Bangla memes are widely used on social media to express humor and social commentary, yet computational analysis of gender bias in Bangla memes remains largely unexplored. In this work, we present a multimodal framework for detecting gender bias in Bangla memes by jointly analyzing textual and visual con tent. We construct a dataset of 6,846 Bangla and Banglish code-mixed memes annotated into three categories: male-biased, female biased, and neutral. For textual representation, we use BanglishBERT, while visual features are extracted using ConvNeXt, and the two modalities are fused for final classification. Our best-performing model, ConvNeXt + BanglishBERT, achieves accuracy of 0.67 and an F1-score of 0.63, outperforming several multimodal baselines. The results demonstrate the effectiveness of multimodal learning for understanding culturally nuanced and code-mixed meme content in low-resource languages. Code and data available at this link

Bias and fairness risks in Large Language Models (LLMs) vary substantially across deployment contexts, yet existing approaches lack systematic guidance for selecting appropriate evaluation metrics. We present a decision framework that maps LLM use cases, characterized by a model and population of prompts, to relevant bias and fairness metrics based on task type, whether prompts contain protected attribute mentions, and stakeholder priorities. Our framework addresses toxicity, stereotyping, counterfactual unfairness, and allocational harms, and introduces novel metrics based on stereotype classifiers and counterfactual adaptations of text similarity measures. We release an open-source Python library, langfair, for practical adoption. Extensive experiments on use cases across five LLMs and five prompt populations demonstrate that fairness risks cannot be reliably assessed from benchmark performance alone: results on one prompt dataset likely overstate or understate risks for another, underscoring that fairness evaluation must be grounded in the specific deployment context.

Multilingual Natural Language Understanding (NLU) systems often struggle to adapt when new languages or new semantic labels are introduced with only a few annotated examples. This challenge is particularly pronounced for low-resource languages, where limited supervision and evolving label spaces make conventional joint-label classification approaches unstable. Most existing multilingual NLU models treat each language-semantic pair as an independent class, entangling linguistic and semantic representations and hindering few-shot adaptation. We propose Dual-Axis Compositional Few-Shot Learning, a framework that explicitly factorizes the representation space into linguistic and semantic embedding axes, enabling independent modeling of language variation and domain-intent semantics. Joint representations are constructed compositionally through multiplicative interaction of axis-specific embeddings, allowing controlled adaptation when either the language set or the semantic label space evolves. The framework integrates factorized prototype learning, axis-structured contrastive alignment, and disentanglement regularization using HSIC-based statistical independence and Jacobian-based cross-axis decorrelation. Experiments on six low-resource Indic languages spanning Indo-Aryan and Dravidian families (Hindi, Bengali, Sanskrit, Assamese, Tamil, and Telugu) demonstrate strong performance under two structured generalization regimes. The model achieves 81.12% accuracy when adapting to few-shot languages with known semantics and 63.5% accuracy when learning new semantic classes from few-shot examples, along with an accuracy of 89.56% on known language and seen semantics. These results show that axis-factorized representations enable stable compositional generalization, offering a promising direction for scalable multilingual NLU in linguistically diverse low-resource settings.

Recursive prompting with large language models enables scalable synthetic dataset generation but introduces the risk of bias amplification. We investigate gender bias dynamics across three generations of recursive text generation using three complementary evaluation frameworks: rule-based pattern matching, embedding based semantic similarity, and downstream task performance. Experiments with three initial bias levels (0.1, 0.3, 0.6) and four mitigation strategies reveal equilibrium dynamics rather than monotonic amplification. The low initial bias amplifies toward the model’s inherent bias level (+ 36%), whereas the high initial bias decays toward it (-26%). Among mitigation methods, contrastive augmentation, which introduces gender-swapped variants, achieves significant downstream bias reduction (98.8% for low initial bias and 91% on average) despite producing higher embedding-based bias scores. This paradox demonstrates that semantic similarity metrics may diverge from behavioral fairness outcomes, highlighting the need for multidimensional evaluation in responsible synthetic data generation.

