Ishita Dasgupta

Timeline

1. February 2018

   Testing compositionality in sentence embeddings

   In February 2018, she posts "Evaluating Compositionality in Sentence Embeddings" as first author, with Demi Guo, Andreas Stuhlmüller, Samuel J. Gershman and Noah D. Goodman. The paper presents a natural language inference dataset that requires some compositionality, finds that InferSent sentence embeddings perform poorly on it, and shows that adding the dataset to training improves test performance on it.

2. 2020

   A Harvard physics PhD

   In 2020, Dasgupta earns her PhD from the Department of Physics at Harvard University, working in the Computational Cognitive Neuroscience Lab with Sam Gershman.

3. 2020–2022

   Princeton postdoc and DeepMind move

   She works as a postdoctoral researcher at Princeton University in the Departments of Psychology and Computer Science, in Tom Griffiths's Computational Cognitive Science Lab, and joins DeepMind's New York research team in December 2020. In 2022, she is a Research Scientist at DeepMind and speaks in the Reasoning in Brains vs Machines session of the ICML workshop Beyond Bayes: Paths Towards Universal Reasoning Systems.

4. July 2022

   Content effects in language models

   In July 2022, she posts "Language models show human-like content effects on reasoning tasks" as first author, with Andrew Lampinen, Stephanie Chan, Felix Hill and others. Tested on natural language inference, syllogisms and the Wason selection task, language models answer more accurately when the semantic content of a task supports the logical inference, as humans do.

5. 2022

   A NeurIPS Outstanding Paper

   With Sreejan Kumar, Carlos G. Correa, Tom Griffiths and others, she coauthors "Using natural language and program abstractions to instill human inductive biases in machines," which receives an Outstanding Paper award at NeurIPS 2022. Co-training meta-learning agents on natural language task descriptions and induced programs guides them toward more human-like inductive biases.

6. February 2023

   Language models for embodied reasoning

   In February 2023, Dasgupta is first author of "Collaborating with language models for embodied reasoning." The study links a pre-trained language-model Planner that issues commands to a simple embodied Actor with a Reporter that communicates with the Planner to inform its next command.

7. 2024–2025

   Gemini 1.5 and 2.5 reports

   In April 2024, she is among the authors of the revised Gemini 1.5 technical report, which describes a multimodal model that recalls and reasons over millions of tokens of context. In July 2025, she is also an author of the Gemini 2.5 technical report covering the Gemini 2.5 Pro and Gemini 2.5 Flash models.

8. As of October 2026

   Staff Research Scientist at DeepMind

   In August 2026, she coauthors "Decoupling Planning and Control for Instructable Agents," which evaluates visual-language-model planners and trained controllers across seven embodied environments, and Google lists her among its affiliated authors at COLM in October 2026. She is a Staff Research Scientist at DeepMind whose research combines computational cognitive science and machine learning.