From Language to Pixels: Task Recognition and Task Learning in LLMs

Janek Falkenstein, Carolin Schuster, Alex Berger, Georg Groh

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Large language models (LLMs) can perform unseen tasks by learning from a few in-context examples. How in-context learning works is still uncertain. We investigate the mechanisms of in-context learning on a challenging non-language task. The task requires the LLM to generate pixel matrices representing images of basic shapes. We introduce a framework to analyze if this task is solved by recognizing similar formats from the training data (task recognition) or by understanding the instructions and learning the skill de novo during inference (task learning). Our experiments demonstrate that LLMs generate meaningful pixel matrices with task recognition and fail to learn such tasks when encountering unfamiliar formats. Our findings offer insights into LLMs’ learning mechanisms to guide future research on their seemingly human-like behavior.

Original languageEnglish
Title of host publicationGenBench 2024 - GenBench
Subtitle of host publication2nd Workshop on Generalisation (Benchmarking) in NLP, Proceedings of the Workshop
EditorsDieuwke Hupkes, Verna Dankers, Khuyagbaatar Batsuren, Amirhossein Kazemnejad, Christos Christodoulopoulos, Mario Giulianelli, Ryan Cotterell
PublisherAssociation for Computational Linguistics (ACL)
Pages27-41
Number of pages15
ISBN (Electronic)9798891761827
StatePublished - 2024
Event2nd Workshop on Generalisation (Benchmarking) in NLP, GenBench 2024 - Miami, United States
Duration: 16 Nov 2024 → …

Publication series

NameGenBench 2024 - GenBench: 2nd Workshop on Generalisation (Benchmarking) in NLP, Proceedings of the Workshop

Conference

Conference2nd Workshop on Generalisation (Benchmarking) in NLP, GenBench 2024
Country/TerritoryUnited States
CityMiami
Period16/11/24 → …

Fingerprint

Dive into the research topics of 'From Language to Pixels: Task Recognition and Task Learning in LLMs'. Together they form a unique fingerprint.

Cite this