Automatic text simplification (ATS) aims to enhance language accessibility for various target groups, particularly persons with intellectual disabilities. Recent advancements in large language models (LLMs) have substantially improved the quality of machine-generated text simplifications, however, existing LLM-based ATS systems do not incorporate preference feedback during post-training, resulting in a lack of personalization tailored to the specific needs of target group persons. In this work, we propose an ATS personalization framework using direct preference optimization (DPO). Specifically, we post-trained LLM-based ATS models using human feedback collected from persons with intellectual disabilities, reflecting their preferences of paired text simplifications generated by mainstream LLMs. Our pipeline for developing personalized LLM-based ATS systems encompasses data collection, model selection, supervised fine-tuning (SFT) and DPO post-training, and result evaluation. Our findings underscore the necessity of active participation of target group persons in designing personalized inclusive AI solutions aligned with human preferences.

This paper presents a methodology that uses LLMs to align multilingual offensive lexicons at the sense level. Lexicons of different structures and origins in Arabic, Bulgarian, Modern Greek, French, and Italian have been aligned directly without pivoting through English. The Modern Greek lexicon is LLM-generated, and the other four lexicons are WordNet-compatible. For inter-language alignment of senses, an LLM-as-a-judge rubric was used over lemma–definition–example triples. The LLM makes 2.87M pairwise comparisons and yields 31 strict global-sense categories. The paper discusses the challenges involved in sense alignment tasks. The resource is available to support downstream applications such as Machine Translation and cross-lingual hate-speech detection.

Social media has amplified public discourse in India while perpetuating caste-based hierarchies. Despite legal protections, caste-based hate speech continues to propagate across digital platforms through culturally embedded expressions that conventional classifiers often struggle to interpret. We propose GYAAN-SAHIT, a knowledge-driven multi-agent framework that addresses this problem through structured debate-based classification. Each agent adopts a distinct ideological and socio-cultural persona, engaging in multi-turn argumentation to reason over context, subtext, and intent. A critic agent then evaluates the coherence of the debate before producing the final classification. The framework further integrates Hindi hate lexicons to ground its reasoning in linguistic and cultural specificity. Experiments show that GYAAN-SAHIT achieves improvement in performance while generating culturally grounded explanations, demonstrating the effectiveness of persona-based multi-agent reasoning for hate speech detection in low-resource and socially complex environments.

Large Language Models (LLMs) perform strongly on many language tasks, but their capability in structurally constrained, accessibility-critical modalities such as Braille remains unclear. We evaluate state-of-the-art LLMs on bidirectional Korean–Braille translation using a human-annotated dataset. Despite expectations that multilingual, instruction-tuned models can generalize to Braille via text representations, we find consistently poor, unstable outputs and substantial disagreement with human judgments. These results point to missing Braille-aware tokenization and weak alignment between Korean and Braille patterns. In contrast, supervised fine-tuning of a small model (T5-small) on the same data yields large and stable gains over zero-shot and prompted LLM baselines across standard metrics (SacreBLEU, ChrF++, CER, BLEU, ROUGE-L, METEOR, CIDEr). Our findings reveal a systematic limitation of current LLMs and demonstrate the effectiveness of modest task-specific supervision.

The rapid expansion of social media platforms has led to a significant increase in the spread of harmful content, including misogynistic, homophobic, and transphobic memes. Detecting such content is challenging because memes often combine textual and visual elements and frequently appear in multilingual and culturally diverse contexts. This study proposes a multimodal transformer-based framework for multilingual harmful meme classification that integrates textual and visual representations to improve detection performance. The proposed architecture employs XLM-RoBERTa for multilingual text encoding and the Swin Transformer for hierarchical visual feature extraction. A cross-attention fusion mechanism is introduced to enable meaningful interaction between textual and visual modalities. The fused representation is then processed through a classification layer to perform multi-class prediction. Experiments are conducted across multiple datasets covering eight languages and three harmful content categories: misogyny, homophobia/transphobia, and hate speech. The model is evaluated using the macro-F1 score and demonstrates consistent improvements over baseline multimodal systems across both high-resource and low-resource languages. The results highlight the effectiveness of transformer-based multimodal architectures in capturing implicit and contextual harmful signals present in memes. The study contributes to the development of robust multilingual systems for harmful content detection and supports efforts toward creating safer and more inclusive online environments.

