Mian Wu

Mian Wu

Modeling the World
About
I'm a senior undergraduate student majoring in Electrical and Computer Engineering at Shanghai Jiao Tong University. I was a visiting student and a member of the Robotic Artificial Intelligence and Learning Lab at UC Berkeley, advised by Sergey Levine. I currently work closely with Sewon Min and Aviral Kumar.
My recent research centers on scalable reinforcement learning with large language models and vision-language models. I am particularly interested in developing RL methods that improve the reasoning capabilities of agents and generalize to more complex free-form generation tasks.

Research

|
Can Video Models Reason Step-by-Step? The Autoregressive Paradigm for Chain-of-Frames
Mian Wu*, Bingda Tang*, Disen Lan, Harold Chen
Submitted to CVPR, 2027
We develop an autoregressive video reasoning framework with on-policy correction, training models to recover from errors in their own generated trajectories rather than solely imitating expert demonstrations. The approach improves long-horizon visual reasoning, reduces error accumulation, and enables more reliable recovery from intermediate mistakes.
Causal Memory for LLM Agents
Mian Wu, Yujia Zheng
Submitted to CLeaR, 2027
We formalize an LLM agent's memory as a temporal causal structure over its past records, recovered from execution logs with causal discovery, to select the records a decision needs and trace the sources it relied on.
Full Visibility Makes Agent Swarms Converge Early and Spread Overfitted Code
Xiaoyang Cao*, Yanlong Liu*, Mian Wu*, Xuan Jiang, Yanchen Liu, Ao Qu, Sraavya Sambara, Yuhan Tang, Awad Abdelhalim, Jinhua Zhao, Paul Pu Liang, Michiel Bakker
Submitted to AAAI 2027 WMAC Workshop
We show that a shared leaderboard makes swarms of 32 LLM coding agents converge on one program lineage about twice as fast as a three-peer follow graph and spreads test-case-overfitted code, ranking first of six conditions on the agents' own grader and fourth on the official ALE-Bench judge.
Beyond the Sampling Frontier: From Evolutionary Discoveries to Generalizable Capability
Mian Wu, Harman Singh, Boxue Yang, Kurt Keutzer, Aviral Kumar, Sewon Min
NeurIPS 2026 EvoRobust Workshop
We propose Evolutionary Self-Improvement (ESI), which runs verifier-guided evolutionary search on training problems the model cannot solve by sampling and fine-tunes the same model on the discovered programs, improving direct generation on unseen hard competitive-programming problems beyond sampling-based self-training.
RLAC: Reinforcement Learning with Adversarial Critic for Free-Form Generation Tasks
Mian Wu, Gavin Zhang, Sewon Min, Sergey Levine, Aviral Kumar
ICLR, 2026
We propose a post-training methodology for improving language models on open-ended generation tasks. Instead of relying on static reward models, RLAC trains a dynamic LLM critic alongside the generator using adversarial gameplay. The critic identifies the most likely failure modes which are then validated externally, adapting as the generator improves. We demonstrate improvements in factual accuracy for text generation and correctness for code generation across multiple benchmarks.
Goal-Driven Reward by Video Diffusion Models for Reinforcement Learning
Qi Wang*, Mian Wu*, Yuyang Zhang*, Mingqi Yuan, Wenyao Zhang, Haoxiang You, Yunbo Wang, Xin Jin, Xiaokang Yang, Wenjun Zeng
CVPR, 2026
We introduce a demo-free reward framework that leverages pretrained video diffusion models to generate goal-conditioned reward signals for RL agents, eliminating the need for hand-crafted reward design or expert demonstrations.
* Equal contribution

Research Interests

Machine Learning, Natural Language Processing, Computer Vision, AI Safety, Reinforcement Learning

Education

B.S. in Electrical and Computer Engineering, Shanghai Jiao Tong University
2021 - 2024, 2025 - 2026

Art

I'm also a drummer who enjoys progressive, metal, and emo rock. I founded the band QaQqvq0.0.
Band photo Album cover
Loading...
...
...
...