Chuyan Zhou

M1 Okazaki Lab, Institute of Science Tokyo

I’m currently a first-year master student at 東京科学大学, Institute of Science Tokyo majoring in Artificial Intelligence, with research interest in LLMs and related areas. Here is my Curriculum Vitae.

Research Interest

Currently, I am most interested or committed in these research topics:

Diffusion LLMs
- Architecture
- Post-training and Scaling
- Agentic RL of dLLMs
Agentic LLMs
- Agentic RL infrastructure and algorithms
- Harness Design
- Recursive Self-Improvement
Post-training of LLMs
- Reinforcement Finetuning, Process Reward Models & Design
- Distillation
Latent/Soft-Token LLMs
Looped Transformers
Efficient LLMs, such asF Decodi
- Fast Decoding Algorithms (Speculative/Jacobi/…), especially those using dLLMs
- Quantization,
- KV Cache Compression/Pruning,
Combination of Connectionist and Probabilistic/Symbolic Methods for NLP
Mechanistic Interpretability
- Sparse Autoencoders
AI for Biology

where the bold items are which I have nonzero experience on before.

Generally, I am interested in researches in fields involving Computational Linguistics, Natural Language Processing, Machine Learning Systems & Theories, Reinforcement Learning and AI for Science.

Misc

C1 level in Japanese & English (JLPT N1 169/180, TOEFL iBT 108/120)
I post blogs here for research, notes and random things such as my traveling:

Here follows my publication list and research project experience.

News

2026.09.28: I officially joined the Okazaki Lab along with the official enrollment.
2026.05.28: I received an offer of admission to 東京科学大学大学院情報理工学院情報工学系知能情報コース (Major of Artificial Intelligence, Department of Computer Science, School of Computing, Institute of Science Tokyo) for master’s studies under Prof. Naoaki Okazaki in the Okazaki Laboratory. I will begin my master’s studies in the Artificial Intelligence course on September 28, 2026.
2026.05.17: Following my received offer to the University of Tokyo’s International Technology Management (G30-TMI) program, I received an offer of placement in my first-choice Matsuo-Iwasawa Laboratory.
2026.04.07: Our paper, GiLT: Augmenting Transformer Language Models with Dependency Graphs, was accepted to the ACL 2026 Main Conference.
2026.03.06: I received an offer of admission to the master’s program in 東京大学大学院工学系研究科技術経営戦略学専攻 (International Technology Management, G30-TMI) at the University of Tokyo.
2026.02.13: By 一般入試 (general entrance examination), I received an offer of admission to the master’s program in 東京大学大学院情報理工学系研究科創造情報学専攻 (Creative Informatics at the University of Tokyo), with placement in my first-choice Nakayama Laboratory.
2025.05: I received the Outstanding Student Award and Scholarship from ShanghaiTech University.
2025.01: I completed a semester at UC Berkeley through the GLOBE Program, earning a GPA of 4.0/4.0.
2023.10: Our team won a silver medal in the Bengali.AI Speech Recognition Challenge on Kaggle, finishing in 24th place.
2023.10: I joined Prof. Kewei Tu’s research group at ShanghaiTech University as an undergraduate researcher.

Publications

GiLT: Augmenting Transformer Language Models with Dependency Graphs

ACL 2026 Main Conference

Tianyu Huang, Yida Zhao, Chuyan Zhou, Kewei Tu [Paper]

Training-free Verifiable Process Reward for LLM Reinforcement Finetuning

Thesis · ShanghaiTech University, BIGAI

2023.10 - 2026.7 (internship)

In modern LLM RL finetuning, ORMs are common of rewarding scheme to predict by the final outcome of the task. Though simple, ORMs suffer reward sparsity causing LLMs’ failure to generate useful thinking process, leading to possibly degenerate performance. PRMs are to predict the reward by intermediate steps of the LLMs, providing more fine-grained/informative feedback.
However, existing PRMs either (1) are black-box neural models, requiring human annotations & training; (2) are training-free/nonparametric by statistical heuristics, while trajectories are segmented into semantically incomplete units; (3) are aggregated into some outcome reward for each trajectory, not really doing credit assignment.
We designed a training-free verifiable process rewarding scheme based on formalized theorem provers, consisting of:
- Forcing LLMs to generating natural language reasoning steps in Horn clause $p\land q\to r$ form, enabling semantically complete segmantation of the reasoning process;
- A training-free process reward by integrating a theorem prover/solver called a verifier;
- An LLM teacher in-the-loop to formalize NL reasoning into verifiers;
- A GRPO-like advantage normalizing algorithm for critic-free value estimation, but for PRMs.
We base experiments on logical/math QA datasets to train reasoning models, and evaluate and compare the performance by accuracy improvement after RL finetuning with our proposed process reward.

Augmenting Transformer Language Models with Dependency Graphs

ACL 2026 Main Conference · ShanghaiTech University

In this project, we proposed a novel method to augment Transformer-based syntactic language models with dependency graphs, especially with word-synchronous semantic dependency graphs, which is to model the dependencies and tokens jointly in a unified framework with a dependency-predicting head so that the original word sequences are kept intact.
Since the use of syntactic information in past works is limited to recursive structures i.e. trees, and graphs are more general structures which we consider have more potential information to be utilized, based on a codebase of Transformer-XL, we implemented the model architecture which operates on the dependency graphs and word sequences jointly.
As a contributor, I conducted baseline experiments on metrics such as perplexity (PTB 1987-89 WSJ Treebanks), SG and BLiMP tests.

More Projects

Reconstruction and Re-evaluation of SFCNN for PLA Scoring

COMPSCI 177 Research Project · ShanghaiTech University

2025.4 - Present

(2025.4) Constructed a PyTorch implementation of the model architecture, training, and evaluation of SFCNN (Scoring Function 3D Convolutional Neural Network) for protein-ligand binding affinity prediction.
(2025.4-6, expected) Developed a novel benchmark for similar models to allow inference directly on the 3D structure of protein-ligand complexes instead of on decoupled protein and ligand structures.

LLM-powered Lecture Generation

COMPSCI 194-196 Research Project · University of California, Berkeley

2024.8 - 2024.12

(2024.8-9) Independently developed the backend framework of the lecture generation pipeline using FastAPI as a deployable web service. This backend includes asynchronous task execution via multithreading, task management via an API powered by Redis databases, and a metadata system for managing generated data.
(2024.9-10) Worked as the main developer to integrate the respective model components on the backend framework.
(2024.10-11) Developed an additional LLM-powered QA agent based on the backend that interacts with LLMs using a long context of generated lectures and a RAG system to dynamically index grounding sources (e.g., textbooks).

Chuyan Zhou

Research Interest

Misc

News

Publications

GiLT: Augmenting Transformer Language Models with Dependency Graphs

Recent Posts

INFOTH Note 23: Parallel Gaussian Channels, WSS, Distributed Source Compression

INFOTH Note 22: AWGN Channel and Shannon Limit

INFOTH Note 21: Polar Codes and Related Theories (Bhattacharyya, Martingale)

INFOTH Note 20: Channel Coding Schemes

INFOTH Note 19: Channel Coding Theorem for DMC 2

Experience

Training-free Verifiable Process Reward for LLM Reinforcement Finetuning

Augmenting Transformer Language Models with Dependency Graphs

More Projects

Reconstruction and Re-evaluation of SFCNN for PLA Scoring

LLM-powered Lecture Generation