Sanjay Adhikesaven

University of California, Berkeley

sanjay.adhikesaven [at] berkeley.edu · LinkedIn · X

I'm a 4th-year EECS undergrad at UC Berkeley, where I'm advised by Sewon Min as part of Berkeley AI Research (BAIR) and the Berkeley NLP Group. I'm interested in how data and training decisions shape language model capabilities, and how we can improve them. I previously worked on training systems and GPU performance at Jane Street.

Publications

• 
  Which Models Are Our Models Built On?
  Auditing Invisible Dependencies in Modern LLMs

  Sanjay Adhikesaven*, Haoxiang Sun*, Sewon Min

  We introduce ModSleuth, an agentic system that reconstructs the dependency graphs behind modern LLMs from papers, model cards, dataset cards, code, configs, and upstream artifacts. It traces how models are used to generate data, filter corpora, evaluate outputs, and guide training decisions, recovering 1,060 source-verified dependencies across four releases and surfacing license-relevant paths, train-evaluation coupling, and documentation mismatches.

  Paper / Demo / Code

  * equal contribution

• 
  Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts

  Jacob Morrison*, Sanjay Adhikesaven*, Akshita Bhagia, Matei Zaharia, Noah A. Smith, Sewon Min

  In Proceedings of COLM 2026

  We present BAR (Branch-Adapt-Route), a recipe for modular post-training. Rather than training a single model on all data at once, BAR trains independent domain experts – each through its own complete training pipeline – and composes them into a unified model via a mixture-of-experts (MoE) architecture. Each expert can be developed, upgraded, or replaced without touching the others.

  Paper / Blog / Models / Code

  * equal contribution

Industry Experience

• Jane Street, Machine Learning Engineer Intern May – Aug 2026

  Training Systems and GPU Performance

• Google, Software Engineer Intern May – Aug 2025

  Agent Evals

• Apple, Machine Learning Engineer Intern Mar – May 2025

  Agentic Search and Memory