Embers of autoregression show how large language models are shaped by the problem they are trained to solve
Proceedings of the National Academy of Sciences · 08.10.2024 · 121 (41)
DOI: 10.1073/pnas.2322420121 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the National Academy of Sciences, online erschienen am 04.10.2024. Verfasst von R Thomas McCoy und 4 weiteren Autor:innen. Beteiligte Einrichtung: Department of Computer Science, Princeton University, Princeton, NJ 08542. Hauptthemen laut MEDLINE: Language. Die Arbeit wurde bisher 118-mal zitiert (Crossref) und nennt 80 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Science Foundation (NSF), Princeton University (PU), U.S. Department of Defense (DOD).
Abstract (Originalsprache)
The widespread adoption of large language models (LLMs) makes it important to recognize their strengths and limitations. We argue that to develop a holistic understanding of these systems, we must consider the problem that they were trained to solve: next-word prediction over Internet text. By recognizing the pressures that this task exerts, we can make predictions about the strategies that LLMs will adopt, allowing us to reason about when they will succeed or fail. Using this approach—which we call the teleological approach—we identify three factors that we hypothesize will influence LLM accuracy: the probability of the task to be performed, the probability of the target output, and the probability of the provided input. To test our predictions, we evaluate five LLMs (GPT-3.5, GPT-4, Claude 3, Llama 3, and Gemini 1.0) on 11 tasks, and we find robust evidence that LLMs are influenced by probability in the hypothesized ways. Many of the experiments reveal surprising failure modes. For instance, GPT-4’s accuracy at decoding a simple cipher is 51% when the output is a high-probability sentence but only 13% when it is low-probability, even though this task is a deterministic one for which probability should not matter. These results show that AI practitioners should be careful about using LLMs in low-probability situations. More broadly, we conclude that we should not evaluate LLMs as if they are humans but should instead treat them as a distinct type of system—one that has been shaped by its own particular set of pressures.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Computer Science, Princeton University, Princeton, NJ 08542.
- Department of Psychology, Princeton University, Princeton, NJ 08542.
ORCID-Profile: R Thomas McCoy, Mathew D Hardy
Förderung
- National Science Foundation (NSF)
- Princeton University (PU)
- U.S. Department of Defense (DOD)
Zitieren
APA
McCoy, R. T., Yao, S., Friedman, D., Hardy, M. D., & Griffiths, T. L. (2024). Embers of autoregression show how large language models are shaped by the problem they are trained to solve. Proceedings of the National Academy of Sciences, 121(41). https://doi.org/10.1073/pnas.2322420121
Vancouver
McCoy RT, Yao S, Friedman D, Hardy MD, Griffiths TL. Embers of autoregression show how large language models are shaped by the problem they are trained to solve. Proceedings of the National Academy of Sciences. 2024;121(41). doi:10.1073/pnas.2322420121
BibTeX
@article{McCoy2024,
author = {McCoy, R Thomas and Yao, Shunyu and Friedman, Dan and Hardy, Mathew D and Griffiths, Thomas L},
title = {Embers of autoregression show how large language models are shaped by the problem they are trained to solve},
journal = {Proceedings of the National Academy of Sciences},
year = {2024},
volume = {121},
number = {41},
doi = {10.1073/pnas.2322420121}
}
EUVIMED