While automatic text summarization has achieved remarkable success in English,extending these capabilities to low-resource languages remains a significantchallenge due to the scarcity of labeled training data. We propose atranslation-augmented approach to multilingual summarization: we systematicallytranslate high-quality English summarization corpora into low-resource targetlanguages using NLLB-200, and use the resulting parallel data to train andevaluate sequence-to-sequence models. We experiment across three typologicallydiverse languages—Swahili, Hausa, and Afrikaans—comparing monolingualfine-tuning (MONO), cross-lingual transfer (XLT), and joint multilingualtraining (TAMT) on mBART-large-50. Monolingual fine-tuning achieves the bestperformance for Swahili (ROUGE-L 13.9) and Afrikaans (ROUGE-L 15.7),surpassing the Lead-3 baseline in both cases, while cross-lingual transferremains strongest for Hausa (ROUGE-L 14.5). We show that native language tokenavailability in mBART-50 is a critical determinant of fine-tuning performance,and characterize the conditions under which the theoretically expectedTAMT > MONO > XLT ordering breaks down. We release our dataset, code, andevaluation infrastructure to support future research on low-resourcemultilingual summarization.

Online platforms continue to witness harmful expressions targeting LGBTQ+ individuals, particularly in the form of homophobic and transphobic comments. While detection of such content has received substantial attention, generating constructive counter-narratives remains comparatively underexplored. In this shared task, we focus on counter-narrative generation in English and Tamil. Participants were provided with social media comments labeled as homophobic or transphobic and were required to generate respectful, contextually appropriate responses that challenge prejudice and promote empathy. Systems were evaluated using both reference-based metrics (Distinct-2 and BERTScore-F1) and rubric-based human evaluation metrics measuring politeness (PRS), quality (QS), and contextual coherence (CCNC). The results demonstrate variation in system performance across languages, with English systems showing stronger lexical diversity and Tamil systems excelling in politeness and contextual coherence. This paper presents dataset statistics, evaluation methodology, system performance analysis, and key observations from the shared task.

We investigate the role of large language models (LLMs) in promoting gender-inclusive language by evaluating their ability to rewrite biased text and generate counterfactual narratives across multiple languages. We introduce a shared task with two subtasks: gender-inclusive rewriting and counterfactual generation. The task covers five languages English, German, Spanish, Tamil, and Kannada reflecting diverse grammatical gender systems and sociocultural contexts. We release curated word-level and sentence-level datasets to support controlled inclusive generation. A total of 50 teams registered for the shared task, and around 8 teams submitted results. Submissions are evaluated using a hybrid framework combining rubric-based automatic scoring with expert human judgment. Finally, we provide an overview of participating systems and discuss key findings and challenges observed across languages.

This paper presents an overview of the Shared Task on detecting homophobia and transphobia in meme datasets across three languages: Hindi, English, and Chinese. With the rapid growth of internet users worldwide, memes have become a widely used medium for expressing humor, satire, and sarcasm on social media platforms. However, their increasing popularity has also facilitated the spread of hate, misinformation, and propaganda targeting specific communities. Hateful memes often attack individuals or groups based on attributes such as physical appearance, language, ethnicity, religion, or sexual orientation. Among those affected, the LGBTQ+ community is particularly vulnerable and frequently targeted on social media platforms. To address this issue, we organized a shared task that focuses on identifying homophobic and transphobic hate in memes. The task aims to encourage the development of automated systems capable of detecting such harmful content across multiple languages. Evaluation was conducted using Macro F1-score as the primary metric. The top performing system achieved a Macro F1-score of 0.8377 for English, 0.8081 for Hindi, and 0.7535 for Chinese, demonstrating promising results for multilingual hate detection in memes.

