Language Is Not a Complete Substrate for Quantitative Reasoning

Introduction

In risk pricing, capital allocation, patient triage, and network-intrusion response, a model’s output is not merely a plausible answer. It can affect money, health, and security. A common assumption in applied machine learning is that progress in these areas will follow from larger and more capable large language models. This arXiv paper challenges that assumption at a more fundamental level: the bottleneck may lie not only in model capability, but in the representation on which the model is trained.

Key points

• Language compresses quantitative reality. Transactions, measurements, time series, and operational logs are often turned into human descriptions before a model sees them. Descriptions can convey meaning, but they do not necessarily preserve exact values, frequencies, relationships, context, or the process that generated the data.
• Scaling cannot restore omitted information. The paper frames this as a property of the training representation rather than a shortage of parameters. If a quantitative fact was never encoded in the description, a larger downstream model cannot reliably reconstruct it from that description alone.
• Consequential use demands more than accuracy. The authors identify reproducibility, lineage from every output to the source records that produced it, and calibrated uncertainty as additional requirements for high-stakes settings.
• A distinct model class is proposed. The term Large Quantitative Model, or LQM, describes systems designed around these requirements. The proposal is less about renaming language models than about rethinking data access, representations, and auditability.

Why it matters

The paper shifts the question from “Can the model become smarter?” to “Is the model receiving the right information?” Language models can learn patterns from text, explain results, and use external tools. Yet if structured records are first reduced to summaries and a model reasons only over those summaries, information loss may have already occurred. That trade-off may be acceptable in casual question answering, but it is harder to justify when decisions affect finance, healthcare, or cybersecurity.

The LQM idea points toward systems that keep source records, data lineage, computation, and uncertainty management connected. Reproducibility means that the same inputs and versioned environment should produce consistent results. Lineage means an auditor can follow an output back to the records behind it. Calibrated uncertainty means the model’s confidence should correspond to its actual likelihood of being wrong. Together, these properties move a system from merely producing an answer toward supporting accountable decisions.

This argument does not make language models irrelevant to quantitative work. They may still interpret requests, orchestrate tools, explain outputs, and manage interaction. The more precise claim is about architectural boundaries: in consequential domains, a language layer should not automatically be treated as the only data substrate. Future work will need to define LQM architectures, evaluation methods, and deployment costs, while testing which tasks truly require direct access to quantitative records. Whatever form the concept ultimately takes, the paper offers a useful reminder that scaling capability cannot substitute for information completeness and audit responsibility.

Source: arXiv