Mental manipulation is a subtle yet pervasive form of abuse in interpersonal communication, making its detection critical for safeguarding potential victims. However, due to manipulation’s nuanced and context-specific nature, identifying manipulative language in complex, multi-turn, and multi-person conversations remains a significant challenge for large language models (LLMs). To address this gap, we introduce the MultiManip dataset, comprising 220 multi-turn, multi-person dialogues balanced between manipulative and non-manipulative interactions, all drawn from reality shows that mimic real-world scenarios. For manipulative interactions, it includes 11 distinct manipulations depicting real-life scenarios. We conduct extensive evaluations of state-of-the-art LLMs, such as GPT-4o and Llama-3.1-8B, employing various prompting strategies. Despite their capabilities, these models often struggle to detect manipulation effectively. To overcome this limitation, we propose SELF-PERCEPT, a novel, two-stage prompting framework inspired by Self-Perception Theory, demonstrating strong performance in detecting multi-person, multi-turn mental manipulation. Our code and data are publicly available at https://github.com/danushkhanna/self-percept .

SemEval SubtaskB, a shared task that is concerned with the detection of text generated by one out of the 5 different models - davinci, bloomz, chatGPT, cohere and dolly. This is an important task considering the boom of generative models in the current day scenario and their ability to draft mails, formal documents, write and qualify exams and many more which keep evolving every passing day. The purpose of classifying text as generated by which pre-trained model helps in analyzing how each of the training data has affected the ability of the model in performing a certain given task. In the proposed approach, data augmentation was done in order to handle lengthier sentences and also labelling them with the same parent label. Upon the augmented data three RoBERTa models were trained on different segments of data which were then ensembled using a voting classifier based on their R2 score to achieve a higher accuracy than the individual models itself. The proposed model achieved an overall validation accuracy of 97.05% and testing accuracy of 76.25%. and our standing was 18th position on the leaderboard.

Large Language Models (LLMs) have significant potential for facilitating intelligent end-user applications in healthcare. However, hallucinations remain an inherent problem with LLMs, making it crucial to address this issue with extensive medical knowledge and data. In this work, we propose a Retrieve-and-Medically-Augmented-Generation with Knowledge Reduction (ReMAG-KR) pipeline, employing a carefully curated knowledge base using cross-encoder re-ranking strategies. The pipeline is tested on medical MCQ-based QA datasets as well as general QA datasets. It was observed that when the knowledge base is reduced, the model’s performance decreases by 2-8%, while the inference time improves by 47%.