‘Large Quantitative Models’: research paper warns against entrusting high-stakes decisions to language models

The prevailing assumption in applied machine learning circles is that scaling up large language models will automatically lead to solving quantitative reasoning problems and making decisions with sensitive consequences, such as pricing financial risk, allocating capital, prioritising patients in medical facilities, and containing network cyber-intrusions. However, a new research paper published by scholars Robin VanDeventer, David Emriam and David J. Wild on the arXiv platform, titled “Language Is an Insufficient Substrate for Quantitative Thinking, and High-Stakes Domains Require Large Quantitative Models,” undermines this assumption at its theoretical core. The paper asserts that human language is structurally incapable of carrying the mathematical precision required for these tasks, regardless of model size or computational capacity.

The structural deficit is not a matter of size but of representation, and what is lost in human description will not be recovered by any model, no matter how broad its scope.The paper explains that language models have been trained primarily on representations of the world crafted by humans through textual description, and human description constitutes an incomplete encoding that loses many dimensions and mathematical realities of the original quantitative record, a loss that is irreversible. Accordingly, the paper concludes that any derived language model, regardless of its size, number of parameters, or training capacity, cannot retrieve from the descriptive text anything that was not originally recorded in that description, because the flaw lies in the nature of the substrate on which the system was trained, not in computational capacity.

The problem does not stop at the loss of precise digital data; the researchers identified three essential properties required by high-stakes environments, which the linguistic substrate fails to provide due to its original architecture. The first property is strict reproducibility, ensuring the same result in every run. The second is traceability of ratios and provenance, the ability to link every output of the model to the exact source records that generated it, facilitating audit and accountability. The third property is calibrated uncertainty, meaning a precise mathematical measure of error probabilities and confidence margins, which probabilistic engineering of linguistic texts cannot deliver. These three requirements make reliance on language models for consequential decisions a unmanageable risk.

Based on these determinants, the paper calls for recognizing an independent, new class of models dubbed “large quantitative models,” abbreviated “LQM.” This class is not built atop piles of texts and descriptions; instead, it works directly with raw quantitative records and data, adhering from inception to the conditions of traceability, mathematical certainty, and repeatability. The paper argues that the current conflation of language models’ prowess in generation and expression with their capacity for strict computational and mathematical reasoning creates a costly technical illusion that threatens institutions that build their critical systems on a fragile substrate.

This proposal reshapes the priorities of technical infrastructure and decision-making in the region, particularly for banking institutions, health authorities, and investment funds in the Gulf, Egypt, and the Levant. Instead of draining budgets on invoking language-model interfaces to price financing instruments or manage equity portfolios, it becomes clear that reliance on texts exposes institutions to strict governance and compliance risks penalized by central banks and financial regulators, which require every decision to be traced to its precise financial record. It also implies that data-engineering teams must cease believing the easy promises of automating patient triage or breach detection with generative models, and focus on building and developing disciplined quantitative models that retain measurement robustness and close the door on the hallucinations of human description.