Hao Mi, Qiang Sheng, Shaofei Wang, Beizhe Hu, Yifan Sun, Zhengjia Wang, Hengqi Zeng, Yang Li, Danding Wang, Juan Cao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Ethical decision-making is a critical aspect of human judgment, and the growing use of LLMs in decision-support systems necessitates a rigorous evaluation of their moral reasoning capabilities. However, existing assessments primarily rely on single-step evaluations, failing to capture how models adapt to evolving ethical challenges. Addressing this gap, we introduce the Multi-step Moral Dilemmas (MMDs), the first dataset specifically constructed to evaluate the evolving moral judgments of LLMs across 3,302 five-stage dilemmas. This framework enables a fine-grained, dynamic analysis of how LLMs adjust their moral reasoning across escalating dilemmas. Our evaluation of nine widely used LLMs reveals that their value preferences shift significantly as dilemmas progress, indicating that models recalibrate moral judgments based on scenario complexity. Furthermore, pairwise value comparisons demonstrate that while LLMs often prioritize the value of care, this value can sometimes be superseded by fairness in certain contexts, highlighting the dynamic and context-dependent nature of LLM ethical reasoning. Our findings call for a shift toward dynamic, context-aware evaluation paradigms, paving the way for more human-aligned and value-sensitive development of LLMs.

Partial multi-label learning (PML) models the scenario where each training sample is annotated with a set of candidate labels, but only a subset of them corresponds to the ground-truths. The key challenge for PML is how to minimize the negative impact of incorrect labels concealed within the candidate ones. Most existing PML solutions require abundant samples to train a noise-robust multi-label predictor. However, due to privacy, safety or ethic issues, we more often have a handful of training samples for the target task. In this paper, we propose an approach named FsPML-DA (Few-shot Partial Multi-Label Learning with Data Augmentation) to simultaneously estimate label confidence, perform data augmentation and induce multilabel classifier. Specifically, FsPML-DA disambiguates the label confidence vector of each PML sample by jointly modeling the feature and semantic similarity, label credibility of other samples and label co-occurrence. Next, FsPML-DA introduces a synthetic feature network to generate more training samples from pairs of given samples with label confidence values. FsPML-DA then leverages original and generated samples to train a noise-tolerant multi-label classifier. Extensive experiments on benchmark datasets show that FsPML-DA performs better than recent competitive PML baselines and few-shot solutions. FsPML-DA can dislodge noisy labels by mining PML data in a sensible way and the proposed data augmentation strategy effectively combats with the scarcity of few-shot training samples.