Zhengyan ShiHi, welcome to my personal page. I am a Senior Researcher at Microsoft Research (MSR). My current research at MSR focuses on teaching language models (LMs) to reason and code. I build learning loops in which LMs not only act but also reason within scalable, self-evolving environments. By allowing models to plan and interact with various environments, I explore how LMs learn from experience/past interactions to continually refine themselves.
I obtained my PhD in Computer Science at University College London (UCL). Before that, I completed an MSc in Data Science (Statistics) with Distinction at UCL and a BSc in Mathematics with First Class Honours from the University of Liverpool and Xi'an Jiaotong-Liverpool University. I have also held research internships at Cohere (London) and Amazon (London & Seattle). Central to my work is the ambition to leverage LMs efficiently and robustly to solve general tasks. To that end, my existing work can be broadly categorized into the following directions:
Google Scholar / Twitter / Github / LinkedIn / Email |
|
arXiv preprint, 2025
Presents BugPilot, a method that instructs code agents to add features which inadvertently break tests, yielding realistic high-quality synthetic data at scale for code agents.
arXiv preprint, 2025
Introduces Gistify, a task requiring coding LLMs to distill large codebases into minimal executable files, highlighting challenges in codebase understanding for current state-of-the-art code agents.
Preprint, 2024
In this work, we identify a critical issue of likelihood over-optimisation in state-of-the-art Direct Alignment Algorithms (DAAs) and explore the relationship between completion likelihood and model performance.
Advances in Neural Information Processing Systems (NeurIPS), 2024
Paper / Github / Community Discussion
We show that in certain scenarios, applying loss to instructions rather than outputs only, which we refer to as Instruction Modelling, could largely improve the performance of instruction tuning on both various NLP and open-ended generation benchmarks. Remarkably, in the most advantageous case, our approach boosts model performance on AlpacaEval 1.0 by over 100%.
International Conference on Learning Representations (ICLR), 2024
Paper / Github / Trending in Community
Improves efficiency of Prompt Tuning in both time and memory by over 20% (when T5-Base is used as the backbone), with better performance. DePT grows more efficient as the model size increases.
Advances in Neural Information Processing Systems (NeurIPS), 2023
Paper / Github / Trending in Community
Combines the idea of the instruction tuning and language modelling. Represents the first work to perform instruction tuning via unsupervised objectives. Boosts prompt-based fine-tuning performance by over 20% in absolute.
Association for Computational Linguistics (Findings of ACL), 2023
Shows Task-adaptive Pre-training (TAPT) as a simple yet effective method for semi-supervised learning (often SoTA performance). Highlights the effectiveness of TAPT even with only a few hundred unlabelled samples (in contrary to the common belief that continued pre-training requires a large amount of unlabelled data).
Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022
Paper / Github / HuggingFace Dataset
Introduces StepGame, a new benchmark for testing multi-hop spatial reasoning in texts. This dataset challenges models to perform robust spatial reasoning across multiple steps, providing a valuable tool for advancing natural language understanding in complex spatial scenarios.
Program Committee: NeurIPS (2023, 2024), ICML (2024), ICLR (2025), AAAI (2023, 2024), COLM (2024), ACL ARR (Feb. 2023 - Jan. 2024), ACL (2023), EMNLP (2022, 2023), EACL (2023), COLING (2023, 2024), ECML/PKDD (2022), KDD (2023), SIGIR (2022, 2023, 2024), ECIR (2024), SDM (2024)