Gender bias in multilingual language generation systems poses serious ethical and social issues, especially in languages with complex morphology. In this study, we propose a lightweight multilingual approach that employs instruction-guided fine-tuning of the mT5-small transformer model for gender-inclusive language generation. The framework accommodates five languages: English, German, Spanish, Tamil, and Kannada. The approach uses the task-prefix rewriting method to transform gender-specific sentences to their gender-neutral versions. The training data from different languages is combined into a single multi-lingual dataset for sequence-to-sequence fine-tuning. Beam search decoding with repetition constraints is used during inference to improve the quality of the output. The system’s performance is measured using GIFI, semantic similarity, and an overall combined score across all languages. Experimental results show that the system can eliminate gender-biased language while retaining semantic meaning in part across languages

The rapid growth of social media has also led to a rise in abusive and harmful content, which negatively affects the online environment for users. The frequent use of offensive language and hate speech contributes to making these platforms increasingly hostile. In particular, homophobic and transphobic remarks target members of the LGBT+ community. Detecting such comments is therefore essential so that they can be flagged promptly and appropriate warnings can be given to users involved in such behaviour. The problem becomes more serious when such content appears in other forms of communication used by younger generations, such as memes. This work tries to address this issue. We propose a method to detect such content using the meme dataset from the LT-EDI 2026 challenge and secured 8th rank for English and 6th rank for Chinese language dataset in the shared task. Our approach uses a multimodal technique that processes both image and text information. The dataset has limited data, which creates a challenge. To handle this, we pre–fine-tune the models on a similar dataset called PrideMM. The proposed multimodal approach achieved Macro F1-scores of 0.24 and 0.57 for English and Chinese memes respectively.

Online hate speech is spreading rapidly, creating significant challenge, particularly in low-resource language such as Tamil. Lack of developed automated content moderation systems makes it difficult to control harmful content effectively. In this study, we propose a computational framework for generating Counter Narratives (CNs) using classical NLP techniques. With this, we leverage TF-IDF features with n-grams to identify the labels as Homophobic or Transphobic. Span detection is performed with TF-IDF features with n-grams and Machine learning models. Counter narratives are then retrieved by computing cosine similarity, ensuring semantic alignment and contextual relevance. Evaluation on the expanded human curated dataset demonstrates that our approach produces contextually appropriate and semantically coherent counter narratives. Notably, the proposed system is submitted at Task 2 shown a overall average score of 80.40 % for Tamil and 77.29 % for English and secured first and fourth rank respectively. GitHub: https://github.com/kannanrrk/Span-Counter-Feature-Based

The detection and response to homophobicand transphobic comments are important challengesin Natural Language Processing. In thispaper, we focus on the detection of span forhomophobic and transphobic comments (Task1) and generation of counter narratives for abusivecomments (Task 2) for the LT-EDI @ ACL2026 shared task. Harmful comments madeonline against the LGBTQ+ community havecreated a hostile environment for users. In thispaper, we have used the transformer model forthe detection of span for homophobic and transphobiccomments and generation of counternarratives. In this task, the detection of the spanof comments containing homophobic and transphobicwords and the generation of counter narrativesfor abusive comments have been doneusing the transformer model. The results showthe efficiency of the transformer model in thedetection of the span of comments and generationof counter narratives. This paper emphasizesthe efficiency of the transformer model increating a safe environment for users.

