Zhihui Xie

Also known as 谢知晖

Zhihui Xie is a PhD researcher at the University of Hong Kong whose work covers critique and reasoning in language models, vision-language alignment and diffusion language models. After bachelor's and master's degrees at Shanghai Jiao Tong University, they were credited among the authors of Reka's 2024 report on the Core, Flash, and Edge models and coauthored the VLFeedback dataset. In 2025, they were first author of the ICML paper on CTRL, a framework that trains critics for code generation with reinforcement learning, and of Dream-Coder 7B, an open diffusion language model for code, and they joined Meta Superintelligence Labs as a Research Intern. They are a third-year PhD student at HKU, advised by Lingpeng Kong and Qi Liu.

Badges: Bridge

Timeline

By 2023

Degrees at Shanghai Jiao Tong

They earn a bachelor's degree in the IEEE Honor Class at Shanghai Jiao Tong University, working with Junchi Yan, and a master's degree there under Shuai Li. Their 2021 SIGIR paper on a comparison-based conversational recommender system with relative bandit feedback is coauthored with Li.
2023

Pretrained Decision Transformer

While at the John Hopcroft Center for Computer Science at Shanghai Jiao Tong University, they coauthor Future-conditioned Unsupervised Pretraining for Decision Transformer with researchers from Tencent AI Lab. The paper proposes the Pretrained Decision Transformer (PDT), which uses future trajectory information as privileged context to predict actions during training, and appears at ICML 2023.
April 2024

Reka Core, Flash, and Edge

They are among the authors credited on Reka's technical report on Reka Core, Flash, and Edge, multimodal language models trained from scratch that process text, images, video, and audio. The report is posted in April 2024, and Reka's cofounder Qi Liu is an assistant professor at the University of Hong Kong.
2024

VLFeedback and Silkie

They coauthor VLFeedback, a vision-language feedback dataset of over 82K multimodal instructions with rationales generated by off-the-shelf models without human annotations. The paper also trains Silkie, a vision-language model fine-tuned by direct preference optimization on the dataset, and appears at EMNLP 2024.
2025

Critic training with CTRL

They are first author of Teaching Language Models to Critique via Reinforcement Learning, published at ICML 2025 with Lingpeng Kong and four other coauthors. Its CTRL framework, short for Critic Training via Reinforcement Learning, trains a critic model to generate feedback that maximizes correction performance for a fixed generator model without human supervision, and enables test-time scaling with up to 106.1% relative improvements across code generation benchmarks.
April 2025–July 2025

Dream 7B and Dream-Coder 7B

In April 2025, they help release Dream 7B, an open diffusion language model from the University of Hong Kong and Huawei Noah's Ark Lab that refines sequences in parallel through iterative denoising. Dream-Coder 7B, an open-source discrete diffusion language model for code generation, follows in July 2025 with Xie as first author.
June 2025

POLARIS and a Meta internship

In June 2025, they coauthor the release of POLARIS, a post-training recipe that boosts 4B models to surpass Claude Opus 4 on AIME25. They join Meta Superintelligence Labs as a Research Intern in summer 2025.
As of October 2026

PhD student at HKU

They are a third-year PhD student at the University of Hong Kong, advised by Lingpeng Kong and Qi Liu in the HKU NLP group. In 2026, they coauthor Self-Distilled Reasoner, whose On-Policy Self-Distillation (OPSD) method lets a single language model act as both teacher and student with different contexts. In September 2026, they help release RecreationWorld, a scalable framework for studying hybrid computer-use agents.