Large language models (LLMs) have rapidly become primary tools for searching and generating information given a carefully designed prompt (may contain few-shot examples). However, these models frequently produce factually incorrect statements that are not consistent with verifiable facts and reliable sources, raising fundamental questions about how these models store, update, and reason with facts. Improving factuality, therefore, requires more than surface-level mitigation strategies: it demands a deeper understanding of how LLMs construct and maintain world models, and how reasoning processes can be guided to remain faithful to the verifiable information. Existing strategies, such as retrieval-augmented generation, training-time alignment, post hoc verification, etc., partly address these challenges but do not provide a holistic account of how facts are internally stored, updated, or grounded in external knowledge sources. My research addresses this gap by studying factuality through the dual lens of reasoning and world modeling, asking how LLMs encode facts, how adversarial or linguistic perturbations compromise factual reasoning, and how interpretability tools can reveal and correct model vulnerabilities. In this work, I aim to develop a framework in which an LLM interacts with an explicit external knowledge source, thereby forming a robust world model for factual evaluation.

Real-world conversations are rich with pragmatic elements, such as entity mentions, references, and implicatures. Understanding such nuances is a requirement for successful natural communication, and often requires building a local _world model_ which encodes such elements and captures the dynamics of their evolving states. However, it is not well-understood whether language models (LMs) construct or maintain a robust implicit representation of conversations. In this work, we evaluate the ability of LMs to encode and update their internal world model in dyadic conversations and test their _malleability_ under linguistic alterations. To facilitate this, we apply seven minimal linguistic alterations to conversations sourced from popular conversational QA datasets and construct a benchmark with two variants (i.e., Manual and Synthetic) comprising yes-no questions. We evaluate nine open and one closed source LMs and observe that they struggle to maintain robust accuracy. Our analysis unveils that LMs struggle to memorize crucial details, such as tracking entities under linguistic alterations to conversations. We then propose a dual-perspective interpretability framework which identifies transformer layers that are _useful_ or _harmful_ and highlights linguistic alterations most influenced by harmful layers, typically due to encoding spurious signals or relying on shortcuts. Inspired by these insights, we propose two layer-regularization based fine-tuning strategies that suppress the effect of the harmful layers.

Saurabh Kumar Pandey, Sachin Vashistha, Debrup Das, Somak Aditya, Monojit Choudhury. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Recent explorations with commercial Large Language Models (LLMs) have shown that non-expert users can jailbreak LLMs by simply manipulating their prompts; resulting in degenerate output behavior, privacy and security breaches, offensive outputs, and violations of content regulator policies. Limited studies have been conducted to formalize and analyze these attacks and their mitigations. We bridge this gap by proposing a formalism and a taxonomy of known (and possible) jailbreaks. We survey existing jailbreak methods and their effectiveness on open-source and commercial LLMs (such as GPT-based models, OPT, BLOOM, and FLAN-T5-XXL). We further discuss the challenges of jailbreak detection in terms of their effectiveness against known attacks. For further analysis, we release a dataset of model outputs across 3700 jailbreak prompts over 4 tasks.

The advent of internet has led to prodigious growth in popularity of social media platforms for people to communicate and opinionate on topics of their interests. And, terroristic events being a topic of national importance, receives enormous response from the citizens. Unfortunately, miscreants with anti-national agendas incite the large available audience on these platforms against the country by inducing anti-national content amid terrorist attacks. The social media platforms being of informal use are commonly observed to have users opinionating using multiple languages in same sentence called code-mixing. The over-arching goal of research done is to identify anti-national code-mix textual content on YouTube in the form of comments on terrorism-related videos. We collected YouTube comments on videos related to terroristic events in Kashmir region of India, which consisted of code-mix comments in Hindi (native language of India) and English languages. The paper presents a novel deep-learning-based transformer model HE-CM-BERT, i.e. Hindi-English code-mix BERT, where we extend the vocabulary of pre-trained multilingual BERT with code-mix vocabulary extracted from the collected data to automate the detection of anti-national code-mix text. The comparative analysis of the proposed model with the state-of-the-art machine learning and deep learning models depicts that it outperforms the existing ones.