Red Teaming for Large Language Models At Scale: Tackling Hallucinations on Mathematics Tasks

TL;DR

The paper addresses the hallucination risk of large language models on elementary mathematics by introducing a scalable red-teaming framework that procedurally generates questions and tests prompting techniques. Using two GPT models (GPT-4 and GPT-3.5-turbo), it shows that while structure-aware prompts and worked examples can modestly improve accuracy, both models remain ill-suited for elementary calculations, with improvements highly dependent on task and prompt context. The study provides detailed experiments (elementary math and algebraic reasoning) and an evaluation scheme focused on final answers rather than intermediate reasoning, highlighting practical implications for AI-assisted education and safety. Overall, the work offers a reproducible methodology for probing numerical reasoning in LLMs and discusses the limits of current prompting strategies at scale.

Abstract

We consider the problem of red teaming LLMs on elementary calculations and algebraic tasks to evaluate how various prompting techniques affect the quality of outputs. We present a framework to procedurally generate numerical questions and puzzles, and compare the results with and without the application of several red teaming techniques. Our findings suggest that even though structured reasoning and providing worked-out examples slow down the deterioration of the quality of answers, the gpt-3.5-turbo and gpt-4 models are not well suited for elementary calculations and reasoning tasks, also when being red teamed.