Gemini 4 Argon
· Evals · Post-training
Designed and built an internal benchmark inspired by SWE-bench, and developed metrics and methods to measure and improve SFT models’ climbability during RL.
Medhini Narasimhan
Staff Research Scientist · Google DeepMind
Hi! I'm a Staff Research Scientist at Google DeepMind, working on Gemini Mid/Post-Training. Previously, I led evaluations for Veo and contributed to its pre-training, image-to-video capability, and post-training. My research interests are in building multimodal models for understanding and generation.
I graduated with a PhD in Computer Science from UC Berkeley in 2023, where I was advised by Prof. Trevor Darrell and was a member of BAIR. Prior to Berkeley, I completed my Master's at UIUC advised by Prof. A. Schwing and Prof. S. Lazebnik, and obtained my Bachelor's from NITK, India.
I am honored to be a Snap Research Scholar (2019) and a Siebel Scholar (2018).
Recent Updates
-
We released Gemini 4 Argon, our strongest coding model yet! Try it out.
-
We launched Gemini Omni, bringing world understanding to multimodal generation. Turn your wildest ideas into videos here.
-
Personal milestone: promoted to Staff Research Scientist at Google DeepMind! 🎉
Selected Industry Work
Gemini Omni
· Evals Lead · Multimodal Generation
Led evaluations across multimodal understanding and generation, including the creation and editing of video from text, images, audio, and video.
Veo 1, 2 & 3
· Pre-training · Post-training · Evaluation · Capabilities
Core contributor across three generations of Veo, with work spanning pre-training, image-to-video capabilities, and post-training. Led Veo’s evaluation efforts, combining human assessments and automated metrics to guide model development. Contributions included training-data optimization and supervised fine-tuning methods to improve generation quality.
Research
ISMB2016
Predicting Symptom Severity and Contagiousness of Respiratory Viral Infections
★ Best Poster Award