MortalMATH: Evaluating the Conflict Between Reasoning Objectives and Emergency Contexts

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Lanzeray, Etienne, Meilliez, Stephane, Ruelle, Malo, Sileo, Damien
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910001204822016
author Lanzeray, Etienne
Meilliez, Stephane
Ruelle, Malo
Sileo, Damien
author_facet Lanzeray, Etienne
Meilliez, Stephane
Ruelle, Malo
Sileo, Damien
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2601_18790
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle MortalMATH: Evaluating the Conflict Between Reasoning Objectives and Emergency Contexts
Lanzeray, Etienne
Meilliez, Stephane
Ruelle, Malo
Sileo, Damien
Computation and Language
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.
title MortalMATH: Evaluating the Conflict Between Reasoning Objectives and Emergency Contexts
topic Computation and Language
url https://arxiv.org/abs/2601.18790