Constrained Auto-Bidding via Generative Response Modeling
arXiv preprint · 2026
Xingdong Zuo.
I am a machine learning research engineer at NAVER, where I apply reinforcement learning to large-scale, real-world decision making — recommender systems and auto-bidding in online advertising auctions.
My research interests include offline RL for recommender systems and auto-bidding, and robot learning with vision-language-action (VLA) models and world action models.
Before NAVER, I completed my M.Sc. in Machine Learning at the University of Freiburg, Germany, and my B.Sc. in Applied Mathematics and Computer Science at Linnaeus University, Sweden.
Publications
Drift-VLA: Fast Vision-Language-Action Policies with One-Step Drifting
Preprint · 2026
Method and System to Provide Information Data with Minimal Infrastructure Resources
Korean Patent Publication No. 10-2025-0124713 (2025.08.20) · NAVER Corp · 2025
HyperCLOVA X Technical Report
arXiv preprint · 2024
Direct Preference-based Policy Optimization without Reward Modeling
NeurIPS · 2023
bibtex
@inproceedings{an2023direct,
title={Direct Preference-based Policy Optimization without Reward Modeling},
author={An, Gaon and Lee, Junhyeok and Zuo, Xingdong and Kosaka, Norio and Kim, Kyung-Min and Song, Hyun Oh},
booktitle={Advances in Neural Information Processing Systems},
volume={36},
year={2023}
} Is Learning World Model Always Beneficial For Reinforcement Learning?
Technical report · 2021
Numerical Simulation of Asano-Khrennikov-Ohya Quantum-like Decision Making Model
NeuroQuantology, 12(4) · 2014
bibtex
@article{zuo2014numerical,
title={Numerical Simulation of Asano-Khrennikov-Ohya Quantum-like Decision Making Model},
author={Zuo, Xingdong},
journal={NeuroQuantology},
volume={12},
number={4},
year={2014}
} Talks & Activities
Bid, Win, Profit: Revenue Optimization with a Generative-AI Ad-Auction Simulator
DAN 25 — NAVER Conference, Seoul · 2025
Hugging Face LeRobot Worldwide Hackathon
Seoul · 2025
Open Source
OpenAI Gym
Authored much of the standard wrapper suite — ClipAction, RescaleAction, FrameStack/LazyFrames, GrayScaleObservation, ResizeObservation, TransformObservation/Reward, FilterObservation, FlattenObservation, RecordEpisodeStatistics, AtariPreprocessing — plus vectorized-environment infrastructure. These wrappers live on in Gymnasium.
PyTorch
Early core contributions, including the DataLoader drop_last flag and nn.utils parameters_to_vector / vector_to_parameters.
LeRobot
Contributions to Hugging Face's robot-learning library, including batched video encoding for faster dataset recording and the CLI for launching LeKiwiHost.
lekiwi-tui
A terminal control center for the LeKiwi robot workflow, built with LeRobot and pyratatui.
rlbidder
A modern, modular reinforcement learning library for auto-bidding — a personal side project.
lagom
A light PyTorch infrastructure for rapid prototyping of reinforcement learning algorithms — modular building blocks for agents, environments, and parallelized experiments.
MazeLab
A customizable framework to create maze and gridworld environments for reinforcement learning, with a simple API for designing your own layouts. Cited in academic papers as a benchmark environment.
VIN_PyTorch_Visdom
PyTorch implementation of Value Iteration Networks (NIPS 2016 best paper), with live training visualization in Visdom.
dm2gym
A thin wrapper that converts DeepMind Control Suite environments into the OpenAI Gym interface.
DeepPILCO
PyTorch implementation of "Improving PILCO with Bayesian Neural Network Dynamics Models" — model-based RL with uncertainty-aware dynamics.
recsys_metrics
An efficient PyTorch implementation of the evaluation metrics used in recommender systems.
gym-recsys
A customizable recommender-system simulator exposed through the OpenAI Gym interface, for training and evaluating RL-based recommenders.
elsewhere
Smaller merged patches in TensorFlow, OpenAI Baselines, Hugging Face TRL, torchmetrics, and pytorch/examples.
Writing
Education
- 2019
- M.Sc. in Machine Learning · University of Freiburg, Germany Master's thesis at the Bosch Center for Artificial Intelligence (BCAI).
- 2015
- B.Sc. in Applied Mathematics & Computer Science · Linnaeus University, Sweden Thesis on the Lindblad equation for quantum-like decision-making models, supervised by Andrei Khrennikov.