Large language models increasingly generate high-quality source code, making reliable detection of machine-generated code essential for maintaining authorship integrity and software accountability. However, detection systems often degrade under distribution shift, particularly across programming languages and application domains. SemEval-2026 Task 13 Subtask A addresses this challenge through a structured OOD evaluation framework that assesses binary machine-generated code detection across unseen languages and application domains. To mitigate this limitation,we propose a robustness-oriented framework that enhances feature-fused UniXcoder representations with supervised contrastive learning, adversarial language-invariant training and uncertainty-aware filtering to promote stable and shift-resilient representations. Our proposed system achieves a macro-F1 of 0.5411 on the official test set and maintains stable performance under severe language–domain shift. Our results demonstrate that domain-level semantic variation is the primary source of degradation under distribution shift, reinforcing the importance of invariance-oriented representations for stable OOD performance

Determining whether large language models (LLMs) perform genuine formal reasoning or rely on semantic heuristics is a key challenge in NLP. Syllogistic reasoning constitutes a theoretically principled evaluation paradigm where validity is fully determined by quantifier structure, allowing systematic analysis of structural inference disentangled from semantic plausibility.SemEval-2026 Task-11, Subtask-1: Disentangling Content and Formal Reasoning in Language Models, establishes a multilingual benchmark designed to rigorously isolate formal logical validity from semantic plausibility effects. The subtask evaluates English syllogistic reasoning under a binary classification setting using Overall Accuracy (ACC) and Total Content Effect (TCE), where lower TCE indicates stronger resistance to content-induced bias.Our proposed approach combines cross-validation, structured aggregation and bias-aware evaluation to optimize the robustness–performance trade-off. It achieves 93.19\% accuracy with a TCE of 3.13, yielding a strong combined score of 38.56 under the official evaluation metric. Condition-wise and multi-run analysis confirms that robustness-focused optimization curbs content-driven errors, reinforcing the necessity of bias-aware training for formal inference

Vaccine-critical memes have emerged as a growing challenge for public health communication, combining images and text to spread misinformation in ways that are difficult to detect automatically. In this paper, we have described our system for the EEUCA 2026 Shared Task on Multimodal Vaccine-Critical Meme Detection, classifying memes from the VaxMeme dataset into Vaccine-Critical, Neutral and Pro-Vaccine categories. We have experimented with multiple text encoders and visual backbones, finding that Twitter-RoBERTa fused with CLIP ViT-L/14 through gated cross-modal attention has achieved a test macro F1 of 0.8357. We have further shown that domain-specific pretraining has outperformed larger general-purpose models, highlighting the importance of domain adaptation over raw model scale. Finally, our system has secured the 3rd position on the shared task leaderboard.

Recovering writing style prompts in low resource languages has been daunting due to diverse morphology, culturally cognizant language patterns and deficient annotated resources. As previous works have predominantly focused on binary sentiment or single attribute transfer, extensive multi-class style classification in under-resourced languages like Telegu has been vastly underexplored. In this study, we have addressed this chasm through the Telugu Prompt-Style Recovery Shared Task at DravidianLangTech@ACL 2026 (Premjith et al., 2026), framing prompt reconstruction as a nine-class classification problem with Formal, Informal, Optimistic, Pessimistic, Humorous, Serious, Inspiring, Authoritative and Persuasive as prompt styles. We have evaluated three input configurations—Change Style, Original Transcripts and Merged input style—while training three transformer based models-MuRIL, XLM-RoBERTa and IndicBERT v2 under identical conditions. Our most promising model, IndicBERT v2 with partial layer freezing and weighted cross-entropy loss, has obtained a macro-F1 of 0.2987 and accuracy of 0.299. The Change Style configuration has significantly outperformed Original and Merged inputs, indicating that explicit style changes have made tonal and meaning cues more distinctive. These results have showcased the importance of language-specific pretraining and careful input design for style-sensitive NLP in low-resource settings, ultimately securing 1st rank on the shared task.

