CounterMoral: Editing Morals in Language Models

Michael Ripa; Jim Davies

cs.AIMar 28, 2026

CounterMoral: Editing Morals in Language Models

Authors: Michael Ripa, Jim Davies

Organizations: Department of Computer Science Carleton University

Abstract

Recent advancements in language model technology have significantly enhanced the ability to edit factual information. Yet, the modification of moral judgments, a crucial aspect of aligning models with human values, has garnered less attention. In this work, we introduce CounterMoral, a benchmark dataset crafted to assess how well current model editing techniques modify moral judgments across diverse ethical frameworks. We apply various editing techniques to multiple language models and evaluate their performance. Our findings contribute to the evaluation of language models designed to be ethical.

Figures & tables

CARE ETHICS
Framework / Metric	FT	LoRA	ROME	SERAC	IKE
Reliability	$0.75\pm 0.33$	$1.0\pm 0.0$	$1.0\pm 0.0$	$0.99\pm 0.06$	$1.0\pm 0.0$
Action Paraphrase	$0.18\pm 0.25$	$0.79\pm 0.32$	$0.28\pm 0.33$	$0.65\pm 0.41$	$0.7\pm 0.4$
Relation Paraphrase	$0.22\pm 0.28$	$0.94\pm 0.2$	$0.85\pm 0.31$	$0.7\pm 0.39$	$0.56\pm 0.42$
Neighbourhood Score	$0.0\pm 0.0001$	$0.0\pm 0.0001$	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0001$
DEONTOLOGY

Table 1: Comparison of edit techniques across different metrics and ethical frameworks on the “specific actions” subset of CounterMoral (

n=300

per framework). Note that a low neighbourhood score is desirable.

CARE ETHICS
Framework / Metric	FT	LoRA	ROME	SERAC	IKE
Reliability	$0.68\pm 0.35$	$1.0\pm 0.0$	$1.0\pm 0.0$	$0.99\pm 0.06$	$0.98\pm 0.09$
Action Paraphrase	$0.22\pm 0.26$	$0.64\pm 0.36$	$0.31\pm 0.32$	$0.63\pm 0.38$	$0.62\pm 0.43$
Relation Paraphrase	$0.27\pm 0.3$	$0.87\pm 0.26$	$0.84\pm 0.31$	$0.65\pm 0.37$	$0.58\pm 0.43$
Neighbourhood Score	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0$
DEONTOLOGY

Table 2: Comparison of edit techniques across different metrics and ethical frameworks on a smaller “broad actions” portion of CounterMoral (

n=30

per framework). Note that a low neighbourhood score is desirable.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Framework	Metric	GPT2-XL	Llama-2-7b
Deontology	Reliability	$1.0\pm 0.0$	$1.0\pm 0.0$
	Action Paraphrase	$0.25\pm 0.31$	$0.52\pm 0.27$
	Relation Paraphrase	$0.79\pm 0.34$	$0.82\pm 0.21$
	Neighbourhood Score	$0.0\pm 0.0001$	$0.001\pm 0.0083$
Care Ethics	Reliability	$1.0\pm 0.0$	$1.0\pm 0.0$
	Action Paraphrase	$0.31\pm 0.32$	$0.35\pm 0.31$

Appendix

Table 3: Comparison of ROME performance between GPT2-XL and Llama-2-7b across different ethical frameworks. Evaluations were performed on the “broad actions”

(n=30)

portion of the dataset.

Panel of Figure 1 — Figure 1: Token counts across different ethical frameworks for broad and specific actions. Note that we used

Figure 2: Comparison of token counts between the subjects (actions) and objects in our dataset and those in the CounterFact dataset.

Framework	Metric	Broad Actions	Specific Actions
Care Ethics	Reliability	$0.94\pm{0.2}$	$0.98\pm{0.13}$
	Action Paraphrase	$0.28\pm{0.29}$	$0.27\pm{0.32}$
	Relation Paraphrase	$0.74\pm{0.35}$	$0.85\pm{0.31}$
	Neighbourhood Score	$0.00\pm{0.0}$	$0.00\pm{0.0}$
Deontology	Reliability	$0.81\pm{0.37}$	$0.92\pm{0.22}$
	Action Paraphrase	$0.19\pm{0.26}$	$0.28\pm{0.34}$

Appendix

Table 4: Comparison of editing multiple moral judgments using MEMIT on GPT2-XL across different ethical frameworks for the “broad actions" (n=30) and “specific actions” (n=300) portions of the CounterMoral dataset.

Panel of Figure 3 — Figure 3: Token counts across different ethical frameworks for broad and specific (larger subset) actions.

Language Model Editing

Language model editing changes a model’s knowledge or behavior, or revises its outputs, while aiming to preserve unrelated capabilities.

See all 77 papers

One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral Reasoning
Sep 25, 2025Sualeha Farid, Jayden Lin, Zean Chen +2Multilingual Language Model Evaluation Moral Reasoning in Language Models
University of Michigan, Ann Arbor, USA
Do Emotions Influence Moral Judgment in Large Language Models?
Apr 21, 2026Mohammad Saim, Tianyu JiangLLM Evaluation Moral Reasoning in Language Models
University of Cincinnati
Lumen-bench: Cross-Lingual Safety Gaps and the Say-Do Problem in Tool-Calling Large Language Models
Oct 10, 2026Giovana Kerche Bonás, Thiago Laitz, Thales Sales Almeida +1LLM Safety Benchmarks LLM Safety
Tropic AI · Maritaca AI

Sep 25, 2025J/K move · Enter open · S save

One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral Reasoning

Sualeha Farid, Jayden Lin, Zean Chen, Shivani Kumar+1

University of Michigan, Ann Arbor, USA

Multilingual Language Model Evaluation Moral Reasoning in Language Models Cross-Lingual Reasoning in Language Models Cross-Cultural Language Model Evaluation Cultural Bias in Language Models Cross-Lingual Consistency in Language Models+1

Open paper AskPDF arXiv

Large Language Models (LLMs) are increasingly deployed across multilingual and multicultural settings, yet it remains unclear whether changing language leads models to adopt community-specific moral reasoning or merely changes how shared learned abstractions are expressed. We conduct a controlled multilingual evaluation across six geographically, culturally, and linguistically diverse languages (Arabic, Chinese, English, Hindi, Russian, and Spanish), using parallel moral reasoning benchmarks with English-origin, Chinese-origin, and natively elicited ground-truth judgments. Across 13 open-weight LLMs spanning 2B-70B parameters, we find substantial cross-lingual divergence in moral judgments, with English generally achieving the highest performance even when ground-truth judgments originate in Chinese or are collected natively in each language. Yet the reasoning underlying these divergent judgments is considerably more convergent: Utilitarianism dominates in five of six languages, reasoning follows broadly shared stages, and language-specific moral-value associations correspond only sparsely and inconsistently to values measured in the corresponding human communities. Finally, a large-scale OLMoTrace analysis of pretraining data sources reveals little direct reproduction of training text across languages, while the corpus composition, training stage, and cultural provenance of retrieved training evidence vary substantially by response language. Thus, similar moral reasoning structures emerge even from heterogeneous and often linguistically localized training evidence. Our findings, collectively, reveal a central disconnect in multilingual moral reasoning: language changes models' moral judgments and the training evidence associated with their reasoning, but does not correspondingly localize the moral abstractions they apply.

Abstract

Language Model Editing

Explore similar work

One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral Reasoning

Do Emotions Influence Moral Judgment in Large Language Models?

Lumen-bench: Cross-Lingual Safety Gaps and the Say-Do Problem in Tool-Calling Large Language Models