Vera Schmitt

Alignment has become the dominant paradigm for governing large language models (LLMs), implemented through post-training optimization, safety fine-tuning, and inference-time steering. These approaches are commonly assumed to translate normative intent into reliable system behavior. As LLMs are increasingly deployed as large-scale decision-support and information systems, however, the limits of this assumption have become evident. This chapter analyzes alignment from a systems and governance perspective, examining how alignment and steering mechanisms interact with scale, infrastructure, and deployment dynamics. We argue that contemporary approaches primarily regulate observable output behavior through probabilistic narrative control rather than enforcing durable epistemic constraints. While effective at shaping refusals, explanations, and interaction norms, they do not guarantee robustness, truthfulness, or stability under distribution shift or strategic pressure. As a result, aligned systems may appear compliant while remaining epistemically fragile. Current legal and ethical governance frameworks, which emphasize documentation, transparency, and design-time intent, struggle to address this gap because they operate upstream of the deployment-time mechanisms that shape behavior in practice. Situating alignment within the global political economy of AI, the chapter shows how capital concentration, infrastructural lock-in, and competitive deployment amplify these limitations by redistributing effective control toward actors with access to compute, data, and platforms. These pressures feed back into model behavior, producing internal failure modes such as sycophancy, over-alignment, and calibration collapse. We conclude by reframing alignment as a bounded interface-level control mechanism, highlighting the need for system-level oversight beyond training-time alignment alone.

This notebook reports the XplaiNLP submission to the CheckThat! 2025 shared task on multilingual subjectivity detection. We evaluate two approaches: (1) supervised fine-tuning of transformer encoders, EuroBERT, XLM-RoBERTa, and German-BERT, on monolingual and machine-translated training data; and (2) zero-shot prompting using two LLMs: o3-mini for Annotation (rule-based labelling) and gpt-4.1-mini for DoubleDown (contrastive rewriting) and Perspective (comparative reasoning). The Annotation Approach achieves 1st place in the Italian monolingual subtask with an F_1 score of 0.8104, outperforming the baseline of 0.6941. In the Romanian zero-shot setting, the fine-tuned XLM-RoBERTa model obtains an F_1 score of 0.7917, ranking 3rd and exceeding the baseline of 0.6461. The same model also performs reliably in the multilingual task and improves over the baseline in Greek. For German, a German-BERT model fine-tuned on translated training data from typologically related languages yields competitive performance over the baseline. In contrast, performance in the Ukrainian and Polish zero-shot settings falls slightly below the respective baselines, reflecting the challenge of generalization in low-resource cross-lingual scenarios.

Fact-checking is necessary to address the increasing volume of misinformation. Traditional fact-checking relies on manual analysis to verify claims, but it is slow and resource-intensive. This study establishes baseline comparisons for Automated Fact-Checking (AFC) using Large Language Models (LLMs) across multiple labeling schemes (binary, three-class, five-class) and extends traditional claim verification by incorporating analysis, verdict classification, and explanation in a structured setup to provide comprehensive justifications for real-world claims. We evaluate Llama-3 models of varying sizes (3B, 8B, 70B) on 17,856 claims collected from PolitiFact (2007-2024) using evidence retrieved via restricted web searches. We utilize TIGERScore as a reference-free evaluation metric to score the justifications. Our results show that larger LLMs consistently outperform smaller LLMs in classification accuracy and justification quality without fine-tuning. We find that smaller LLMs in a one-shot scenario provide comparable task performance to fine-tuned Small Language Models (SLMs) with large context sizes, while larger LLMs consistently surpass them. Evidence integration improves performance across all models, with larger LLMs benefiting most. Distinguishing between nuanced labels remains challenging, emphasizing the need for further exploration of labeling schemes and alignment with evidences. Our findings demonstrate the potential of retrieval-augmented AFC with LLMs.

Yurii Shynkarov, Veronika Solopova, Vera Schmitt. Proceedings of the Fourth Ukrainian Natural Language Processing Workshop (UNLP 2025). 2025.