The deployment of Large Language Models(LLMs) has intensified concerns regarding thepropagation of societal stereotypes encodedwith web-scale training corpora. This pa-per presents a dual-paradigm framework spe-cially designed to address multilingual gender-inclusvity and counterfactual generation. Formultilingual gender-neutral text transformation,a fine-tuned mT5 encoder–decoder model per-forms controlled sentence rewriting with mini-mal edits while preserving semantic fidelity andgrammatical fluency. For counter-narrative gen-eration, the Llama-3 8B decoder-only model isemployed to generate empathetic and persua-sive responses through structured prompt-basedgeneration. The framework is evaluated usingdatasets from the LT-EDI ACL 2026 sharedtask across multiple languages, including En-glish, Tamil, Kannada, German, and Spanish.Experimental results demonstrate strong effec-tiveness in identifying and neutralizing gendermarkers, particularly in morphologically richlanguages, while the counter-narrative compo-nent achieves high performance in politeness,coherence, and relevance. Overall, the pro-posed approach contributes toward the develop-ment of responsible and inclusive multilingualNLP systems.

We describe the IHLC team’s submissionto the LT-EDI ACL 2026 Shared Task onGender-Inclusive Language Generation andCounterfactual/Counter Narrative Generation.Our English-only system applies an activationsteeringapproach combined with carefullyengineered prompt templates to producegender-neutral rewrites and empathetic counternarratives.We summarize system design, experimentalsetup, evaluation protocol used bythe shared task, and report results for both subtasks(Task A: Gender Inclusive Language Generation– Average = 80.00%, Rank 3; TaskB: Counter Narrative Generation – Average =78.12%, Rank 6). We also analyze strengthsand failure modes observed in automatic andhuman-checked evaluations and highlight directionsfor improvement.

This paper presents our submissions to the LT-EDI@ACL 2026 Shared Task on Gender Inclusive Language Generation. The task focuses on controlled text rewriting that reduces gender bias while keeping the original meaning and fluency intact. We participated in boththe subtasks and treated them independently, training separate instances of the instruction-tuned encoder–decoder model on the respective training datasets. Scores are calculated based on averages across different rubrics, including Gender Assumption (GA), Gender Neutrality (GN), and Quality Relevance (QR) for Task A, and Politeness and Respectful (PR), Contextual Counter-Narrative Coherence (CCNC), and Quality and Relevance (QR) for Task B.For Subtask A (Gender-Inclusive Language Generation) in the English dataset, an average score of 43.7917 could be achieved. For Subtask B (Counterfactual Generation), we achieved an average score of 82.6241. Overall, the experiments indicate that full finetuning of instruction-tuned transformers provides an effective way to produce sentence in gender-neutral form and also producing counter-factual sentences for biased one, wheneach subtask is optimized on its own data.

Online platforms increasingly host hate speechtargeting marginalized communities, includ-ing homophobic and transphobic commentsdirected at LGBTQ+ individuals. Counter-narratives provide a constructive way to re-spond to harmful speech by promoting em-pathy, factual clarification, and respectful di-alogue.In this work, we participate in the Shared Taskon Counter-Narrative Generation on Homopho-bic and Transphobic Comments at LT-EDI @ACL 2026. We adopt a zero-shot promptingapproach using large language models accessedthrough publicly available AI tools, includingGPT-4o, Gemini 1.5 Pro, and Llama-3 SonarLarge via Perplexity AI. Instead of traininga task-specific model, we design a structuredprompt that guides the models to generate re-spectful, concise, and contextually appropriatecounter-narratives.Experiments were conducted on English andTamil comments provided by the organiz-ers. Results demonstrate that prompt-basedgeneration can produce meaningful multilin-gual counter-narratives without additional train-ing. Our approach highlights the potential oflarge language models as lightweight tools forcounter-speech generation in multilingual on-line environments.

Over the past decade, the rapid advancement of LLMs has significantly improved natural language generation. However, these models often inherit and amplify gender biases present in large-scale training data, leading to stereotypical associations, androcentric language, and misgendering. Such biases can negatively impact applications in education, healthcare, legal systems, and automated content generation. In this paper, we address this issue as defined in the shared task LT-EDI on Gender-Inclusive Language Generation. The task focuses on rewriting gender-biased sentences into inclusive, gender-neutral alternatives while preserving meaning. We propose a retrieval-augmented framework combining lexical replacement, semantic retrieval, and controlled instruction-tuned generation. An edit-distance constraint and self-evaluation step ensure minimal, coherent, and bias-free outputs. We also present zero-shot adaptation for low resource language. The implementation code available here https://github.com/SupriyaChanda/gilg-ltedi-acl2026.git.

