Recent advances in large language models (LLMs) have significantly enhanced automated legal reasoning. Yet, it remains unclear whether their performance reflects genuine legal reasoning ability or artifacts of data contamination. We present a comprehensive empirical study of tax law reasoning approaches and implement a contamination detection protocol to rigorously assess LLM reliability. We show that performance can be inflated by contamination. Building on this analysis, we conduct a systematic evaluation, comparing monolithic LLMs with hybrid systems that translate statutory text into formal representations and delegate inference to symbolic solvers. We build a novel test suite designed to probe generalization to unseen documents via case and rule variations. Our findings indicate that legal reasoning is inherently compositional and that neuro-symbolic frameworks offer a more reliable and robust foundation for legal AI, as well as improved generalization to unobserved situations.

In-depth logical reasoning in domains such as legal analysis remains challenging for large language models (LLMs), exhibiting brittle and inconsistent performance on tasks requiring multi-step deductive inference. This limitation is expected, as LLMs primarily rely on learned distributional associations rather than performing explicit step-by-step logical inference through formal rule application. To address this gap, we propose a novel framework, called PercepLog, that incorporates a classical deductive reasoning algorithm, backward chaining, within an outer control loop, while utilizing LLMs for primitive operations such as substitution and unification. By following the backward chaining algorithm, the framework benefits from the interpretability and explainability of symbolic reasoning procedures while retaining the flexibility of LLM perception for processing natural-language legal text. We evaluate the proposed method on statutory reasoning tasks across a range of legal domains as well general logical reasoning. The results demonstrate that our approach enables more robust, explainable, and principled reasoning over unstructured legal text, where exhaustive formalization is often brittle and prone to hallucination. Our method yields average relative performance improvements of 58% over direct LLM reasoning and 16% over formal LLM-based translation and Prolog inference pipelines.

Highly imbalanced textual datasets continue to pose a challenge for supervised learning models. However, viewing such imbalanced text data as an anomaly detection (AD) problem has advantages for certain tasks such as detecting hate speech, or inappropriate and/or offensive language in large social media feeds. There the unwanted content tends to be both rare and non-uniform with respect to its thematic character, and better fits the definition of an anomaly than a class. Several recent approaches to textual AD use transformer models, achieving good results but with trade-offs in pre-training and inflexibility with respect to new domains. In this paper we compare two linear models within the NMF family, which also have a recent history in textual AD. We introduce a new approach based on an alternative regularization of the NMF objective. Our results surpass other linear AD models and are on par with deep models, performing comparably well even in very small outlier concentrations.