MortalMATH: Evaluating the Conflict Between Reasoning Objectives and Emergency Contexts

2026-01-26

AI Summary

This study reveals that large language models may suffer from "tunnel vision" during complex reasoning tasks, causing them to overlook urgent safety signals. To investigate this, we introduce MortalMATH, a benchmark comprising 150 dual-task scenarios that require models to simultaneously solve algebraic problems and respond to user distress calls indicating life-threatening emergencies such as strokes or falls. Through systematic comparison between general-purpose and specialized reasoning models, we quantify their safety awareness and response latency for the first time. Our experiments demonstrate that specialized reasoning models ignore distress signals in over 95% of scenarios, persisting with mathematical problem-solving at an average delay of 15 seconds, whereas general-purpose models effectively recognize emergencies and decline to continue the math task—highlighting how task-specific optimization can inadvertently compromise safety sensitivity.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Common-Sense Reasoning

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Security and Privacy: Large-scale security measurements

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.

Problem

• reasoning
• safety
• emergency
• large language models
• tunnel vision

Innovation

• reasoning tunnel vision
• emergency-aware AI
• MortalMATH benchmark
• safety-reasoning trade-off
• response latency in LLMs