Detecting hate speech in memes is challenging due to their multimodal nature and subtle, culturally grounded cues such as sarcasm and context. While recent vision-language models (VLMs) enable joint reasoning over text and images, end-to-end prompting can be brittle, as a single prediction must resolve target, stance, implicitness, and irony. These challenges are amplified in multilingual settings. We propose a prompted weak supervision (PWS) approach that decomposes meme understanding into targeted, question-based labeling functions with constrained answer options for homophobia and transphobia detection in the LT-EDI 2026 shared task. Using a quantized Qwen3-VLM to extract features by answering targeted questions, our method outperforms direct VLM classification, with substantial gains for Chinese and Hindi, ranking 1st in English, 2nd in Chinese, and 3rd in Hindi. Iterative refinement via error-driven LF expansion and feature pruning reduces redundancy and improves generalization. Our results highlight the effectiveness of prompted weak supervision for multilingual multimodal hate speech detection.

Counter Narrative (CN) generation via Large Language Models (LLMs) offers a scalable approach to combating hate speech by producing targeted responses that challenge harmful content. However, existing methods typically require costly post-training or fine-tuning to improve narrative diversity and quality. We introduce a fine-tuning-free prompt optimization technique that enhances Counter Narrative effectiveness without additional model training, making it both resource-efficient and readily deployable. We conduct extensive evaluation on hate speech datasets spanning English and Tamil, employing both reference-based metrics and rubric-based LLM-as-a-judge scoring to capture multiple dimensions of narrative quality. Experiments across multiple LLMs demonstrate that our approach consistently outperforms vanilla prompting baselines, exhibits strong transferability across models, and adapts seamlessly to new evaluation metrics—requiring no architectural or procedural changes. Our findings suggest that carefully optimized prompting strategies can match or exceed the performance of more resource-intensive approaches, offering a practical path toward scalable hate speech intervention.

The problem of harmful online discourse against the LGBTQ+ community is still a concern on social media platforms. Although hate speech detection is a well-explored area, the task of constructive counter-narrative generation is still an emerging field of research, especially in the multilingual and low-resource settings. Counter-narratives are designed to counter harmful discourse with respectful and empathetic responses, as opposed to mere content deletion. In this paper, the model proposes a zero-shot multilingual system for counter-narrative generation in English and Tamil. The proposed system employs the pretrained google/flan-t5-base transformer model guided by rubric-aligned prompts to encourage politeness, contextual relevance, and non-toxic response generation. The system operates in a zero-shot setting without task-specific fine-tuning and uses beam search decoding for controlled response generation. On the English test data, the system scored an overall score of 70.33 per cent with a contextual coherence score of 81.82 per cent. On the Tamil test data, the system scored an overall score of 33.57 per cent with significantly lower scores on coherence and quality. These findings indicate that structured prompting can facilitate safe and coherent generation in English, but also underscore the challenges of zero-shot multilingual models in low-resource language scenarios.

Social media is now an important platform for communication and interaction. At the same time, the amount of abusive and harmful content online has also increased. Offensive language and hate speech are making these platforms less safe and less welcoming for users. Many of these contents include homophobic and transphobic remarks aimed at the LGBT+ community. Such behaviour damages healthy discussions and can negatively affect users. For this reason, it is important to detect these contents early so they can be flagged and removed to maintain a healthy online well-being. The issue becomes more difficult when harmful messages appear in popular formats like memes. Memes are widely used by younger users to communicate online. Because they combine images and text, detecting offensive meaning becomes challenging. In this work, we attempt to address this problem. We develop a method to identify such content using the meme dataset released for the LT-EDI 2026 challenge and secured rank 5 in the shared task. We propose a Zero-shot learning based method employing two LLMs (Qwen2.5-VL-3B-Instruct and Meta-Llama-3-8B-Instruct) to generate descriptions and classify such memes. We achieved a macro F1-score of 0.55 for the English language meme.

