arXiv subjects

Anne Arzberger

Publications and source records attributed to Anne Arzberger.

3 recordsLinked to original sources

arXiv2026

Beyond a Global Norm: Personalizing Toxicity Sensitivity in Language Models Without Retraining

Reducing toxicity is often framed as a global alignment problem, yet perceptions of harmful language are subjective and context-dependent. We present the first comparative evaluation of training-free methods for aligning language generation to user-specific toxicity sensitivities across three inference-time intervention stages: pre-decoding (prompt conditioning and rewriting), in-decoding (token, logit, and representation steering), and post-decoding (candidate re-ranking). Evaluated against toxicity sensitivity targets derived from the PRISM dataset, all methods reduce alignment error by 28-47%. However, the results reveal a fundamental trade-off between alignment effectiveness, personalization, and general language quality, showing how toxicity sensitivity alignment is an inherently multi-objective problem.

cs.CL↗

arXiv2026

"Label from Somewhere": Reflexive Annotating for Situated AI Alignment

AI alignment relies on annotator judgments, yet annotation pipelines often treat annotators as interchangeable, obscuring how their social position shapes annotation. We introduce reflexive annotating as a probe that invites crowd workers to reflect on how their positionality informs subjective annotation judgments in a language model alignment context. Through a qualitative study with crowd workers (N=30) and follow-up interviews (N=5), we examine how our probe shapes annotators' behaviour, experience, and the situated metadata it elicits. We find that reflexive annotating captures epistemic metadata beyond static demographics by eliciting intersectional reasoning, surfacing positional humility, and nudging viewpoint change. Crucially, we also denote tensions between reflexive engagement and affective demands such as emotional exposure. We discuss the implications of our work for richer value elicitation and alignment practices that treat annotator judgments as situated and selectively integrate positional metadata.

cs.HC↗

arXiv2026

Co-Constructing Alignment: A Participatory Approach to Situate AI Values

As AI systems become embedded in everyday practice, value misalignment has emerged as a pressing concern. Yet, dominant alignment approaches remain model centric, treating users as passive recipients of prespecified values rather than as epistemic agents who encounter and respond to misalignment during interactions. Drawing on situated perspectives, we frame alignment as an interactional practice co-constructed during human AI interaction. We investigate how users understand and wish to contribute to this process through a participatory workshop that combines misalignment diaries with generative design activities. We surface how misalignments materialise in practice and how users envision acting on them, grounded in the context of researchers using Large Language Models as research assistants. Our findings show that misalignments are experienced less as abstract ethical violations than as unexpected responses, and task or social breakdowns. Participants articulated roles ranging from adjusting and interpreting model behaviour to deliberate non-engagement as an alignment strategy. We conclude with implications for designing systems that support alignment as an ongoing, situated, and shared practice.

cs.HC↗