About me

Hi, I’m Avani Gupta.

I draw inspiration from human learning to build efficient (both sample-efficient and generalizable), interpretable, and aligned AI systems. My research explores how abstract conceptual knowledge and social learning, two foundations of human intelligence, can be translated into autonomous agents. I am particularly interested in training paradigms for generalized open-ended learning that enable common-sense understanding of the world, while ensuring transparency, safety, and alignment with human values.

Currently, I am working as an AI Engineer in the MBZUAI Research Office, where I translate industry problems into academic research, bridging my industry experience with academic research rigor. I’m also working with Prof. Abdalla Swikir on multi-agent reinforcement learning with a focus on understanding how explicit peer learning can give rise to emergent pro-social behaviors.

Here is my CV if you need one: Curriculum Vitae

Masters thesis: Leveraging Human-Centered Explanations for Model Improvement and Evaluation (thesis link)

Publications: Please see my publications page or Google Scholar for a full, up-to-date list.

Research Journey: I began my research at the Centre for Visual Information Technology (CVIT), where I completed my B.Tech (honours) and Master’s by Research under my advisor Prof. P. J. Narayanan. I worked on human-centered abstract concept-based learning for introducing knowledge priors and interpretable control in neural networks, with applications in debiasing and removing spurious correlations (NeurIPS 2023). During my internships at IBM Research, I worked with Dr. Renuka Sindhgatta on business process automation, where I used reinforcement learning to model next-best-action recommendations for complex organizational processes. In my industry research roles, I worked across the full LLM training lifecycle in the clinical domain with Dr. Shadab Khan and Dr. Marco Pimentel, as well as on code/tools-specific LLMs. I emphasised transparency and trust in my works: from analysing clinical coherence of patient embeddings from my trained foundation models to open-source releases such as Med42, a 70B clinical LLM and papers on our work. I also built production AI assistants serving thousands of users, ensuring safety through custom moderation and jailbreak prevention.

Outside work, I maintain an active lifestyle with regular workouts and enjoy reading philosophy, scientific literature, and watching documentaries. I find flow states (or in the “zone” state) while jotting down new research ideas, practicing Kathak, and creating fine-arts (like sketching and drawing).

My life moto:

Do what you believe in, do your best in whatever you do, and let outcomes follow.

News

2025

• October 2025 — Paper titled CAV Styler: Interpretable & Controllable Style Transfer accepted to ICVGIP 2025.
• September 2025 — Work on Building Trust in Clinical LLMs: Bias Analysis & Dataset Transparency accepted to EMNLP 2025.
• July 2025 — Paper on Prototype Guided Backdoor Defense via Activation Space Manipulation accepted to ICCV 2025.

2024

• January 2024 — Paper Predicting Business Process Events in Presence of Anomalous IT Events published at CODS-COMAD 2024.

2023

• December 2023 — Contributed to release of Med42 (70B Clinical LLM) on HuggingFace.
• September 2023 — Paper Concept Distillation: Leveraging Human-Centered Explanations for Model Improvement accepted to NeurIPS 2023.
• September 2023 — Successfully defended my Master’s Thesis at IIIT Hyderabad.

2022

• December 2022 — Received Best Paper Award and Oral Presentation at ICVGIP 2022 for work titled Interpreting Intrinsic Image Decomposition using Concept Activations.
• October 2022 — Paper CitRet: A Hybrid Model for Cited Text Span Retrieval accepted to COLING 2022.