The automated detection of LGBTQ+ phobia in social media memes is essential for fostering inclusive digital environments, yet it remains challenging due to the complex interplay of visual metaphors and multilingual text. We participated in the "Homophobia and Transphobia Meme Classification" shared task at LT-EDI 2026, evaluating a multimodal architecture across English, Hindi, and Chinese tracks. Our system employs a late-fusion strategy: XLM-RoBERTa encodes OCR-extracted text into a representation h_t ∈ ℝ⁷⁶⁸ , while CLIP extracts visual features h_v ∈ ℝ⁵¹². These are concatenated into a joint vector z = [h_t ⊕ h_v] ∈ ℝ¹²⁸⁰ and processed via a non-linear multilayer perceptron to capture cross-modal interactions.The system demonstrated robust performance in high-resource contexts, securing 3rd rank in the Chinese track (Macro F1: 0.7371) and 4th rank in the English track (Macro F1: 0.6121). In contrast, the Hindi track results (Macro F1: 0.1616) revealed significant challenges related to script complexity and class imbalance. These findings underscore the effectiveness of global transformer-based models for multimodal reasoning while highlighting the ongoing need for specialized linguistic refinement in low-resource and diverse script environments

This paper presents our system for the LT-EDI@ACL 2026 workshop on meme classification of homophobia and transphobia in English, Hindi, and Chinese. Detecting harmful content in memes is challenging because meaning often emerges from the interaction between visual elements and short textual cues, particularly in multilingual settings. To address this, we build a multimodal pipeline using CLIP ViT-L/14 visual embeddings, EasyOCR text extraction, TF–IDF lexical features, and a multinomial logistic regression classifier. We further incorporate two optional expert modules, a LoRA-adapted Qwen2-VL model and a CLIP zero-shot classifier, and combine predictions using weighted majority voting. The system is intentionally lightweight and reproducible, demonstrating that strong pretrained transfer features paired with explicit OCR can provide robust multilingual meme moderation without extensive fine-tuning. On the official leaderboard, our submission ranks 1st in Hindi, 3rd in English, and 5th in Chinese.

We present our approach to LT-EDI@ACL 2026 on counter-narrative generation for homophobic and transphobic comments. Generating high-quality counter-narratives in multilingual and low-resource settings remains challenging, particularly when data imbalance and script variation affect model performance. To address these issues, we explore multiple modeling strategies built around Gemma 3 12B with QLoRA fine-tuning, including data rebalancing and alternative input strategies for Tamil. Our findings show that task-specific fine-tuning combined with native-script Tamil produces more stable and higher-quality outputs than large few-shot prompts or transliteration-basedinputs. On the official leaderboard, our system ranks second in English with an overall score of 86.35% and sixth in Tamil with 63.77%,highlighting both the effectiveness of targeted fine-tuning and the challenges of low-resource counter-narrative generation.

This paper describes the system developed byTeamV for the LT-EDI 2026 Shared Task onCounter-Narrative Generation on Homophobic Transphobic Comments. The shared taskcomprises two subtasks: (1) Hate Speech SpanDetection in English, Tamil, and Hindi, and (2)Counter-Narrative Generation in English andTamil. Our system leverages the reasoning andmultilingual capabilities of a large proprietarylanguage model (Qwen3-Max) through rigor-ous few-shot in-context learning (ICL) and ro-bust post-processing mechanisms. Our submit-ted system demonstrated state-of-the-art perfor-mance on the official CodaBench leaderboard.In Task 1, our approach achieved 1st Placeacross all three languages, securing macro F1scores of 0.5338 in English, 0.5272 in Tamil,and 0.5478 in Hindi. For Task 2, our generatedcounter-narratives ranked 1st globally in En-glish with an overall average score of 87.47%and 5th in Tamil. We present our promptingmethodology, robust span-matching pipeline,detailed official results, and an analysis of themodel’s performance across diverse languages.