Student Researcher
Working on multi-agent systems.
Experience
Research Scientist (Intern)
Working on the stability of LLM RL post-training.
PhD Candidate
Research on AI self-improvement, advised by Ilija Bogunovic.
Ph.D. Candidate · UCL AI Centre · University College London
Hi — I'm Xiaohang, currently a Student Researcher at Google DeepMind (London) and a PhD candidate at UCL advised by Ilija Bogunovic. I have also been fortunate to work closely with Yaodong Yang.
My central research is AI self-improvement: studying how a learning system can keep improving from signals it generates itself. As people learn by practicing and refining rather than following an answer key, such a system keeps learning by interacting with the open world, instead of being capped by external supervision that is expensive and limited to its supervisor's level.
Self-play in games. RMDO (ICML 2023) grows a strategy pool with best responses to the current equilibrium, converging to a Nash equilibrium in extensive-form games; AdaDO (JMLR) makes this expansion adaptive and sample-efficient.
Imagined self-play for robust decision-making. ARDT (NeurIPS 2024) trains a sequence model against an adversary imagined from offline data; CART (NeurIPS 2025 Workshop) extends this to stochastic games.
Self-play alignment for LLMs. RSPO (ICML 2026) treats preference optimization as a game the model plays against itself, with plug-and-play regularization and provable Nash convergence; RMOD (ICLR 2026) brings the same game view to controlled decoding (test-time alignment), against a worst-case weighting of objectives.
Self-distillation for diffusion LLMs. wd1 (ICLR 2026) is ratio-free reinforcement learning for dLLM reasoning; GDSD (NeurIPS 2026) recasts RL as distilling a reward-guided version of the denoiser back into itself, completely likelihood-free and ELBO-free.
Before the PhD, I received an MSc in Data Science from UCL and a BSc in Mathematics from Sun Yat-sen University (SYSU).
Working on multi-agent systems.
Working on the stability of LLM RL post-training.
Research on AI self-improvement, advised by Ilija Bogunovic.
Research direction
A common loop — act, evaluate, update — instantiated across four problem families, from expanding strategy pools in games to distilling a denoiser from its own guided signals.
Where self-improvement is literal: iteratively expanding a strategy pool with best responses to its equilibrium, converging to global Nash equilibrium sample-efficiently.
Worst-case-aware agents for adversarial environments: sequence models that condition on returns against imagined adversaries.
Self-play carried into alignment: a policy improved against itself to approximate the Nash equilibrium of the preference optimization game.
The newest instantiation: post-training a denoiser on signals it already generates, without unstable likelihood-ratio estimates.
Invited talk on No-Regret Self-Improvement at the UQ-DS-Seminar, hosted by Ruihong Qiu.
GDSD accepted at NeurIPS 2026.
Joined Google DeepMind (London) as a Student Researcher, working on multi-agent systems.
RSPO accepted at ICML 2026; GDSD preprint released on arXiv.
wd1 and Robust Multi-Objective Controlled Decoding accepted at ICLR 2026.
Invited talk on Diffusion Language Models at Tencent LightSpeed Studio.
Invited talk on Diffusion Language Models and its Post-training at Mind Lab, Macaron AI.
Robust Adversarial RL in Stochastic Games accepted at the NeurIPS 2025 Reliable ML Workshop.
wd1 report released on arXiv.
Invited talk on ARDT at the Google DeepMind Game Theory Group (London).
ARDT accepted at NeurIPS 2024.
Selected recent work. Full author lists on each paper page.
Neural Information Processing Systems (NeurIPS) 2026 · Diffusion LLM
International Conference on Machine Learning (ICML) 2026 · Alignment
International Conference on Learning Representations (ICLR) 2026 · Diffusion LLM
International Conference on Learning Representations (ICLR) 2026 · Alignment
NeurIPS 2024 · Offline RL · invited talk, DeepMind Game Theory Group
NeurIPS 2025 Reliable ML Workshop
International Conference on Machine Learning (ICML) 2023 · Game theory
Journal of Machine Learning Research (JMLR)
IJCAI 2021 · RL theory