Hope speech has played a vital role in online communities, yet most NLP work has focused on English and a few high-resource languages, leaving code-mixed varieties like Tulu largely unexplored. In the Shared Task on Hope Speech Detection in Code-Mixed Tulu at DravidianLangTech@ACL 2026, we have tackled two subtasks: (i) coarse-grained classification into Encouraging, Discouraging, Uninvolved and Blended categories (Task 1) and (ii) fine-grained classification into Optimistic, Realistic, Inspiring, Fading and Hopelessness (Task 2).We have fine-tuned three multilingual transformer encoders XLM-RoBERTa-base, MuRIL and mBERT on the official training splits. In Task 1, a three-way soft-voting ensemble of all three models has yielded the best performance with a macro F1 of 0.58, securing 1st place. In Task 2, XLM-RoBERTa-base alone has outperformed both MuRIL and mBERT, achieving a macro F1 of 0.42 and also securing 1st place.

Abusive language targeting women has been a serious problem on Tamil social media and building systems to detect it automatically is harder than it looks. Tamil is morphologically complex, people have written it mixed with English in ways no dictionary has accounted for and a lot of the hostility has been indirect enough that has slipped past models trained on surface patterns. In the Shared Task on Abusive Tamil Text Targeting Women on Social Media DravidianLangTech@ACL 2026, we have worked on classifying Tamil YouTube comments as Abusive or Non-Abusive. We have trained three transformer models four times each with different learning rates, giving us 12 models total. Their predicted probabilities have been averaged to make the final decision. The 12-model ensemble has achieved a macro F1 of 0.8086, outperforming all individual models and securing 4th place in the shared task. Combining Tamil-specialized and multilingual transformer models has outperformed any single-architecture approach.

Memes, originally crafted for humor or cultural commentary, have evolved into powerful tools for spreading harmful content, particularly misogynistic ideologies. These memes sustain damaging gender stereotypes, further entrenching social inequality and encouraging toxic behavior across online platforms. While progress has been made in detecting harmful memes in English, identifying misogynistic content in Chinese remains challenging due to the language’s complexities and cultural subtleties. The multimodal nature of memes, combining text and images, adds to the detection difficulty. In the LT-EDI@LDK 2025 Shared Task on Misogyny Meme Detection, we have focused on analyzing both text and image elements to identify misogynistic content in Chinese memes. For text-based models, we have experimented with Chinese BERT, XLM-RoBERTa and DistilBERT, with Chinese BERT yielding the highest performance, achieving an F1 score of 0.86. In terms of image models, VGG16 outperformed ResNet and ViT, also achieving an F1 score of 0.85. Among all model combinations, the integration of Chinese BERT with VGG16 emerged as the most impactful, delivering superior performance, highlighting the benefit of a multimodal approach. By exploiting these two modalities, our model has effectively captured the subtle details present in memes, improving its ability to accurately detect misogynistic content. This approach has resulted in a macro F1 score of 0.90355, securing 3rd rank in the task.

Ensuring a safe and inclusive online environment requires effective hate speech detection on social media. While detection systems have significantly advanced for English, many regional languages, including Malayalam, Tamil and Telugu, remain underrepresented, creating challenges in identifying harmful content accurately. These languages present unique challenges due to their complex grammar, diverse dialects, and frequent code-mixing with English. The rise of multimodal content, including text and audio, adds further complexity to detection tasks. The shared task “Multimodal Hate Speech Detection in Dravidian Languages: DravidianLangTech@NAACL 2025” has aimed to address these challenges. A Youtube-sourced dataset has been provided, labeled into five categories: Gender (G), Political (P), Religious (R), Personal Defamation (C) and Non-Hate (NH). In our approach, we have used mBERT, T5 for text and Wav2Vec2 and Whisper for audio. T5 has performed poorly compared to mBERT, which has achieved the highest F1 scores on the test dataset. For audio, Wav2Vec2 has been chosen over Whisper because it processes raw audio effectively using self-supervised learning. In the hate speech detection task, we have achieved a macro F1 score of 0.2005 for Malayalam, ranking 15th in this task, 0.1356 for Tamil and 0.1465 for Telugu, with both ranking 16th in this task.