Can General-Purpose Language Models Emulate a General-Purpose Computer In-Context?

Vasilis Papageorgiou, Jy-yong Sohn, Kangwook Lee, Dimitris Papailiopoulos

Keywords: large language models, general-purpose computing, AI alignment

TL;DR: We investigate whether it is possible for looped pretrained LLMs to in-context emulate a general-purpose computer, without the use of external mechanisms (such as memory or interpreters)?

Abstract: Several recent works have drawn parallels between modern Large Language Models (LLMs) and general-purpose computers, suggesting that language serves as their programming interface. In this study, we test part of this analogy; specifically, we investigate whether a pretrained LLM can emulate a memory-bounded, reduced instruction-set based computer by executing random programs through looped inference calls. All this within the model's own context window, and without the aid of external mechanisms such as associative memory or interpreters. The abstraction level of these programs is based on two general-purpose computational models - the SUBLEQ One-Instruction Set Computer (OISC) and the Minsky counter machine. Our prompts are carefully designed in a data-agnostic manner, and we conduct studies to examine failure modes related to the emulated computer functionality. Our findings indicate that certain models are capable of efficiently executing general-purpose instructions, despite not being explicitly trained for such a task. This suggests intriguing implications for AI alignment, as some models demonstrate the ability to \emph{autonomously} emulate the operation of a general-purpose computer.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9120