Predicting continuous Valence and Arousal scores across diverse languages poses significant challenges due to typological differences and the difficulty of modeling affective intensity. We introduce AdaptStance, a parameter-efficient framework designed for the SemEval-2026 Task 3 benchmark. To address cross-lingual disparities, AdaptStance routes inputs through resource-specific pipelines: direct regression with a hybrid concordance loss for high-resource languages, and an auxiliary multi-task mechanism to stabilize regression in low-resource and non-Western contexts. Architectural analysis reveals that decoupling task heads benefits morphologically related languages, whereas joint representations act as crucial regularizers for distant language families. Ultimately, this lightweight approach achieves competitive performance over generative baselines, demonstrating the efficacy of targeted architectural alignment while identifying Valence as the primary bottleneck in continuous affect prediction. Our code is available on GitHub.

The growing adoption of large language models for code generation poses challenges for code quality, security, and authorship verification—particularly when test conditions involve unseen programming languages, generators, or application domains. We present our system, which combines three code-pretrained transformer encoders (CodeT5p-220M, CodeBERT, UniXcoder) with a structure-first Flow-Augmented AST (FA-AST) encoder implemented as a Gated Graph Neural Network. On Subtask A our best single model achieves macro F1 of 0.559; a post-competition layered rank-fusion ensemble across all three encoders raises this to 0.643. On Subtask C we obtain 0.585 officially; a three-stage ensemble combining neural probabilities with LightGBM-based features and class-priority routing raises this to 0.652. Our contributions include a language-agnostic structural detector, a diversity-driven rank-fusion strategy exploiting low inter-model correlation for binary classification, and a meta-learner stacking pipeline for multi-class detection under distribution shift.

While Automatic Speech Recognition (ASR) systems have shown impressive performance in languages having sufficient annotated speech data like English, their performance is still limited for low-resource, dialect rich languages like Tamil. Tamil poses further challenges because of its extremely high regional variation in dialects that manifest in varying vocabulary, pronunciations, and even syntactic structures. To address these challenges, we present a unified framework WhisTam based on the Whisper medium model, which performs speech transcription and dialect classification jointly within a single system. Our method is evaluated against speech samples from four regional dialects and achieves a macro F1-score of 0.53 and a Word Error Rate (WER) of 0.55 for dialect classification and transcription respectively, ranking 2nd in the dialect classification task and 3rd in the transcription task in the DravidianLangTech@ACL 2026 shared task on Dialect-based Speech Recognition and Classification in Tamil. These findings emphasize the challenges in dialectal Tamil ASR as well as the promise of multi-task learning for low-resource languages. Our implementation is publicly available at: https://github.com/rwd51/DravidianLangTech-Wave2Word.

Conventional research on speech recognition modeling relies on the canonical form for most low-resource languages while automatic speech recognition (ASR) for regional dialects is treated as a fine-tuning task. To investigate the effects of dialectal variations on ASR we develop a 78-hour annotated Bengali Speech-to-Text (STT) corpus named Ben-10. Investigation from linguistic and data-driven perspectives shows that speech foundation models struggle heavily in regional dialect ASR, both in zero-shot and fine-tuned settings. We observe that all deep learning methods struggle to model speech data under dialectal variations, but dialect specific model training alleviates the issue. Our dataset also serves as a out-of-distribution (OOD) resource for ASR modeling under constrained resources in ASR algorithms. The dataset and code developed for this project are publicly available.

Each new generation of English-oriented Large Language Models (LLMs) exhibits enhanced cross-lingual transfer capabilities and significantly outperforms older LLMs on low-resource languages. This prompts the question: Is there a need for LLMs dedicated to a particular low-resource language? We aim to explore this question for Bengali, a low-to-moderate resource Indo-Aryan language native to the Bengal region of South Asia. We compare the performance of open-weight and closed-source LLMs such as LLaMA-3 and GPT-4 against fine-tuned encoder-decoder models across a diverse set of Bengali downstream tasks, including translation, summarization, paraphrasing, question-answering, and natural language inference. Our findings reveal that while LLMs generally excel in reasoning tasks, their performance in tasks requiring Bengali script generation is inconsistent. Key challenges include inefficient tokenization of Bengali script by existing LLMs, leading to increased computational costs and potential performance degradation. Additionally, we highlight biases in machine-translated datasets commonly used for Bengali NLP tasks. We conclude that there is a significant need for a Bengali-oriented LLM, but the field currently lacks the high-quality pretraining and instruction-tuning datasets necessary to develop a highly effective model.