VOOZH about

URL: https://arxiv.org/abs/2305.18654

⇱ [2305.18654] Faith and Fate: Limits of Transformers on Compositionality

Computer Science > Computation and Language

arXiv:2305.18654 (cs)
[Submitted on 29 May 2023 (v1), last revised 31 Oct 2023 (this version, v3)]

Title:Faith and Fate: Limits of Transformers on Compositionality

View PDF
Abstract:Transformer large language models (LLMs) have sparked admiration for their exceptional performance on tasks that demand intricate multi-step reasoning. Yet, these models simultaneously show failures on surprisingly trivial problems. This begs the question: Are these errors incidental, or do they signal more substantial limitations? In an attempt to demystify transformer LLMs, we investigate the limits of these models across three representative compositional tasks -- multi-digit multiplication, logic grid puzzles, and a classic dynamic programming problem. These tasks require breaking problems down into sub-steps and synthesizing these steps into a precise answer. We formulate compositional tasks as computation graphs to systematically quantify the level of complexity, and break down reasoning steps into intermediate sub-procedures. Our empirical findings suggest that transformer LLMs solve compositional tasks by reducing multi-step compositional reasoning into linearized subgraph matching, without necessarily developing systematic problem-solving skills. To round off our empirical study, we provide theoretical arguments on abstract multi-step reasoning problems that highlight how autoregressive generations' performance can rapidly decay with\,increased\,task\,complexity.
Comments: 10 pages + appendix (40 pages)
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Cite as: arXiv:2305.18654 [cs.CL]
(or arXiv:2305.18654v3 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2305.18654
arXiv-issued DOI via DataCite

Submission history

From: Nouha Dziri [view email]
[v1] Mon, 29 May 2023 23:24:14 UTC (4,815 KB)
[v2] Thu, 1 Jun 2023 20:50:32 UTC (5,660 KB)
[v3] Tue, 31 Oct 2023 16:35:07 UTC (5,117 KB)
Full-text links:

Access Paper:

  • View PDF
  • TeX Source

Current browse context:

cs.CL
< prev   |   next >
Change to browse by:

References & Citations

BibTeX formatted citation

Data provided by:

Bibliographic and Citation Tools

Bibliographic Explorer (What is the Explorer?)
Connected Papers (What is Connected Papers?)
Litmaps (What is Litmaps?)
scite Smart Citations (What are Smart Citations?)

Code, Data and Media Associated with this Article

alphaXiv (What is alphaXiv?)
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub (What is DagsHub?)
Gotit.pub (What is GotitPub?)
Hugging Face (What is Huggingface?)
ScienceCast (What is ScienceCast?)

Demos

Replicate (What is Replicate?)
Hugging Face Spaces (What is Spaces?)
TXYZ.AI (What is TXYZ.AI?)

Recommenders and Search Tools

Influence Flower (What are Influence Flowers?)
CORE Recommender (What is CORE?)

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.