Enhancing Multilingual Counterfactual Generation through Alignment-as-Preference Optimization

2026-05-12

AI Summary

Existing methods for generating counterfactual explanations in non-English languages struggle to simultaneously achieve high validity—defined as successfully flipping the model’s prediction—and minimality, which requires minimal perturbation to the input, due to an inherent trade-off between these two objectives. This work proposes Macro, a novel framework that formulates multilingual counterfactual generation as a preference alignment problem. Macro constructs quantifiable preference pairs using a composite scoring function and fine-tunes multilingual large language models via Direct Preference Optimization (DPO), eliminating the need for translation or supervised fine-tuning. Experiments across four models and seven languages demonstrate that Macro improves average validity by 12.55% while preserving minimality, significantly outperforming chain-of-thought and translation-based baselines, reducing generation errors, and enhancing cross-lingual consistency.

Abstract

Self-generated counterfactual explanations (SCEs) are minimally modified inputs (minimality) generated by large language models (LLMs) that flip their own predictions (validity), offering a causally grounded approach to unraveling black-box LLM behavior. Yet extending them beyond English remains challenging: existing methods struggle to produce valid SCEs in non-dominant languages, and a persistent trade-off between validity and minimality undermines explanation quality. We introduce Macro, a preference alignment framework that applies Direct Preference Optimization (DPO) to multilingual SCE generation, using a composite scoring function to construct preference pairs that effectively translate the trade-off into measurable preference signals. Experiments across four LLMs and seven typologically diverse languages show that Macro improves validity by 12.55\% on average over the chain-of-thought baseline without degrading minimality, while avoiding the severe minimality violations of the translation-based baseline. Compared to supervised fine-tuning, Macro achieves superior performance on both metrics, confirming that explicit preference optimization is essential for balancing this trade-off. Further analyses reveal that Macro increases cross-lingual perturbation alignment and mitigates common generation errors. Our results highlight preference optimization as a promising direction for enhancing multilingual model explanations.

Problem

• counterfactual generation
• multilingual
• validity
• minimality
• preference optimization

Innovation

• preference optimization
• counterfactual generation
• multilingual alignment
• Direct Preference Optimization (DPO)
• minimal perturbation