Fusion Matters: Length-Aware Analysis of Positional-Encoding Fusion in Transformers

Transformers require positional encodings to represent sequence order, yet most prior work focuses on designing new positional encodings rather than examining how positional information is fused with token embeddings. In this paper, we study whether the fusion mechanism itself affects performance, particularly in long-sequence settings. We conduct a controlled empirical study comparing three canonical fusion strategies--element-wise addition, concatenation with projection, and scalar gated fusion--under identical Transformer architectures, data splits, and random seeds. Experiments on three text classification datasets spanning short (AG News), medium (IMDB), and long (ArXiv) sequences show that fusion choice has negligible impact on short texts but produces consistent gains on long documents. To verify that these gains are structural rather than stochastic, we perform paired-seed analysis and cross-dataset comparison across sequence-length regimes. Additional experiments on the ArXiv dataset indicate that the benefit of learnable fusion generalizes across multiple positional encoding families. Finally, we explore a lightweight convolutional gating mechanism that introduces local inductive bias at the fusion level, evaluated on long documents only. Our results indicate that positional-encoding fusion is a non-trivial design choice for long-sequence Transformers and should be treated as an explicit modeling decision rather than a fixed default.

Profiling German Text Simplification with Interpretable Model-Fingerprints

While Large Language Models (LLMs) produce highly nuanced text simplifications, developers currently lack tools for a holistic, efficient, and reproducible diagnosis of their behavior. This paper introduces the Simplification Profiler, a diagnostic toolkit that generates a multidimensional, interpretable fingerprint of simplified texts. Multiple aggregated simplifications of a model result in a model's fingerprint. This novel evaluation paradigm is particularly vital for languages, where the data scarcity problem is magnified when creating flexible models for diverse target groups rather than a single, fixed simplification style. We propose that measuring a model's unique behavioral signature is more relevant in this context as an alternative to correlating metrics with human preferences. We operationalize this with a practical meta-evaluation of our fingerprints' descriptive power, which bypasses the need for large, human-rated datasets. This test measures if a simple linear classifier can reliably identify various model configurations by their created simplifications, confirming that our metrics are sensitive to a model's specific characteristics. The Profiler can distinguish high-level behavioral variations between prompting strategies and fine-grained changes from prompt engineering, including few-shot examples. Our complete feature set achieves classification F1-scores up to 71.9 %, improving upon simple baselines by over 48 percentage points. The Simplification Profiler thus offers developers a granular, actionable analysis to build more effective and truly adaptive text simplification systems.

Local Language Models for Context-Aware Adaptive Anonymization of Sensitive Text

Qualitative research often contains personal, contextual, and organizational details that pose privacy risks if not handled appropriately. Manual anonymization is time-consuming, inconsistent, and frequently omits critical identifiers. Existing automated tools tend to rely on pattern matching or fixed rules, which fail to capture context and may alter the meaning of the data. This study uses local LLMs to build a reliable, repeatable, and context-aware anonymization process for detecting and anonymizing sensitive data in qualitative transcripts. We introduce a Structured Framework for Adaptive Anonymizer (SFAA) that includes three steps: detection, classification, and adaptive anonymization. The SFAA incorporates four anonymization strategies: rule-based substitution, context-aware rewriting, generalization, and suppression. These strategies are applied based on the identifier type and the risk level. The identifiers handled by the SFAA are guided by major international privacy and research ethics standards, including the GDPR, HIPAA, and OECD guidelines. This study followed a dual-method evaluation that combined manual and LLM-assisted processing. Two case studies were used to support the evaluation. The first includes 82 face-to-face interviews on gamification in organizations. The second involves 93 machine-led interviews using an AI-powered interviewer to test LLM awareness and workplace privacy. Two local models, LLaMA and Phi were used to evaluate the performance of the proposed framework. The results indicate that the LLMs found more sensitive data than a human reviewer. Phi outperformed LLaMA in finding sensitive data, but made slightly more errors. Phi was able to find over 91% of the sensitive data and 94.8% kept the same sentiment as the original text, which means it was very accurate, hence, it does not affect the analysis of the qualitative data.

Prompt Tuning without Labeled Samples for Zero-Shot Node Classification in Text-Attributed Graphs

Node classification is a fundamental problem in information retrieval with many real-world applications, such as community detection in social networks, grouping articles published online and product categorization in e-commerce. Zero-shot node classification in text-attributed graphs (TAGs) presents a significant challenge, particularly due to the absence of labeled data. In this paper, we propose a novel Zero-shot Prompt Tuning (ZPT) framework to address this problem by leveraging a Universal Bimodal Conditional Generator (UBCG). Our approach begins with pre-training a graph-language model to capture both the graph structure and the associated textual descriptions of each node. Following this, a conditional generative model is trained to learn the joint distribution of nodes in both graph and text modalities, enabling the generation of synthetic samples for each class based solely on the class name. These synthetic node and text embeddings are subsequently used to perform continuous prompt tuning, facilitating effective node classification in a zero-shot setting. Furthermore, we conduct extensive experiments on multiple benchmark datasets, demonstrating that our framework performs better than existing state-of-the-art baselines. We also provide ablation studies to validate the contribution of the bimodal generator. The code is provided at: https://github.com/Sethup123/ZPT.

