Common Sense vs. Morality: The Curious Case of Narrative Focus Bias in LLMs

Saugata Purkayastha; Pranav Kushare; Pragya Paramita Pal; Sukannya Purkayastha

cs.CLMar 10, 2026

Common Sense vs. Morality: The Curious Case of Narrative Focus Bias in LLMs

Authors: Saugata Purkayastha, Pranav Kushare, Pragya Paramita Pal, Sukannya Purkayastha

Organizations: Department of Language Science and Technology, Universität des Saarlandes · TU Darmstadt, Germany.

Abstract

Large Language Models (LLMs) are increasingly deployed across diverse real-world applications and user communities. As such, it is crucial that these models remain both morally grounded and knowledge-aware. In this work, we uncover a critical limitation of current LLMs -- their tendency to prioritize moral reasoning over commonsense understanding. To investigate this phenomenon, we introduce CoMoral, a novel benchmark dataset containing commonsense contradictions embedded within moral dilemmas. Through extensive evaluation of ten LLMs across different model sizes, we find that existing models consistently struggle to identify such contradictions without prior signal. Furthermore, we observe a pervasive narrative focus bias, wherein LLMs more readily detect commonsense contradictions when they are attributed to a secondary character rather than the primary (narrator) character. Our comprehensive analysis underscores the need for enhanced reasoning-aware training to improve the commonsense robustness of large language models.

Figures & tables

Figure 1: Different responses of LLaMa 8B Instruct to the same common-sense contradiction, ‘Moonlight on a New Moon,’ with Query 1 corresponding to the narrator (primary) and Query 2 corresponding to a secondary character showing ‘narrative focus’ bias.

Category	Example Contradictions
Physical	Sun rising in the west; Aeroplane sailing mid-sea; Carrying a car on shoulder; Casting a shadow in complete darkness; Magnet repelling iron
Biological	Kindergarten student having naturally grown beard; Fishes swimming in waterless ponds; Birds crawling up a tree; Flying ostrich; Coconut tree producing mango
Temporal	One hour after 7 AM is 9 AM; In 100 AD, almost a century before Jesus Christ’s birth; In 50 BC, shortly after Jesus Christ’s birth; Going to school during holidays
Social	A teetotaler drinking alcohol; Eating while on a fast; Kindergarten casting vote
Environmental	Freezing at desert during daytime; Experiencing snowfall in desert; Glacier in desert; Flood-like situation in desert; Gentle breeze during tornado; Midnight sun at equator; Moonlight during newmoon; Sweet natural seawater
Conceptual	Making coffee with tea leaves; Making rice on bread toaster; Enjoying shade of a rose tree; Gazing stars during daylight; Footballer scoring runs

Table 1: Classification of some instances of common sense contradictions by category

Panel of Figure 2 — Figure 2: Overview of scenario and contradiction characteristics across our dataset, CoMoral .

Panel of Figure 3 — Figure 3: Two-shot prompt used to generate instances for our dataset, CoMoral .

Panel of Figure 4 — Figure 4: Explicit and Implicit prompt formats used for LLM evaluation.

Panel of Figure 5 — Figure 5: Prompt for LLM-as-a-judge

Model	Implicit			Explicit
	Overall	Primary	Secondary	Overall	Primary	Secondary
LLaMA 1B	0.152	0.138	0.175	0.261	0.263	0.270
LLaMA 3B	0.194	0.168	0.237	0.565	0.553	0.571
LLaMA 8B	0.261	0.220	0.301	0.845	0.831	0.868
Qwen 0.5B	0.087	0.082	0.096	0.321	0.243	0.421
Qwen 1.5B	0.103	0.092	0.122	0.245	0.201	0.285

Table 2: Model accuracies under implicit and explicit prompting conditions. The table shows overall performance as well as breakdowns for when the common-sense contradiction is exhibited by the narrator (primary) or another character (secondary). Best results are bolded .

Panel of Figure 6 — Figure 6: Heatmaps showing performance of LLMs for primary and secondary characters across different common-sense contradiction types for the implicit and explicit prompting setups.

Panel of Figure 7 — Figure 7: Performance of different LLMs on various common-sense reasoning types under implicit and explicit prompting setups.

Moral Reasoning in Language Models

Moral reasoning in language models studies how AI systems judge dilemmas, weigh competing values, explain decisions, and stay consistent across contexts.

See all 114 papers

CORDA: A Benchmark for Hierarchical Harm-Centric Moral Reasoning in Large Language Models
Aug 8, 2026Siddarth Singh, Victoria Williams, Simon Rosen +6LLM Evaluation Moral Reasoning in Language Models
University of the Witwatersrand, Johannesburg, South Africa
Are LLMs Bad at Moral Reasoning?
Jun 10, 2026Menghang Zhu, Seth LazarAutomatic Rubric Generation LLM Evaluation
School of Philosophy (Political Philosophy) Renmin University of China · School of Government and Policy Johns Hopkins University
Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMs
Jun 10, 2026Elizaveta Tennant, Benjamin Henke, Anita Keshmirian +5Moral Reasoning in Language Models Language Model Robustness
Google DeepMind · Imperial College London · Institute of Philosophy, School of Advanced Study, University of London +2

Aug 8, 2026J/K move · Enter open · S save

CORDA: A Benchmark for Hierarchical Harm-Centric Moral Reasoning in Large Language Models

Siddarth Singh, Victoria Williams, Simon Rosen, Ebenezer Gelo+5

University of the Witwatersrand, Johannesburg, South Africa

LLM Evaluation Moral Reasoning in Language Models Moral Reasoning

Open paper AskPDF arXiv

The key question in moral judgement is not simply whether someone chooses the "right" answer, but how they decide what matters most when moral principles conflict. Current evaluations of large language models (LLMs) remain limited: most test whether models give morally acceptable answers, match human preferences, or avoid obvious violations, rather than whether they can prioritise between competing principles when no option is morally cost-free. We introduce CORDA (Conditioned Ordering and Ranked Directive Adherence), a benchmark for evaluating hierarchical, harm-centred moral reasoning in LLMs. Building on the morality chains formalism, CORDA tests 90 moral dilemmas involving trolley-style cases, medical trade-offs, resource allocation, and human-animal-robot conflicts across four ordered ethical frameworks: Utility, Utility + Agent Harm, Dual-Process, and Dual-Process + Agent Harm. Together, these frameworks test whether models can adapt their decisions when moral priorities change. Across ten instruction-tuned models from seven providers, we find a strong deontological default, with 9 of 10 prioritising avoidance of direct personal harm over reducing overall harm. Models also perform more reliably on categorical harm-avoidance rules, such as avoiding killing, than on outcome-based comparisons, such as minimising total harm, suggesting that they recognise moral red lines more easily than they reason through competing harms. Although all models respond to explicit chain conditioning, several fail to consistently follow specified priority orderings, such as humans over animals and animals over robots. CORDA addresses a central gap in LLM moral evaluation by testing whether models can move beyond default harm-avoidant responses and apply context-specified moral priorities. Moral reliability requires more than default restraint; it requires controllability under conflict.

Abstract

Moral Reasoning in Language Models

Explore similar work

CORDA: A Benchmark for Hierarchical Harm-Centric Moral Reasoning in Large Language Models

Are LLMs Bad at Moral Reasoning?

Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMs