Amrutha Saseendran
AI Research Scientist
I am an AI researcher and GenAI Tech Lead with 8+ years of experience across industrial and academic research, currently working at AstraZeneca in Cambridge (UK). My work centers on large language models, multimodal AI, and generative methods, with a focus on building reliable and deployable AI systems for scientific and clinical use.
I hold a PhD in machine learning, with research contributions in generative modeling and representation learning, and have published in leading venues including NeurIPS, ICML, ICLR, and ICCV. My background combines foundational ML research with applied work in industry, bridging methodological development and real-world deployment.
More recently, I have been leading technical work on agentic AI systems, including multi-agent reasoning and long-horizon workflows, with the aim of integrating LLM-based agents into complex scientific pipelines in a robust and responsible manner.
Recent Publications
Segment Anyword: Mask Prompt Inversion for Open-Set Grounded Segmentation
ICML 2025We introduce Segment Anyword, a training-free method for open-set image segmentation that combines diffusion-based visual prompts with linguistic structure to consistently segment objects across diverse text descriptions, outperforming existing methods on challenging benchmarks.
Decore: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations
EMNLP 2025We propose DeCoRe, a training-free decoding strategy that reduces LLM hallucinations by contrasting outputs from base and retrieval-head-masked models, improving contextual faithfulness across summarisation, instruction following, and open-book QA tasks.
Balancing Act: Diversity and Consistency in Large Language Model Ensembles
ICLR 2025We propose DMoA, a dynamic LLM ensemble that formalizes the trade-offs between diversity and consistency through a unified framework, revealing task-dependent reasoning requirements and delivering state-of-the-art results on BigBench Hard.
News & Events
Speaking at GenAI London 2025
I will be speaking at GenAI London 2025 on “Agentic AI for Science: From Discovery to Design,” sharing practical insights from applying Genai/Agentic AI approaches in scientific R&D.
Latent Refinement Decoding in Hugging Face Top-3 Papers
Our collaborative work with King’s College London (KCL) was featured among the top 3 daily papers on Hugging Face.
Paper Accepted at EMNLP 2025
Excited to share that "Decore: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations" will be presented at EMNLP 2025.
Paper Accepted at ICLR 2025
Our work on "Balancing Act: Diversity and Consistency in Large Language Model Ensembles" has been accepted at ICLR 2025!
Joined AstraZeneca as GenAI Research Scientist
Started leading the GenAI initiatives at AstraZeneca's Center for AI, focusing on LLM-based solutions for biopharma R&D.
Ph.D. Defense - Summa Cum Laude
Successfully defended my Ph.D. thesis on "Optimization of the Latent Space of Deep Generative Models" at Universität Siegen.
Paper Accepted at NeurIPS 2022
Our work "Trading off Image Quality for Robustness is not Necessary with Deterministic Autoencoders" accepted at NeurIPS.
Paper Accepted at NeurIPS 2021
"Shape Your Space: A Gaussian Mixture Regularization Approach to Deterministic Autoencoders" accepted at NeurIPS 2021.