Anomaly detection in textual data remains an underexplored problem, hindered by fragmented benchmarks, scarce labeled data, and methods that fail to model the distribution of normal text in a continuous and generative manner. We propose Flocat, an unsupervised approach that reformulates this task through a source-reversed flow matching paradigm: rather than learning to generate, we learn a flow from inlier text embeddings toward a compact isotropic Gaussian target, whose radius is explicitly constrained to ensure that the distance to the centroid serves as a geometrically consistent anomaly score. Built on a Diffusion Transformer (DiT) backbone, Flocat operates at both token and sentence level through a level-agnostic interpolation mechanism, enabling fine-grained signal interpretability. We evaluate on 10 datasets spanning text classification, spam detection, sentiment analysis, human- vs. machine-generated text and depression screening, consistently achieving state-of-the-art AUC-ROC across all benchmarks and text representations.

This paper describes the system we submitted to the CRF:filling 2026 shared task. We propose a modular, LLM-based framework including an LLM as rewriter, which enhances the original clinical note from the perspective of each target CRF item; an LLM extractor, which retrieves the relevant value using a k-shot prompting strategy; and an LLM as a judge, which determines whether the clinical note contains evidence to support a given answer, defaulting to ’unknown’ otherwise. We evaluated our system on the English portion of the dataset; our complete framework achieves a macro-F1 of 0.64 on the development set. Our analysis reveals that while the rewriting step effectively generates correct factual information, it also increases false positives. The judge component mitigates this by adopting a conservative prediction strategy that substantially reduces false positives at the cost of a moderate reduction in true positives, yielding higher precision and better alignment with the shared task metric. On the test set, a light version of our system ranked 21 out of 32 public submissions, achieving a macro-F1 of 0.45.

Simplifying complex texts is essential to ensure equitable access to information, particularly for individuals with cognitive impairments. The Easy-to-Read (ETR) initiative provides a framework to make content more accessible for these individuals. However, manually creating such texts remains time-consuming and resource-intensive. In this work, we investigate the potential of large language models (LLMs) to automate the generation of ETR content. To address the scarcity of aligned corpora and the specific constraints of ETR, we propose a multi-task learning (MTL) approach that trains models jointly on text summarization, text simplification, and ETR generation. We explore two complementary strategies: multi-task retrieval-augmented generation (RAG) for in-context learning (ICL), and MTL-LoRA for parameter-efficient fine-tuning (PEFT). Our experiments with Mistral-7B and LLaMA-3-8B, conducted on ETR-fr, a new high-quality dataset, show that MTL-LoRA consistently outperforms all other strategies in in-domain settings, while the MTL-RAG-based approach achieves better generalization in out-of-domain scenarios. Our code is publicly available at https://github.com/FrLdy/ETR-PEFT-Composition.