Inference Scaling Laws for LLMs Bigger is not always better
Year: 2025
Type: article
Abstract: This work addresses a fundamental question: if LLMs are unreliable at mathematics, how can they still be useful for mathematical tasks? The answer lies in reframing their role—from calculator to code generator. Large Language Models are unreliable at direct mathematical computation—they hallucinate answers, fail multi-step reasoning, and cannot verify their own work. However, LLMs can serve a different role: generating code for verified computational engines . This work evaluates the “LLM as code generator” paradigm, where models translate natural language into Wolfram Mathematica syntax that a computer algebra system executes with guaranteed correctness.
Author: White, Melvyn
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7164825151