TZTadeusz Zbiegień

Tadeusz Zbiegień

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Tadeusz Zbiegień has a total of 12 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    DSA Risk Assessment -Social Media Mental Harm Automatic Annotation Case Study

    by Tadeusz Zbiegień, Tejas Goyal, Jaromir Savelka, Ewa Ilczuk, Małgorzata Kuśmierczyk, Krystyna Mokrzycka, Przemysław Pałka

    The EU's Digital Services Act (DSA) requires Very Large Online Platforms (VLOPs) to identify and mitigate systemic risks to users' mental well-being through annual risk assessments. However, analyzing these lengthy, heterogeneous reports is labor-intensive for regulators and researchers. This study evaluates the capability of four LLMs (GPT 5.2, GPT5mini, Claude Sonnet 4.5, and Claude Opus 4.5) to automatically annotate mental-healthrelated content in DSA risk reports. Using a human-annotated benchmark of 14 VLOP risk assessments from 2023-2024, we developed a hierarchical annotation schema distinguishing between risk types (internal/external) and mitigation measures (platform-level/external). Our asynchronous, zero-shot annotation pipeline reduced processing time from approximately four hours of human labor per document to under two minutes. Results show Claude Opus achieved the highest sentence-level coverage (56-65%), though all models exhibited systematic biases, such as over-triggering on general risk categories, while under-detecting specific mitigation measures. Although not yet suitable for fully automated compliance assessment, the proposed pipeline establishes a potential foundation for future research on computational oversight and empirical analysis of platform governance under the DSA framework.

  • Open Access

    Generating and Verifying Expert-Level Legal Reasoning: A Proof-of-Concept Study in Polish Criminal Law

    by Tadeusz Zbiegień, Kamil Mamak, Mikołaj Małecki, Michał Rachalski, Benjamin Shishko, Przemysław Pałka

    LLMs are increasingly likely to be used in legal work, including criminal-law settings where unsupported reasoning may have serious consequences. Existing legal benchmarks and exam-style evaluations suggest that LLMs can perform well on some tasks, but do not adequately test expert-level, open-ended, jurisdiction-specific legal analysis. They also tell us little about whether such systems can reliably assess the quality of their own outputs in agentic workflows. To address these gaps, we designed an expert-authored Polish criminal-law vignette and a 54-item rubric and evaluated four LLMs together with a multi-agent pipeline. This paper presents the results of two experiments: case-solving performance and verifier performance. We find that chain-of-thought prompting substantially improves answer generation, with GPT CoT and Claude CoT achieving the strongest results, while the best multi-agent run is competitive but does not exceed the strongest CoT systems. By contrast, verifier performance lags behind generation: apart from GPT CoT and Claude CoT, most systems assess their own answers too leniently relative to expert judgment. We therefore argue that solving a legal case and reliably evaluating such a solution are distinct capabilities. This distinction matters for agentic legal systems, where a weak verifier may reinforce rather than correct erroneous output. Although preliminary and limited to a single vignette, this study shows the value of expert-authored, rubric-based evaluation for legal AI and highlights verifier reliability as an important bottleneck for agentic legal systems.

  • Open Access

    Investigating Expert-Based Prompt Engineering for Legal Entailment Tasks

    by Cor Steging, Ludi van Leeuwen, Tadeusz Zbiegień

    Legal reasoning is complex and multi-faceted, requiring a broad set of skills. By employing domain knowledge from legal experts, we design five elements that can be included in prompts for large language models that could aid in legal reasoning tasks. We use additional legal guidelines, 1-shot prompting, dictionary definitions, knowledge representations of legal articles, and IRAC-style prompting. We investigate the effect of each prompt element on the model’s performance on a legal entailment task. Certain prompt elements can improve performance, depending on the context and the model. For the smaller models, increasing the number of prompt elements improves performance on average. For any particular combination of model and sub-task, only using a subset of the prompt elements seems to work best. For the most advanced reasoning model we evaluate, using a selection of prompt elements increases average performance across all evaluated sub-tasks. Results indicate that the problem space of the legal entailment task may be too large for a single model and prompt. In future research, we therefore aim to investigate the capabilities of an ensemble of specialized models.