April 3 seminar: Manish Bhattarai

April 3, 2026

Harnessing Unsupervised Learning to Mitigate Challenges in Scientific Discovery

Manish Bhattarai, Los Alamos National Laboratory

3:00 pm, UNM Centennial Engineering Center, Room 1026
Online Guests: Contact Prof. Santhanam <bsanthan@unm.edu> for a Zoom link

Abstract: Generative AI and LLMs have revolutionized scientific research, boosting productivity, accelerating discoveries, and enhancing efficiency. However, significant challenges persist: LLMs frequently generate “hallucinations”, plausible yet inaccurate information—and are susceptible to adversarial attacks that can spread toxic content and misinformation. Their high computational demands also raise questions of practicality and accessibility.

In this talk, I will present our novel solution that leverages Unsupervised Learning, focusing on Matrix/Tensor factorizations coupled with denoising diffusion models. First, I’ll demonstrate how our approach uses contrastive learning to refine representations, effectively minimizing hallucinations and increasing reliability in tasks like code translation and information retrieval. I’ll also illustrate how our tensor methods compress LLMs, enhancing performance and making deployment feasible on resource-constrained hardware. Furthermore, I’ll show how combining Tensor factorizations with denoising diffusion strengthens model defenses against adversarial attacks, improving robustness in multimodal applications.

Bio: Dr. Manish Bhattarai is a Staff Scientist at Los Alamos National Laboratory, where he leads cutting-edge research at the intersection of generative AI, natural language processing, reinforcement learning, high-performance computing, and tensor methods. He earned his Ph.D. in Electrical Engineering from the University of New Mexico in 2020. His research focuses on building trustworthy, efficient, and scalable AI systems, with contributions spanning explainability, adversarial robustness, and fine-tuning methods for large language models. His work has appeared in leading AI venues such as CVPR, NeurIPS, ICML, and ICLR. Dr. Bhattarai has also made significant contributions to large-scale scientific computing, including the development of high-performance tensor libraries for decomposing exabyte-scale datasets. This work helped earn his team the prestigious 2021 R&D 100 Award. At Los Alamos, he applies AI and tensor-based methods to challenging problems in materials science, chemistry, biology, and multimodal learning, working across data modalities that include text, images, DNA sequences, molecular structures, and signals. Beyond his research, he is a dedicated mentor who has guided more than 20 students through programs such as the ISTI Advanced Machine Learning School, Cybersecurity School, NSF-MSGI, and GRA initiatives.