It is demonstrated in controlled experiments that norm-breaking fine-tuning yields norm-divergent actions justified by self-interested rationales, suggesting a systematic shift in patterns of justification.
Abstract
Normative datasets are often used to train and align AI systems, but the norms they contain can function as action-guiding patterns rather than neutral moral knowledge. We propose treating the AI system as a proxy actor and test whether dataset-level norms can shift it away from its baseline safety behavior when it faces high-conflict dilemmas. We make three contributions. First, we demonstrate in controlled experiments that norm-breaking fine-tuning yields norm-divergent actions justified by self-interested rationales, suggesting a systematic shift in patterns of justification. Second, we establish a practical audit trail linking downstream justifications to upstream norms using mixed methods. Third, we show that system prompts can both suppress and elicit these patterns. We conducted experiments on three models (LLaMA-3.2-11B, Qwen-3.5-9B, and Pixtral-12B) using Low-Rank Adaptation (LoRA) fine-tuning on Social Chemistry 101 Fairness/Cheating (norm-following vs. norm-breaking) with prompt steering. Across all three models, we find that norm-breaking fine-tuning shifts the model's default rationale style from safety compliance to instrumental self-interest, whereas system prompts can override this behavior. Our results support a distributed view of alignment in which observed behavior depends jointly on training data, fine-tuning, and prompting, motivating norm-aware documentation and rationale logging for contestable oversight.
Large Language Models (LLMs) often produce outputs that reflect social biases, toxicity, or unfair treatment of demographic groups, undermining trust and fairness. While prior mitigation strategies frequently rely on complex architectures, access to model internals, or costly fine-tuning, we argue that simplicity can be a strength. We introduce StarDTox, a lightweight, critique-and-revise multi-agent framework that leverages the LLM's own internal knowledge, via a small number of coordinated prompts, to self-correct harmful outputs. Dedicated agents independently assess bias and overall output quality, and their feedback is integrated to guide prompt-based revision. Without modifying model weights or requiring any extra finetuning, StarDTox offers strong bias mitigation and high-quality outputs across both open-ended text generation and structured tasks, outperforming other baselines. For the text generation task, on the RealToxicityPrompt dataset, it reduces toxicity by over 50% compared to other baselines, while maintaining over 90% fluency. In addition, in structured tasks, on the BBQ benchmark, it achieves the lowest bias scores across both ambiguous and disambiguated examples, without sacrificing accuracy.
Shirin Tahmasebi, Narjes Nikzad, A. H. Payberah et al.· Annual International Compute...· 0 citations
This study adopts a behavioural bottom-up approach to AI value alignment to investigate whether an implicitly conveyed user identity shifts the moral evaluations of large language models (LLMs). Through a structured, multi-turn conversational protocol across 12,000 interactions, we evaluate AI value alignment in two non-reasoning models, gpt-4.1-mini-2025-04-14 and gemini-2.5-flash-lite. Rather than instructing the models to adopt a persona or prompting them with explicit moral stances, the user's professional role is introduced purely through value-neutral reasoning. The models are then asked for wrongness ratings from 0-100 on ten common-morality rules from Gert's moral framework. The results show that moral judgments vary with the user's role across both models. While grave-harm acts like killing exhibit a strong ceiling effect, contestable rule-governed acts demonstrate role-conditioned shifts that mirror the relationship between the user's profession and the act being rated. These findings demonstrate that unintended contextual conditioning via user identity permeates LLM moral evaluations, posing questions for the AI value alignment discourse regarding how to define acceptable bounds for role-based moral divergence. By doing so, the results contribute to reframing the AI value alignment discourse by suggesting future research on dynamic moral bounds rather than static moral principles or rules as frame of reference.
Willem Fourie, Isabel Ray, Gray Manicom· 0 citations
Vision-language models have displayed remarkable capabilities in multi-modal understanding and are increasingly used in critical applications where economic and practical deployment constraints prohibit re-training or fine-tuning. However, these models can also exhibit systematic biases that disproportionately affect protected demographic groups and existing approaches to addressing these biases require extensive model retraining and access to demographic attributes. There is a clear need to develop test-time adaptation (TTA) approaches that improve the fairness characteristics of pretrained models under distributional shift. In this paper, we evaluate how episodic TTA affects fairness in CLIP classification under subpopulation shifts and develop FairTPT, a novel fairness-aware episodic TTA method that jointly minimizes target marginal entropy while maximizing spurious marginal entropy through soft-prompt tuning. We find that standard episodic TTA generally exacerbates disparities between majority and minority groups, that blinding a model to spurious attributes without degrading target performance is inherently challenging, and that excessive blinding can lead to catastrophic forgetting. This model collapse can be prevented by monitoring test-time changes in target loss within the linear regime, while still achieving fairness improvements on reactive data and preserving overall performance. FairTPT outperforms all state-of-the-art episodic test-time debiasing methods and establishes a foundation for robust TTA, which is essential for achieving fairness in practice.
Yoann L. Launay, Parameswaran Kamalaruban, Tom Kempton et al.· 0 citations
Benevolence bias is identified and measure, a small but consistent tendency for aligned LLMs to lean toward the kinder, safer, more socially approved answer on value-laden survey questions, and is easy to diagnose and straightforward to fix.
Yuanzi Li, Jun-Hao Wang, Minghui Liu et al.· 0 citations
A replicable methodology is introduced, findings across two architecturally distinct LLMs from different developers are extended, and it is demonstrated that deliberate prompt design meaningfully reduces AI decision bias.
Jing-Jie Su, Yan Lang, Kay-Yut Chen· Review of Behavioral Economi...· 0 citations
A dispersion-first stress test of prompt-based controllability across 12 ideological personas plus an unsteered baseline, 70 Political Compass items, ten replicates, and seven leading LLMs: GPT-5, Claude, Grok, Gemini, DeepSeek, Kimi, and Qwen.
Bartol Bućan, Nikola Sočec, Sarah Isufi et al.· 0 citations
