Why My T-1000 Needs to Borrow My TI-81: The Calculator Paradox in Large Language Models
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Large language models do not compute mathematical results; they predict what mathematical results look like. This paper identifies a failure mode we term the Calculator Paradox: LLMs fail most dangerously not on overtly difficult mathematical problems but on problems that appear non-mathematical yet require genuine computation—temporal calculations, implicit comparisons, and linguistically simple quantitative assertions where neither party expects failure. We present evidence from a live research session and a controlled demonstration experiment, including a phrasing ablation that isolates perceived difficulty as a causal variable. We introduce a two-component measure of implicit computation density—C₁ (operation density) and C₂ (decision-boundary sensitivity)—to characterize the hidden quantitative load carried by natural-language assertions and its downstream brittleness. We argue that the critical vulnerability is the model’s inability to distinguish arithmetic it can perform from arithmetic it merely pattern-matches: the threshold judgment itself constitutes the bug. We propose a default-to-verification architecture with progressive verification tiers, inverting the standard tool-invocation paradigm. We distinguish the Calculator Paradox from hallucination as a separate failure type and situate it within a unified framework of metacognitive collapse in autoregressive systems, encompassing the author’s prior analyses of Novelty Blindness, the Temporal Gap in LLM Memory Systems, and the AI Depth Asymmetry.Keywords: large language models, arithmetic reasoning, tool use, computational co-processors, calculator paradox, token prediction, threshold judgment, implicit computation density, metacognitive collapse, default-to-verification
Fields of Study & Subjects
Natural Language Processing Techniques
Topic Modeling
Artificial Intelligence Applications
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 28 March 2026
- Language
- EN