Eric Chu

I am a Staff Research Scientist at Google DeepMind. I've been a core contributor across Gemini 1.0 – 3.1, PaLM 2, and Bard, leading efforts in RL, coding, synthetic data, and most recently self-improving reasoning agents.

I'm most interested in AI systems that reshape how people learn and how science gets done. My work spans three pillars — generative models, alignment and human-AI collaboration, and AI for science — that connect back to that goal.

I completed my PhD at MIT, advised by Deb Roy in the Media Lab and Jacob Andreas in CSAIL, where I developed novel deep learning models and studied human behavior. I interned at FAIR with Jason Weston and Stephen Roller, and Google Brain with Peter J. Liu. Before that, I was at UC Berkeley, where I got into machine learning via research in computational biology.

Selected Work & Research

For a list of my papers, see Google Scholar.

Generative models and deep learning

• Large language models at Google DeepMind (2022 — Present)
  Multi-author technical report, incl. Eric Chu
  Gemini 2.5 Gemini 1.5 Gemini 1 PaLM 2 Bard
• Coding large language models launched around Google I/O (2023)
  Bard coding Cloud (DuetAI) Search (CodeTips) AI-powered Colab FunSearch
• Language models trained on media diets can predict public opinion (2021)
  Eric Chu, Jacob Andreas, Steve Ansolabehere, Deb Roy
  Paper Cited at US Congressional AI Hearing
• BIG-Bench: "navigate" logical and spatial reasoning task (2020)
  Collaborative benchmark, incl. Eric Chu, Sneha Priscilla Makini
  Paper @ TMLR Code Task highlighted in BBH-Hard
• MeanSum: first end-to-end neural model for unsupervised, multi-document, abstractive summarization (2018)
  Eric Chu*, Peter Liu*
  Paper @ ICMLSlidesCode
• Learning personas from dialogue with attentive memory networks (2017)
  Eric Chu*, Prashanth Vijayaraghavan*, Deb Roy
  Paper @ EMNLPData
• Deep audio and visual models for learning emotional arcs in movies (2016)
  Eric Chu, Deb Roy
  Paper @ ICDM Data Thesis NYTimes coverage Variety coverage

• Many-Shot In-Context Learning (2024) | Paper @ NeurIPS
• Reasoning via sequential scratchpads (using T5 on Fermi questions, before CoT) (2021)
• Continual language learning via adapters and model merging (2021)
• Influence functions for tracing toxic generations in language models (2020)
• Generative models with hierarchical plans (2019) | slides
• Speech recognition and speech synthesis for adaptive literacy tutoring (2016) | Code 1 2 3 4

Alignment and human-AI collaboration

• Learning from human feedback at Google: initiated and led RLHF x code (2022 — 2023)
• Code safety at Google: led generative code safety efforts (2022)
• Parents' online school reviews reflect several racial and socioeconomic disparities in K–12 education (2021)
  Nabeel Gillani, Eric Chu, Doug Beeferman, Rebecca Eynon, Deb Roy
  Paper @ AERA MIT news
• Are visual explanations useful? A case study in model-in-the-loop prediction (2019)
  Eric Chu, Deb Roy, Jacob Andreas
  Paper
• Games for fairness and interpretability (2019)
  Eric Chu, Nabeel Gillani, Sneha Priscilla Makini
  Paper @ WWW, ICLR Trustworthy Workshop
• AI & Equality: co-created class on AI and equality (2017)
  MIT course

• Evolving Evocative 2D Views of Generated 3D Objects (2021) | Paper @ NeurIPS Creativity Workshop
• Model version control and interpretable model diffs (2020)
• Artistic influence GAN (2019) | Paper @ NeurIPS Creativity Workshop
• CAD tool for designing topological sculptures (2013)

AI for science

A returning interest, going back to undergrad research at Berkeley.

• Agents for augmenting and automating research (2024 — Present)
• Leveraging superhuman AIs as pedagogical partners (2021)
  Slides

• Joint representation learning of movies and fMRI brain activity, building on the emotional arcs dataset (2020)
• Mathematical reasoning, uncertainty, and numeracy in language models (2019)
• Machine learning for protein structure prediction research (2013)
• Magnetic particle imaging research (2011)