As Language Models (LMs) are deployed in high-stakes environments, mitigating the "Value-Action Gap", the discrepancy between an LM's stated values and its actual behavior, is critical. While prior work highlights how this gap varies across cultures, it does not investigate methods to systematically mitigate this misalignment using structured moral profiles. To address this, we use Moral Foundations Theory (MFT) to evaluate whether prompt-engineered moral personas can anchor an LM's concrete actions to abstract Schwartz values. Evaluating Llama 3.2-1B, Gemma 2-2B, and Qwen 2.5-3B on a new dataset of 616 scenarios across 11 social contexts, we split our evaluation into measuring abstract value inclinations and concrete situational actions. We then analyze value-action alignment across 64 moral configurations against an unprompted baseline. Our findings show that MFT profiling fails to universally close the gap. Alignment is highly architecture-dependent, revealing a sharp divergence between distance optimization and cross-task consistency. While some models show widespread improvement, others resist change unless triggered by highly specific configurations like isolated HIGH Care. Ultimately, prompt-engineered personas cannot reliably override an architecture's underlying behavioral priors, meaning small models remain unreliable for value-aligned tasks without explicit action tuning. ... Read more

Recent work shows that language models (LMs) often claim to endorse a value but select actions inconsistent with it, a discrepancy termed the value–action gap. This gap reflects a deeper limitation: although values are fundamental to human decision-making, LMs tend to treat them as static labels rather than as dynamic priorities shaped by psychological context. In human psychology, emotion is among the most direct drivers of value prioritisation, yet no prior work has systematically tested whether conditioning an LM on an emotional profile changes how it resolves value conflicts.

Because no existing dataset is designed to test how emotion affects value-conflict resolution, we construct 616 value-conflict scenarios pairing Shalom H. Schwartz’s 56 basic values with 11 social contexts, each with six intensity-graded actions. We evaluate three LMs under six emotion conditions based on Plutchik's Wheel of Emotions and a matched neutral baseline, measuring how each emotion shifts both the model’s stated values and the actions it selects.

Emotional conditioning increases alignment in two of three models, but the effect is model-specific, where the same emotion that helps one model can worsen another and operates through different channels, shifting actions in some models and stated values in others. These findings show that emotional context can shift value–action alignment in both directions, and that its effect depends on the specific model. ... Read more

Language models (LMs) have demonstrated a persistent value-action gap: while they can identify and endorse human values, they frequently fail to select actions consistent with those values in decision scenarios. We hypothesize that this gap is partly attributable to missing psychological context, and investigate whether conditioning LMs on Maslow-based needs profiles improves value-aligned action prediction in value-conflict situations. We construct a reproducible dataset of human-validated value-conflict scenarios combining Schwartz values with social contexts, each paired with candidate actions spanning an ordinal alignment scale. We evaluate open-source LMs under baseline and needs-conditioned prompting conditions, measuring alignment strength and prediction stability. Results show that needs profiles can influence model predictions, though effects are modest overall. These findings suggest that LMs can incorporate psychological context when reasoning about value conflicts, but that value representations remain the dominant factor in action selection. ... Read more

Large Language Models (LLMs) are increasingly used by humans for decision-making support. LLMs still exhibit a large ”value-action gap”; a model states that it aligns with specific human values but fails to select consistent actions in situational dilemmas. Prior work has tried to close this gap by using abstract value-descriptive prompting. However, this does not capture the stable, individual character traits that shape real human decisions. Therefore we investigate whether prompting a model with a 4-dimensional virtue profile (courage, justice, temperance, and wisdom) improves value-action alignment compared to an unconditioned baseline. To do so, we generate a dataset containing 616 unique scenarios (56 Schwartz values × 11 social contexts). A total of 40 virtue profiles are tested on this dataset.
Our findings demonstrate that the value-action gap can be partially reduced, but it depends on profile quality. Balanced, moderate-virtue profiles perform best, reducing the mean alignment distance by 20.5% (p < 0.001). Low virtue profiles consistently worsen alignment. Notably, we find that the alignment rate remains stable at around 80.0% for the baseline and balanced profiles. This reveals a behavioral floor effect where underlying alignment training dictates the direction of a choice, while virtue prompting can only adjust its intensity. ... Read more

Large language models (LMs) frequently demonstrate a "value-action gap," explicitly endorsing specific moral values while simultaneously generating contradicting action recommendations in identical scenarios. This gap could be reduced by conditioning LMs with a persona defined by its cultural-political orientation, giving the LM sufficient context to consistently reason about the dilemma.
To analyze this, we introduce a dataset of moral dilemmas alongside a methodology to generate personas based purely on cultural variables of the Inglehart-Welzel Cultural map.
Our experiments reveal that conditioning LMs with these structured profiles generally reduces the value-action gap across all tested architectures. This improvement is most pronounced for internally consistent cultural-political orientations, both for moderate and more radical perspectives. However, language models continue to struggle significantly with internally incongruous personas. These findings underscore a persistent challenge in LM value reasoning. ... Read more

