We present NarSiL (Narrative Similarity Learners), our system for SemEval-2026 Task 4 Track A on Narrative Story Similarity. NarSiL employs a two-stage architecture: a Mixture-of-Experts (MoE) initial classifier that also leverages supermajority voting across three large language models (Gemma-3-12B, GPT-3.5-turbo-instruct, and Gemini-2.5-Flash) over multiple runs, followed by a structured three-pathway fallback for ambiguous cases. The three pathways correspond directly to the task’s three core similarity components, abstract theme, narrative outcome, and course of action. Each path yields a similarity score corresponding to its respective component, and the scores are then combined through a weighted aggregation step. NarSiL achieves 64.25% accuracy on the official test set. An improved score of 70.25% is obtained by considering only the supermajority voting of GPT, followed by the previously described fallback.

Large Language Models (LLMs) have demonstrated impressive results in natural language processing (NLP) tasks, however, their brittleness against subtle input perturbations continues to pose a significant challenge. Existing research on robustness has predominantly focused on standard text-based perturbations and the use of invisible characters and homoglyphs, while overlooking the impact of stylized characters increasingly prevalent on social media. To address this, we propose TinyAttack, a novel adversarial attack framework designed to exploit vulnerabilities in LLMs through Unicode-based stylistic transformations. TinyAttack utilises five Unicode variants to modify the visual rendering of text without altering its underlying semantic or syntactic structure. Our comprehensive evaluation on both open-source (Llama, Mistral, Gemma, Qwen) and closed-source LLMs (Gemini, GPT) demonstrates their susceptibility to these stylized inputs, with performance drops ranging from 29-92% and 6-88.5%, respectively, across all tasks.Our code is available at https://github.com/TRAI-group/TinyAttack.

Stance classification is a core task in argument mining and subjectivity analysis, crucial for understanding public discourse and opinion dynamics on social media. Despite their impressive few-shot capabilities, Large Language Models (LLMs) remain sensitive to prompt construction, including the selection and ordering of in-context examples. In this paper, we propose a Topic-Guided prompting method for argument stance classification that dynamically integrates topic-specific information into the few-shot context. We evaluate our method on five LLMs across three datasets spanning formal debates and user-generated online comments. Our extensive evaluation shows that our proposed Topic-Guided prompting outperforms standard few-shot prompting and state-of-the-art example selection strategies. Further analysis indicates that our method reduces the bias towards the ’support’ class observed in several models, resulting in more balanced predictions across stances and thus a more robust approach to stance classification.

The highly formal and negotiated language of United Nations (UN) resolutions presents unique challenges for argument mining. This paper describes our system submitted to the ArgMining 2026 Shared Task: Reconstructing the Reasoning in United Nations Resolutions. Adhering to the strict constraint of utilising open-weight models with at most 8 billion parameters, we propose a hybrid, compute-efficient architecture powered by Qwen3-8B. For the preambular-operative classification, we implement a set of deterministic rules related to the specificity of UN documents, supplemented by an LLM-based multi-label classifier for thematic dimensions and a directed-graph extraction approach for argumentative relation prediction.