LLMs-Integrated Automatic Hate Speech Recognition Using Controllable Text Generation Models

This paper proposes an automatic speech recognition (ASR) model for hate speech using large language models (LLMs). The proposed method integrates the encoder of the ASR model with the decoder of the LLMs, enabling simultaneous transcription and censorship tasks to prevent the exposure of harmful content. Instruction tuning of the LLM to mask hate-related words with specific tokens requires an annotated hate speech dataset, which is limited. We generate text samples using an LLM with the Chain-of-Thought (CoT) prompting technique guided by cultural context and examples and then convert them into speech samples using a text-to-speech (TTS) system. However, some of them contain non-hate speech samples with hate-related words, which degrades the censorship performance. This paper filters the samples which text classification models correctly label as hate content. By adjusting the threshold for the number of correct answer models, we can control the level of hate in the generated dataset, allowing us to train the LLMs through curriculum learning in a gradual manner. Experimental results show that the proposed method achieves a masking accuracy of 58.6\% for hate-related words, surpassing previous baselines. We also confirm that the curriculum training contributes to the efficiency of both transcription and censorship tasks.

Concepts from Representations: Post-hoc Concept Bottleneck Models via Sparse Decomposition of Visual Representations

Deep learning has achieved remarkable success in image recognition, yet their inherent opacity poses challenges for deployment in critical domains. Concept-based interpretations aim to address this by explaining model reasoning through human-understandable concepts. However, existing post-hoc methods and ante-hoc concept bottleneck models (CBMs), suffer from limitations such as unreliable concept relevance, non-visual or labor-intensive concept definitions, and model or data-agnostic assumptions. This paper introduces Post-hoc Concept Bottleneck Model via Representation Decomposition (PCBM-ReD), a novel pipeline that retrofits interpretability onto pretrained opaque models. PCBM-ReD automatically extracts visual concepts from a pre-trained encoder, employs multimodal large language models (MLLMs) to label and filter concepts based on visual identifiability and task relevance, and selects an independent subset via reconstruction-guided optimization. Leveraging CLIP's visual-text alignment, it decomposes image representations into linear combination of concept embeddings to fit into the CBMs abstraction. Extensive experiments across 11 image classification tasks show PCBM-ReD achieves state-of-the-art accuracy, narrows the performance gap with end-to-end models, and exhibits better interpretability.

Legal experts disagree with rationale extraction techniques for explaining ECtHR case outcome classification

Interpretability is critical for applications of large language models in the legal domain which requires trust and transparency. While some studies develop task-specific approaches, other use the classification model's parameters to explain the decisions. However, which technique explains the legal outcome prediction best remains an open question. To address this challenge, we propose a comparative analysis framework for model-agnostic interpretability techniques. Among these, we employ two rationale extraction methods, which justify outcomes with human-interpretable and concise text fragments (i.e., rationales) from the given input text. We conduct comparison by evaluating faithfulness-via normalized sufficiency and comprehensiveness metrics along with plausibility-by asking legal experts to evaluate extracted rationales. We further assess the feasibility of LLM-as-a-Judge using legal expert evaluation results. We show that the model's "reasons" for predicting a violation differ substantially from those of legal experts, despite highly promising quantitative analysis results and reasonable downstream classification performance. The source code of our experiments is publicly available at https://github.com/trusthlt/IntEval.

AfroScope: A Framework for Studying the Linguistic Landscape of Africa

Language Identification (LID) is the task of determining the language of a given text and is a fundamental preprocessing step that affects the reliability of downstream NLP applications. While recent work has expanded LID coverage for African languages, existing approaches remain limited in (i) the number of supported languages and (ii) their ability to make fine-grained distinctions among closely related varieties. We introduce AfroScope, a unified framework for African LID that includes AfroScope-Data, a dataset covering 713 African languages, and AfroScope-Models, a suite of strong LID models with broad language coverage. To better distinguish highly confusable languages, we propose a hierarchical classification approach that leverages Mirror-Serengeti, a specialized embedding model targeting 29 closely related or geographically proximate languages. This approach improves macro F1 by 4.55 on this confusable subset compared to our best base model. Finally, we analyze cross linguistic transfer and domain effects, offering guidance for building robust African LID systems. We position African LID as an enabling technology for large scale measurement of Africas linguistic landscape in digital text and release AfroScope-Data and AfroScope-Models publicly.

ReinPath: A Multimodal Reinforcement Learning Approach for Pathology

Interpretability is significant in computational pathology, leading to the development of multimodal information integration from histopathological image and corresponding text data.However, existing multimodal methods have limited interpretability due to the lack of high-quality dataset that support explicit reasoning and inference and simple reasoning process.To address the above problems, we introduce a novel multimodal pathology large language model with strong reasoning capabilities.To improve the generation of accurate and contextually relevant textual descriptions, we design a semantic reward strategy integrated with group relative policy optimization.We construct a high-quality pathology visual question answering (VQA) dataset, specifically designed to support complex reasoning tasks.Comprehensive experiments conducted on this dataset demonstrate that our method outperforms state-of-the-art methods, even when trained with only 20% of the data.Our method also achieves comparable performance on downstream zero-shot image classification task compared with CLIP.

SyntaxMind at BLP-2025 Task 1: Leveraging Attention Fusion of CNN and GRU for Hate Speech Detection

This paper describes our system used in the BLP-2025 Task 1: Hate Speech Detection. We participated in Subtask 1A and Subtask 1B, addressing hate speech classification in Bangla text. Our approach employs a unified architecture that integrates BanglaBERT embeddings with multiple parallel processing branches based on GRUs and CNNs, followed by attention and dense layers for final classification. The model is designed to capture both contextual semantics and local linguistic cues, enabling robust performance across subtasks. The proposed system demonstrated high competitiveness, obtaining 0.7345 micro F1-Score (2nd place) in Subtask 1A and 0.7317 micro F1-Score (5th place) in Subtask 1B.