Language models are inadequate for complex quantitative decisions

Language Is an Insufficient Substrate for Quantitative Reasoning, and Consequential Domains Need Large Quantitative Models

Artificial IntelligenceMachine Learning

Summary

Using language models to make important decisions based on numbers, like pricing or medical triage, is limited because language can’t fully capture precise data. The authors explain that when people describe data in words, some details are lost forever, so no computer model trained on those words can perfectly recover the original figures. They also identify three key needs for trusted quantitative decision-making: consistent results, traceability back to original data, and meaningful measures of uncertainty. Because language models can’t provide these by design, the authors say a new type of model focused on quantitative data is needed.

What this means in practice

  • For financial risk teams: Develop models that ensure precise, auditable calculations of risk directly from numerical data rather than text descriptions.
  • For clinical decision support teams: Design systems that trace every medical recommendation back to original patient records with clear confidence levels.

A position paper. It proposes an approach and reports no results.

Authors

Reuben Vandeventer, David Imrem, David J. Wild

Abstract

The prevailing assumption in applied machine learning is that progress on consequential quantitative decisions such as pricing risk, allocating capital, triaging patients, or containing a network intrusion will follow from progress in large language models (LLMs). A language model is trained on a representation of the world that was produced by human description; description is a lossy encoding of the quantitative record, and the loss is irreversible: no downstream model, at any scale, can recover from a description what the description did not encode. We formalize this as a property of the representation on which a model is trained rather than of the model capacity, and we identify three further properties that consequential settings demand of a model and that a language substrate cannot supply by construction: reproducibility, lineage from every output back to the source records that produced. it, and calibrated uncertainty. We argue that these properties define a distinct model class, which we call the Large Quantitative Model (LQM).
View PDFOpen arXiv