Junyu (James) Guo

01 / About

About me

I am Junyu Guo (James), a Ph.D. candidate at the University of California, Berkeley, where I am fortunate to be advised by Prof. Javad Lavaei. My research focuses on designing efficient decision-making algorithms for agents operating under uncertainty, with particular emphasis on data-driven sequential decision-making for Large Language Models (LLMs). I am also fortunate to work with Prof. Costas Spanos (Berkeley EECS), Prof. Ming Jin and Dr. Shangding Gu, as well as Liyuan Liang and Yuchen Fang from Prof. Lavaei’s group.

Prior to Berkeley, I completed my undergraduate studies in the Department of Mathematical Sciences at Tsinghua University in Beijing, China, where I was advised by Prof. Chenxu Li and Prof. Yiwen Shen on efficient simulation for portfolio allocation in incomplete markets.

During my time at Tsinghua, I was also an exchange student at Cornell University, collaborating with Prof. Andreea Minca and Prof. Qiaomin Xie on reinforcement learning for mean field games.

Beyond research, I am an avid football fan (Manchester City and FC Barcelona), and I enjoy traveling, music, hiking, and playing soccer.

02 / Focus

Research interests

Reinforcement Learning

Efficient algorithms for sequential decision-making under uncertainty, with emphasis on meta-RL for quick adaptation to non-stationary environments.

LLM Reasoning

Improving the decision-making and reasoning of large language models through uncertainty estimation and quantification.

Trustworthy AI

AI systems that are interpretable, explainable, fair and privacy-preserving.

03 / Research

Selected papers

All research

2026

PreprintLLM reasoning

When Context Changes: Understanding Update Failures in LLMs

Junyu Guo, Yuchen Fang, Shangding Gu, Costas Spanos, James Demmel, Javad Lavaei

LLM agents can answer with an old value of a variable even when the updated one is still in context, a failure we call stale binding. We introduce the Controlled In-Context Memory (CICM) benchmark, trace the failure to attention drifting toward old values, and correct most of these errors by redirecting attention without further training.

arXiv
2026

PreprintLLM reasoning

LLMs Should Express Uncertainty Explicitly

Junyu Guo, Shangding Gu, Ming Jin, Costas Spanos, Javad Lavaei

Trains language models to state their uncertainty in two ways: a verbalized confidence score after reasoning, and explicit uncertainty markers during reasoning. This reduces overconfident errors and gives downstream systems such as retrieval-augmented generation a signal to act on.

arXiv
2025

NeurIPS 2025Safe RL

Don't Trade Off Safety: Diffusion Regularization for Constrained Offline RL

Junyu Guo, Zhi Zheng, Donghao Ying, Ming Jin, Shangding Gu, Costas Spanos, Javad Lavaei

DRCORL uses a diffusion model to regularize policy learning in offline safe reinforcement learning, and gradient manipulation to resolve conflicts between reward and safety objectives, while keeping inference fast.

arXiv NeurIPS
2025

NeurIPS MATH-AI Workshop 2025LLM reasoning

Meta Thinker: Thinking What AI Thinks

Junyu Guo, Shangding Gu, Costas J. Spanos, Javad Lavaei

Asks whether LLMs can pick the right reasoning style for a problem on their own. After comparing five reasoning paradigms on mathematical, logical and commonsense benchmarks, we propose a meta-thinking prompt algorithm that selects or synthesizes a style from the input, improving both accuracy and token efficiency.

OpenReview Poster page Code
2025

PreprintLLM reasoning

StyleBench: Evaluating Thinking Styles in Large Language Models

Junyu Guo, Shangding Gu, Ming Jin, Costas Spanos, Javad Lavaei

A benchmark that evaluates five reasoning strategies across diverse tasks and model scales, showing when structured reasoning helps and when it only adds cost or new failure modes.

arXiv Code