About

Sarath Shekkizhar is a Staff Research Scientist at Salesforce. He joined Salesforce through its acquisition of Tenyx, where he was part of the founding team building voice AI for customer support. His current research studies large language models: training and post-training methods, reasoning and alignment, the geometry of neural representations, and the behavior of multi-agent systems.

Research

The central question is how a model's internal representation relates to the behavior observed during training, reasoning, and interaction. Recent work examines identity failures when language-model agents talk to one another, interaction awareness through user-turn generation, and the limits of counterparty modeling in negotiation. Earlier work used representation geometry to study reasoning, toxicity, and continual learning in language models.

Before working on language models, Sarath developed methods for graph signal processing and machine learning. This work included non-negative kernel graphs, data selection for scalable graph learning, and methods for interpreting neural network outputs through graph structure. The common setup is empirical: define a measurable mechanism, test it under controlled changes, and report where the explanation succeeds or fails.

Background

Sarath received a Ph.D. in Electrical Engineering from the University of Southern California, advised by Antonio Ortega. He also holds M.S. degrees in Computer Science and Electrical Engineering from USC and a B.Tech. in Electronics and Communication from the National Institute of Technology, Tiruchirappalli. His recognitions include the IEEE Rising Star in Signal Processing at ICASSP 2023 and an IEEE Best Student Paper Award at ICIP 2020.

The CV contains the complete education, employment, patent, award, and academic-service record. The publications page lists papers and preprints, and the blog contains longer technical arguments and experiment-driven notes.