Deeksha Varshney | ADHRI-NLP Lab

I am an Assistant Professor at the School of Artificial Intelligence and Data Science (SAIDE), Indian Institute of Technology Jodhpur (IIT Jodhpur), Rajasthan, India. I lead the ADHRI-NLP Lab (AI for Development on Human Reasoning and Impact) where we work on cutting-edge problems in Natural Language Processing and AI for Sustainable Impact. I earned my Ph.D. from Indian Institute of Technology Patna, India in 2023. I am grateful to have Prof. Asif Ekbal as my Ph.D. supervisor. My doctoral research focused on Enhancing Dialogue Generation Models using Multi-source Heterogeneous Information.

Prior to joining IIT Jodhpur, I worked as a Research Fellow at the National University of Singapore (NUS) under Prof. Gianmarco Mengaldo, focusing on Climate NLP and Sustainable Finance.

I am currently actively working on building Interpretable Models, where I am exploring Latent Subspace Geometry and Optimization to uncover the internal "architecture of ideas" within Large Language Models. Through the lens of Mechanistic Interpretability, I investigate how high-dimensional activations are structured geometrically, moving beyond traditional assumptions to explore whether neural representations form complex, non-linear manifolds. To address the resulting vulnerabilities, I am exploring the frontier of Latent Space Reasoning, a paradigm that moves beyond traditional text-based Chain-of-Thought (CoT) by investigating how models perform multi-step inference as continuous trajectories or discrete tokens within a Latent Manifold. This approach allows us to study a "pure" form of reasoning that bypasses the constraints of natural language syntax, which we can leverage for LLM Hallucination Detection and Mitigation and to develop surgical Trustworthy AI interventions. By applying Representation Engineering, I am currently working to reverse-engineer internal structures like "refusal directions" and specific "safety neurons" to strengthen model guardrails against harmful content, such as for Misogynistic content, at the representation level. Recognizing the unique needs of a linguistically diverse country like India, which encompasses over 22 official languages, I am dedicated to compressing and refining these sophisticated reasoning and safety mechanisms into compact Multilingual Language Models and Small Language Models (SLMs). I am utilizing efficiency-driven techniques like Layer Pruning and Knowledge Distillation to ensure these models remain resource-efficient without sacrificing performance. My focus is on creating high-performance tools that enable the development of localized NLP products for NLP in Education, Healthcare, and Climate, that are small enough to be accessible while remaining smart enough to be trusted by diverse communities worldwide.

My research further extends to Multimodal Data Fusion for the Healthcare domain, where I integrate diverse modalities such as Electronic Health Records (EHR) and Medical Imaging to build more comprehensive diagnostic systems. I am also developing robust Edge AI Models for specialized, high-stakes applications, including Anomaly Detection in Surveillance Videos and resource-efficient NLP solutions for Sustainability.

Through my Research Lab at IIT Jodhpur, my group actively works on building robust, interpretable, and socially impactful AI systems across multilingual and multi-domain settings.

I am actively looking for motivated students interested in joining my group as Ph.D. scholars, Research associates and Interns at IIT Jodhpur. Please feel free to email me, with a statement of purpose.

Apply now (PhD) : Deadline 20 April 2026