Josh Barua

Hey there! I'm a research fellow at the Cambridge-Boston Alignment Initiative, advised by Aaron Mueller. Before that, I was a predoctoral researcher at UT Austin with Kyle Mahowald. And before that, I was an undergrad at UC Berkeley 🐻, where I was fortunate to get my start in research with Alane Suhr as part of the Berkeley NLP Group. I've also spent great summers interning at CMU LTI with Graham Neubig and the MIT Lincoln Lab.

Current local minima

Reasoning in LLMs

Recently, I've been exploring the underlying principles of reasoning in LLMs: to what extent it generalizes, how we can control it, and what the role of human-like reasoning is in model computation.

Steepest directions for gradient updates

Human–model collaboration & oversight

How can humans learn from, collaborate with, and oversee "superhuman" models? Can we design guardrails that better scale with model capabilities?

Science of language models

I am interested in understanding the structure of representations and the kinds of solutions neural networks learn. Can we find simple explanations in small models and controlled testbeds that reliably predict more complex phenomena?

In addition to research, I spent three fun semesters teaching the data structures & algos class (CS 61B) at Berkeley. In my free time, I enjoy climbing rocks, throwing discs, and sending funny cat videos to Angela.

[Jan 2026] Our paper Long Chain-of-Thought Reasoning Across Languages was accepted to ICLR 2026!! See you in Brazil 🇧🇷

[Sep 2025] Spending the next year at UT Austin working with Kyle Mahowald 🤘

[Jul 2025] Our paper Long Chain-of-Thought Reasoning Across Languages was accepted to the ScalR Workshop at COLM! See you in Montréal 🇨🇦

[May 2025] I started my research internship at CMU LTI working with Graham Neubig!

[Sep 2024] Our paper Using Language Models to Disambiguate Lexical Choices in Translation was accepted to EMNLP 2024! See you in Miami 🌴

Bayesian Cue Combination Explains Modality Preferences in Vision-Language Models
Josh Barua, William Rudman, Kyle Mahowald
Under Review

Long Chain-of-Thought Reasoning Across Languages
Josh Barua, Seun Eisape, Kayo Yin, Alane Suhr
ICLR 2026 / ScalR @ COLM 2025
code

Using Language Models to Disambiguate Lexical Choices in Translation
Josh Barua, Sanjay Subramanian, Kayo Yin, Alane Suhr
EMNLP 2024
code

Improving Medical Visual Representations via Radiology Report Generation
Keegan Quigley, Miriam Cha, Josh Barua, Geeticka Chauhan, Seth Berkowitz, Steven Horng, Polina Golland
ArXiv 2023