Carrie Yuan

I am a second year PhD student at the University of Washington, Allen School of Computer Science and Engineering, starting Fall 2025. I am fortunate to be supervised by Prof. Natasha Jaques. Previously I was a Software Development Engineer II at Amazon AGI Data Prep team, focusing on developing data preparation tools and infrastructure used for large language model (LLM) training. I graduated from Carnegie Mellon University with a B.S. in Neuroscience and Computer Science (Fun fact: I was the only person graduated with a second major in CS in the Neuroscience department).

Email  |  CV (Updated: June 2026)  |  LinkedIn  |  Twitter  |  Github |  Research Philosophy

Read  |  Run  |  Travel

Also, check out my friends

• [April 2026]

  I co-hosted Dan Liebling's talk on Science AI: After the "Bitter Lesson" at the DUB Seminar.

• [March 2026]

  I gave a talk on AgenticRed and using self-evolving algorithms on AI Safety at the Knowledge Sharing session hosted by Google's ML Red Team.

• [Feb 2026]

  I drafted a MARL-based water resource management prototype at Pale Blue Futures: AI for Water Resilience Research Workshop, at University of Toronto.

• [Nov 2025]

  I presented DiZCo: Planning Zero-Shot Coordination in World Models at the Allen School’s annual Industry Affiliates Research Showcase.

• [Nov 2025]

  I gave a talk at Ballard High School, Seattle as part of Seattle Public School Hour of Code program, part of CSEdWeek program!

• [Sept 2025]

  I started my PhD at University of Washington. An exciting chapter of my life!

• [July 2025]

  I attended the Cooperative AI Summer School in Marlow, UK, 9-13 July 2025.

• [June 2025]

  I gave a talk on MARL's potential usage in addressing climate change at Cooperation for Climate Change Action Workshop at HEC Lausanne, 5-6 June 2025.

• [June - Oct 2025]

  I worked as Research Intern with Dr. Goran Radanović at the Max Planck Institute for Software Systems (MPI-SWS) in Saarbrücken, Germany.

• [April 2025]

  I presented InvestESG at ICLR 2025, Singapore.

• [March 2025]

  I accepted the CSE PhD offer from UW.

• [Jan 2025]

  InvestESG was accepted by ICLR 2025!

• [Dec 2024]

  I graduated as Master of Science in Computational Linguistics from University of Washington.

• [Dec 2024]

  I presented InvestESG at NeurIPS 2024 Workshop on Tackling Climate Change with Machine Learning.

• [Oct 2024]

  Preliminary result of InvestESG was accepted by NeurIPS 2024 Workshop on Tackling Climate Change with Machine Learning.

• [Oct 2024]

  Submitted my first first-authored paper, InvestESG, to ICLR 2025.

• [June 2024]

  My first paper, CASHER, was accepted by RSS 2024 Data Generation for Robotics (DGR) Workshop!

• [Sep 2022]

  Starting part-time master program at University of Washington.

• [July 2021]

  Relocating to Seattle and starting to work for Amazon.

• [May 2021]

  Graduated with a B.S in Neuroscience and Computer Science from Carnegie Mellon Univeristy (2 majors in 3 years). Go Tartan!

• [August 2020]

  Attending Computational Neuroscience summer course at NeuroMatch.

• [August 2019]

  Received Summer Undergraduate Research Fellowships (SURF) to conducting neuroscience research under the supervision of Dr. Alison Barth.

Research

I am broadly interested in AI for Science, multi-agent systems, and reinforcement learning. .

We present MoDA (Mode-conditioned Diversity Alignment), an online post-training RL algorithm that jointly optimizes generation quality and diversity, inspired by the coordination perspective in multi-agent reinforcement learning (MARL). MoDA trains a single shared LLM policy conditioned on abstract numbered roles, where each role acts as an agent competing to produce outputs distinct from the others. This formulation encourages mode-conditioned agents to explore complementary regions of the high-quality output space without requiring hand-crafted personas or architectural modifications.

Keywords: Multi-Agent Reinforcement Learning, Diversity, Alignment, Language Models

We evaluate large language model (LLM) agents as behaviorally grounded simulators of companies and investors in a continuous-action climate-finance dilemma built on the InvestESG platform, and benchmark them against fully rational profit-driven reinforcement learning (RL) agents, a centralized social planner, and human participants in the same game.

