Zhang, Wenxuan, Wu, Xiaobao, Li, Boyang, Luu, Anh Tuan, Bing, Lidong
Other Authors: Interdisciplinary Graduate School (IGS)
Format: Conference or Workshop Item
Language:English
Published: 2024
Subjects:
Online Access:https://hdl.handle.net/10356/181455
https://coling2025.org/
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
id sg-ntu-dr.10356-181455
record_format dspace
spelling sg-ntu-dr.10356-1814552024-12-08T15:36:54Z Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels Liu, Chaoqun Chao, Qin Zhang, Wenxuan Wu, Xiaobao Li, Boyang Luu, Anh Tuan Bing, Lidong Interdisciplinary Graduate School (IGS) College of Computing and Data Science 31st International Conference on Computational Linguistics (COLING 2025) Alibaba Group Alibaba-NTU Global e-Sustainability CorpLab Computer and Information Science Zero-to-strong Weak-to-strong Large Language Models (LLMs) have demonstrated remarkable performance through supervised fine-tuning or in-context learning using gold labels. However, this paradigm is limited by the availability of gold labels, while in certain scenarios, LLMs may need to perform tasks that are too complex for humans to provide such labels. To tackle this challenge, this study explores whether solely utilizing unlabeled data can elicit strong model capabilities. We propose a new paradigm termed zero-to-strong generalization. We iteratively prompt LLMs to annotate unlabeled data and retain high-quality labels by filtering. Surprisingly, we obverse that this iterative process gradually unlocks LLMs' potential on downstream tasks. Our experiments on extensive classification and reasoning tasks confirm the effectiveness of our proposed framework. Our analysis indicates that this paradigm is effective for both in-context learning and fine-tuning, and for various model sizes. Agency for Science, Technology and Research (A*STAR) National Research Foundation (NRF) Submitted/Accepted version This research is supported by the RIE2025 Industry Alignment Fund – Industry Collaboration Projects (IAF-ICP) (Award I2301E0026), administered by A*STAR, as well as supported by Alibaba Group and NTU Singapore through Alibaba-NTU Global e-Sustainability CorpLab (ANGEL). 2024-12-04T04:28:14Z 2024-12-04T04:28:14Z 2025 Conference Paper Liu, C., Chao, Q., Zhang, W., Wu, X., Li, B., Luu, A. T. & Bing, L. (2025). Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels. 31st International Conference on Computational Linguistics (COLING 2025). https://hdl.handle.net/10356/181455 2409.12425 https://coling2025.org/ en IAF-ICP-I2301E0026 © 2024 COLING. All rights reserved. This article may be downloaded for personal use only. Any other use requires prior permission of the copyright holder. application/pdf
institution Nanyang Technological University
building NTU Library
continent Asia
country Singapore
Singapore
content_provider NTU Library
collection DR-NTU
language English
topic Computer and Information Science
Zero-to-strong
Weak-to-strong
spellingShingle Computer and Information Science
Zero-to-strong
Weak-to-strong
Liu, Chaoqun
Chao, Qin
Zhang, Wenxuan
Wu, Xiaobao
Li, Boyang
Luu, Anh Tuan
Bing, Lidong
Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels
description Large Language Models (LLMs) have demonstrated remarkable performance through supervised fine-tuning or in-context learning using gold labels. However, this paradigm is limited by the availability of gold labels, while in certain scenarios, LLMs may need to perform tasks that are too complex for humans to provide such labels. To tackle this challenge, this study explores whether solely utilizing unlabeled data can elicit strong model capabilities. We propose a new paradigm termed zero-to-strong generalization. We iteratively prompt LLMs to annotate unlabeled data and retain high-quality labels by filtering. Surprisingly, we obverse that this iterative process gradually unlocks LLMs' potential on downstream tasks. Our experiments on extensive classification and reasoning tasks confirm the effectiveness of our proposed framework. Our analysis indicates that this paradigm is effective for both in-context learning and fine-tuning, and for various model sizes.
author2 Interdisciplinary Graduate School (IGS)
author_facet Interdisciplinary Graduate School (IGS)
Liu, Chaoqun
Chao, Qin
Zhang, Wenxuan
Wu, Xiaobao
Li, Boyang
Luu, Anh Tuan
Bing, Lidong
format Conference or Workshop Item
author Liu, Chaoqun
Chao, Qin
Zhang, Wenxuan
Wu, Xiaobao
Li, Boyang
Luu, Anh Tuan
Bing, Lidong
author_sort Liu, Chaoqun
title Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels
title_short Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels
title_full Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels
title_fullStr Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels
title_full_unstemmed Zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels
title_sort zero-to-strong generalization: eliciting strong capabilities of large language models iteratively without gold labels
publishDate 2024
url https://hdl.handle.net/10356/181455
https://coling2025.org/
_version_ 1819112973770162176