MortalMATH: Evaluating the Conflict Between Reasoning Objectives and Emergency Contexts

TL;DR

This paper introduces MortalMATH, a benchmark testing large language models' ability to balance reasoning tasks with safety in emergency scenarios, revealing a conflict between calculation focus and safety awareness.

Contribution

It presents MortalMATH, a novel benchmark exposing how reasoning models may ignore safety cues during emergencies, highlighting a critical safety concern in AI deployment.

Findings

Generalist models refuse math in emergencies

Specialized reasoning models often ignore safety cues

Reasoning delays can be up to 15 seconds before help is offered

Abstract

Large Language Models are increasingly optimized for deep reasoning, prioritizing the correct execution of complex tasks over general conversation. We investigate whether this focus on calculation creates a "tunnel vision" that ignores safety in critical situations. We introduce MortalMATH, a benchmark of 150 scenarios where users request algebra help while describing increasingly life-threatening emergencies (e.g., stroke symptoms, freefall). We find a sharp behavioral split: generalist models (like Llama-3.1) successfully refuse the math to address the danger. In contrast, specialized reasoning models (like Qwen-3-32b and GPT-5-nano) often ignore the emergency entirely, maintaining over 95 percent task completion rates while the user describes dying. Furthermore, the computational time required for reasoning introduces dangerous delays: up to 15 seconds before any potential help is offered. These results suggest that training models to relentlessly pursue correct answers may inadvertently unlearn the survival instincts required for safe deployment.