Keywords: Computational Social Science, NLP for Social Good, Multi-Agent, MARL

We introduce AgenticRed, an automated pipeline that leverages LLMs' in-context learning to iteratively design and refine red-teaming systems without human intervention. Rather than optimizing attacker policies within predefined structures, AgenticRed treats red-teaming as a system design problem, and it autonomously evolves automated red-teaming systems using evolutionary selection and generational knowledge. Red-teaming systems designed by AgenticRed consistently outperform state-of-the-art approaches, achieving 96% attack success rate (ASR) on Llama-2-7B, 98% on Llama-3-8B and 100% on Qwen3-8B on HarmBench. Our approach generates robust, query-agnostic red-teaming systems that transfer strongly to the latest proprietary models, achieving an impressive 100% ASR on GPT-5.1, DeepSeek-R1 and DeepSeek V3.2.

Keywords: Red-Teaming, AI Safety, Open-endedness, Evolutionary Algorithms

Although a variety of transformers have been proposed for symbolic music generation in recent years, there is still little comprehensive study on how specific design choices affect the quality of the generated music. In this work, we systematically compare different datasets, model architectures, model sizes, and training strategies for the task of symbolic piano music generation. To support model development and evaluation, we examine a range of quantitative metrics and analyze how well they correlate with human judgment collected through listening studies. Our best-performing model, a 950M-parameter transformer trained on 80K MIDI files from diverse genres, produces outputs that are often rated as human-composed in a Turing-style listening survey.

Keywords: music information retrieval, music generation, transformer, symbolic music

We introduced, InvestESG, a novel multi-agent reinforcement learning (MARL) benchmark designed to study the impact of Environmental, Social, and Governance (ESG) disclosure mandates on corporate climate investments. The benchmark models an intertemporal social dilemma where companies balance short-term profit losses from climate mitigation efforts and long-term benefits from reducing climate risk, while ESG-conscious investors attempt to influence corporate behavior through their investment decisions.

Keywords: multi-agent reinforcement learning, climate change, ai for climate

We propose Crowdsourcing and Amortizing Human Effort for Real-to-Sim-to-Real (CASHER), a pipeline for scaling up data collection and learning generalist policies where human effort scales sublinearly with the number of environments where data is collected. The key idea is to crowdsource digital twins of real-world scenes using 3D reconstruction techniques and collect large-scale data in these simulation scenes, rather than in the real-world. Data collection in simulation is initially driven by reinforcement learning bootstrapped with human demonstrations.

Keywords: Data Scaling in Robotics; Reinforcement Learning; Grasping and Manipulation

Awards

Amazon AGI DataPrep Team, Peer Recognition Awards, 2023

Amazon AGI DataPrep Team, Peer Recognition Awards, 2022

Carnegie Mellon University, University Honor (summa cum laude), 2021

Summer Undergraduate Research Fellowships (SURF), 2019

Services

Conference Reviewer: AMLC(2024), ICML(2025), ICLR(2026), NeurIPS(2026)

Workshop Reviewer: ICLR 2026 Workshop on Recursive Self-Improvement

Social RL Reading Group Organizer (2026-Present)

GPSS Senator of Paul G. Allen Center for Computer Science & Engineering (2025-Present)

DUB Community Day Session Chair (2025)

Allen School Pre-Application Mentorship Service (PAMS) Mentor (2025)

WEIRDLab Tutorial Organizer (2024)

Research Philosophy

I like to model everything. Some might call a reductionist. But I cannot resist the appeal of describing the mundane everyday complex phenomenons, such as weather, grocery price, or interpersonal relations, in terms of elegant mathematical symbols and functions with certain input space and output space. Studying computer science and mathematics allows me to do that in a more rigorous fashion. It teaches me a language to describe a problem space so that my ideas can be communicated universally. Studying neuroscience allows me to see both the complexiticity and simplicity of the world I live in. We can describe our minds as simple as an ensemble of electrical pulses, but the complexity lies in that our coordination, our neuroplasticity and biological efficiency (compared to its power-hungry machine counterpart).

I believe that real science does not have boundary of subjects. There are techniques, or tools, (I like to imagine things in a visual sense) that might be potentially helpful to solve a problem, and these tools might be labeled with academic subjects such as mathematics, economics, cognitive science, psychology, biology and physics. We should be open to learn to any tools that would be useful to our goal.

Also, check out my friends:

• Hesper Yin
• Yutong Zhu
• Ábel Ilyés-Kun

Template is from Jon Barron's website.