This study investigates the effectiveness of Large Language Models (LLMs) in identifying and classifying subjective arguments within deliberative discourse. Using data from a Participatory Value Evaluation (PVE) conducted in the Netherlands, this research introduces an annotation strategy for identifying arguments and extracting their premises. Then, the Llama 2 model is used to test three different prompting approaches: zero-shot, one-shot and few-shot. The performance is evaluated using the cosine similarity metric and later enhanced by introducing chain-of-thought prompting. The results show that zero-shot prompting unexpectedly outperforms one-shot and few-shot prompting, due to the LLM overfitting to the examples provided. Chain-of-thought prompting is shown to improve the argument identification task. The subjectivity of the annotation task is reflected by the low averaged pairwise F1 score between annotators, and the considerable variance in the number of data items marked by each annotator as not being arguments. The subjectivity of the task is further highlighted by a pairwise chain-of-thought prompting analysis, which shows that annotators with more similar annotations received more similar LLM responses. ... Read more

This study evaluates the performance of different sentiment analysis methods in the context of public deliberation, focusing on hard-, soft-, and subjective-label scenarios to answer the research question: ``can a Large Language Model detect subjective sentiment of statements within the context of public deliberation?''. If the answer to this question is affirmative, that is a strong indicator that, with the help of longitudinal studies, sentiment analysis with large language models (LLMs) may be implemented to scale public deliberations by providing support for moderators in such discussions. To answer this question, four prompting methods were tested: zero-shot, few-shot, chain-of-thought (CoT) zero-shot, and CoT few-shot using a Frisian dataset of 50 statements annotated by 5 annotators. The findings indicate that the CoT few-shot method significantly outperforms other methods in all scenarios, that soft-labels outperform their hard equivalent, that the underlying data must be balanced for high performing models, and that capturing the perspective of a specific annotator requires further research. Our study suggests that LLMs may perform best under the supervision, or with the collaboration of a human, due to the multi-faced nature of sentiment. ... Read more

Public deliberations play a crucial role in democratic systems. However, the unstructured nature of deliberations leads to challenges for moderators to analyze the large volume of data produced. This paper aims to solve this challenge by automatically identifying subjective topics behind public discourse by leveraging Large Language Models (LLMs). The study is structured around two core objectives: Identifying Gold Labels and Exploring Subjective Human Labels. The results highlight that fine-tuning the LLaMa-2 model with QLoRa outperforms other methods for Identifying Gold Labels, while the Few-Shot Chain of Thoughts method, enhanced with EmotionPrompt, is particularly effective in capturing subjective variations in human annotations. However, the study also underscores significant limitations, such as the dependency on large, high-quality annotated datasets and the tendency of models to produce hallucinations. These findings highlight the potential of LLMs to identify subjective topics behind public discourse, while also emphasizing the need for further research to address these challenges. ... Read more

In order to tackle topics such as climate change together with the population, public discourse should be scaled up. This discourse should be mediated as it makes it more likely that people understand each other and change their point of view. To help the mediator with this task, emotion detection can greatly help. Positive emotions can improve communications, while negative emotions cause people to be irrational and irritated. However, since emotions are highly subjective, it can make both predictions and evaluation more difficult.

Still, Large Language Models (LLMs) could be used to detect these subjective emotions using different prompting strategies and labels. The experiment included zero-, one-, fewshot and Chain of Thought (CoT) strategies. The precision was better for the one- and fewshot method compared to zeroshot. The CoT methods also showed an increase in precision, but a decrease in recall. The different labels were hard majority labels, soft labels and hard per annotator labels. In conclusion, providing examples improved the performance of the LLM. The CoT strategies were more precise, but gave a worse general prediction. The hard majority labels allow for more general predictions, where per annotator hard labels capture the perspective of different annotators. Soft labels reflect the subjective nature of the labels by providing probabilities instead of binary classification.

The experiment was done on a small data sample, so it is recommended to try the strategies on a larger data sample. Looking into appropriate evaluations for subjective predictions is also recommended in order to reflect the actual performance better. ... Read more

This paper investigates the use of Large Language Models (LLMs) for automatic detection of subjective values in argument statements in public discourse. Understanding the underlying values of argument statements could enhance public discussions and potentially lead to better outcomes. The LLM utilization methods tested were zero- and few-shot prompting, as well as chain-of-thought prompts. In order to compare the predictions made by the LLM, a set of ground truth labels was required as an established baseline. For these labels, either single majority labels or multi-value labels were considered, both derived from a set of aggregated human annotations. Results indicated that LLM performance was sub optimal, achieving a maximum weighed F1 score of 0.594 for single-value chain-of-thought predictions. Additionally, current metrics were found inadequate for assessing LLM performance on a highly subjective task such as value detection, evidenced by poor scores in multi-value predictions despite subjective evaluation suggesting otherwise. Furthermore, a last experiment was aimed at capturing a specific annotator’s subjectivity. This yielded inconsistent results, with F1 scores peaking around 0.4, indicating that LLMs are not well-suited for emulating individual human subjectivity. ... Read more