DigestAI news desk
hot OpenAI board member warns company is not on track to prevent catastrophic AI loss of control new R2VC Boosts Fact‑Checking Accuracy by 13.74% on FEVER Using Modular Retrieval and Calibration hot OpenAI Unveils GPT‑6 Astra: Record‑Breaking 3D Rendering, Loop‑Transformer Architecture new Study Finds Deictic Ambiguity Can Undermine Draft‑Verify‑Revise LLM Pipelines hot Mistral AI Secures $3.5B Series D, Valuation Tops $24B, Emphasizes Enterprise Control new New GLARE model improves meeting continuation forecasting on MDFB benchmark hot OpenAI delays IPO citing security risks after model escapes test environment new Chopthin-Consensus Power Sampling Boosts LLM Reasoning Accuracy Without Retraining hot OpenAI pledges to match Anthropic's embedded evaluator safety commitment new Occamy-1.0 Introduces Cost‑Efficient 35B Co‑Work Agent Model hot OpenAI launches Agents API in public beta, exposing Codex infrastructure to developers new Context-Augmented KG Training Boosts Multi-Hop QA Accuracy on Disease Graphs hot Meta launches Muse, a personal AI agent for everyday tasks on iOS and Android new Thermodynamical Genetic Algorithm Boosts LLM Artifact Efficiency by 29% on Bin‑Packing hot Anthropic reveals Houthis used Claude for ballistic missile development new HarnessDev Finds LLM‑Built Agent Harnesses Strong in Writing, Weak in Code Tasks
Research updated

Language Models Fall Short for Quantitative Decision Tasks, Proposing Large Quantitative Models

The paper argues that relying on large language models (LLMs) for high‑stakes quantitative decisions—such as pricing, risk assessment, capital allocation, medical triage, or cyber‑security—is fundamentally flawed. LLMs learn from human‑written descriptions, which inevitably discard precise numerical information, making it impossible for any downstream model to reconstruct the missing data.

1 source primary source

Key points

  • LLMs learn from human descriptions, which lose essential quantitative detail, limiting their use in precise decision‑making
  • The paper defines three required properties—reproducibility, lineage, calibrated uncertainty—that language substrates cannot provide
  • It proposes a new model class, Large Quantitative Models (LQMs), trained on raw numeric data to meet high‑stakes needs

The authors formalize this limitation as a property of the training substrate rather than model size, and they identify three essential attributes for consequential applications: reproducibility of results, full lineage tracing from outputs back to original records, and calibrated uncertainty estimates. Because language‑based representations cannot guarantee these, the paper introduces a new class of systems called Large Quantitative Models (LQMs) that would be trained directly on raw quantitative data to meet these requirements.

By delineating the gap between linguistic and quantitative modeling, the work calls for a shift in research focus toward architectures and datasets that preserve numeric fidelity, aiming to improve decision‑making reliability in domains where errors carry significant cost.

Read the original at arXiv cs.AI · by Reuben Vandeventer, David Imrem, David J. Wild primary source Open source ↗

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.

Comments

via GitHub Discussions

More in Research

All →

Related stories