Hello
I develop methods for large language models, reinforcement learning, and multimodal scientific foundation models.
I am a Computer Science PhD student at the University of California, Irvine. My research spans large language models, reinforcement learning, multimodal learning, and genomic foundation models, with an emphasis on reliable evaluation and scientific applications.
I enjoy moving between methods and systems: designing objectives, building distributed training and evaluation infrastructure, and turning research questions into reproducible benchmarks. Before UCI, I earned an M.S. in Computer Engineering from UIUC and a B.S. in ECE with a Data Science minor from the University of Michigan–Shanghai Jiao Tong University Joint Institute.
M.S. · Computer Engineering
B.S. · ECE, Data Science minor
University of Michigan–Shanghai Jiao Tong University Joint Institute · Outstanding Graduate
LLMs + reinforcement learning
Genomic AI
Knowledge distillation
Multimodal reasoning
AI evaluation
Broader technical interests
My earlier work and continuing interests also include bioinformatics, molecular optimization, pangenome, protein and single-cell foundation models, biomedical lay summarization, tokenization and positional encodings, normalization, graph neural networks, diffusion models, computer vision, distributed machine learning, and computer architecture.
Updates
News
Recent research, releases, and milestones.
-
Joined Pandita AI (diagram.ai)
Working as an LLM Research Scientist Intern on verifiable evaluation and post-training for scientific diagrams.
-
MUGO accepted to KDD 2026
Differentiable combinatorial optimization for causal variant discovery in the non-coding genome.
-
A Very Big Video Reasoning Suite accepted to ICML 2026
A large-scale suite of procedurally generated tasks for long-horizon video reasoning.
-
Team XSZ selected for an oral presentation at BioLaySumm 2025
Section-wise retrieval, LLM generation, and reinforcement learning for biomedical lay summaries.
-
ACM SIGBio Best Paper Award
For work on L2 normalization and geodesic distance in high-dimensional single-cell visualization.
Current work
Research
Selected ongoing projects. Unreleased work is intentionally described at a high level.
Under review · Science Advances
OmegaGenome
Toward better sub-million-scale expert models from large genomic language models.
I lead this effort to distill genomic foundation models into deployable task experts. The work spans end-to-end training, controlled evaluation, and efficiency analysis across classification and base-resolution sequence prediction.
- Knowledge distillation across diverse genomic teacher architectures
- 18 classification tasks plus base-resolution multi-track regression
- Reproducible training, benchmarking, and systems-efficiency pipelines
Teacher
Genomic foundation model
general sequence representations
Student
Compact expert model
task-specific and deployable
DNAThinker
A reasoning-grounded multimodal foundation model for cell-type-conditioned DNA generation, editing, and pair prediction. I lead the model, data, and evaluation effort.
On-policy distillation for long-horizon agents
Studying stable knowledge transfer in multi-turn agent trajectories, supported by multi-GPU training and evaluation infrastructure built with FSDP, vLLM, Ray, and SLURM.
DSV: verifiable rewards for scientific diagrams
Developing semantic verifiers for scientific diagrams and using their structured feedback as rewards for reinforcement-learning post-training and iterative repair.
Scalable, shape-aware distillation
Running controlled multi-domain studies across NLP and single-cell models, with multi-seed baselines, negative controls, and reproducible cluster-scale experiments.
Pipe-Déjàvu
Hardware-aware, latency-predictable differentiable search for faster configuration and convergence of distributed machine-learning pipeline parallelism.
MIRACLE
Interpretable multi-task learning for identifying shared epigenetic regulation across autoimmune diseases, with site–gene–pathway structure built into the analysis.
AgenticVBench · game understanding
I designed two proposals for long-horizon video understanding: race-telemetry reconstruction in SuperTuxKart and action-ledger reconstruction in Minecraft. Each pairs seeded generators with machine-exact ground truth, deterministic graders, and anti-shortcut calibration.
Kart proposal #73 Minecraft proposal #74 Kart branch Minecraft branch
Method lab
Distillation, from objective to rollout
Inspect the manuscript-derived OmegaGenome loss anatomy and established on-policy curricula for multi-turn agents. Unreleased results are intentionally omitted.
Selected work
Publications
Peer-reviewed and accepted work, followed by selected public preprints.
KDD 2026
MUGO: Differentiable Combinatorial Optimization for Causal Variant Discovery in the Non-coding Genome
Recasts causal-variant discovery as efficient differentiable search and evaluates it across multiple molecular modalities and tissues.
BioLaySumm 2025 · Oral
Team XSZ at BioLaySumm2025: Section-Wise Summarization, Retrieval-Augmented LLM, and Reinforcement Learning Fine-Tuning for Lay Summaries
A section-aware pipeline combining retrieval-augmented generation and reinforcement learning for readable, faithful biomedical summaries.
Preview: Figure 1 from the paper (CC BY 4.0).
ACM BCB 2024 · SIGBio Best Paper Award
L2 Normalization and Geodesic Distance for Enhanced Information Preservation in Visualizing High-dimensional Single-cell Sequencing Data
Studies how normalization and geodesic structure improve information preservation in high-dimensional biological visualization.
Method preview adapted from Figures 1–2; background from Figure 5.
Interactive method note Why normalization changes the geometry Trace x → z → θ → w → p → embedding; κ visibly reshapes the toy neighborhood.
KDD AIDSH 2024 · Oral
REINVENT-Transformer: Molecular de-novo design through Transformer-based Reinforcement Learning
Combines a Transformer policy with oracle-feedback reinforcement learning for goal-directed molecular generation.
Paper controls
Two published methods, made tangible
Inspect the optimization mechanics and reported ablations behind two selected papers. Illustrative quantities and published measurements are explicitly separated.
Published method MUGO · differentiable variant selection Inspect temperature and straight-through optimization.
Published method Multi-objective RL for biomedical lay summaries Inspect the public reward design and validation ablation.
Earlier & public preprint work
- 2024 · Preprint MIRACLE: Multi-task learning based Interpretable Regulation of Autoimmune diseases through Common Latent Epigenetics
- 2020 · ICVRV
Practice
Experience
Research roles, engineering systems, and teaching.
Jul 2026—Present
LLM Research Scientist Intern · Pandita AI / diagram.ai
Developing semantic verification and reinforcement-learning post-training for mathematical and scientific diagram generation.
Aug 2024—Jul 2025
Research Scientist Intern · XtalPi
Developed and benchmarked antibody binding-affinity models, with leakage-resistant data splits, balanced objectives, encoder integration, and reproducible evaluation.
May—Aug 2023
Software Engineer Intern · Amazon Web Services
Built a React/TypeScript rate-card workflow and Java/AWS backend integrations for VMware Cloud on AWS; added data-validation and anomaly-detection tooling.
May—Aug 2022
Multimodal Cognitive Computing Algorithm Intern · Shanghai AI Laboratory
Worked on multimodal target detection with zero-shot depth estimation and multimodal neural architecture search.
Jul—Oct 2022
Open-Source Software Developer · 4Paradigm / OpenMLDB
Built an automated feature-generation and selection workflow with Python and OpenMLDB SQL, then contributed it upstream and presented the system to the open-source community.
OpenMLDB contribution OpenMLDB repository Meetup talk Code Camp talk
Watch the OpenMLDB Meetup talk
Nov 2021—Jun 2022
Deep Learning Software Engineer Intern · Intel Corporation
Contributed model-compression and quantization work and documentation to Intel Neural Compressor; studied inference-server architecture and helped implement C++ inference tooling.
Dec 2020—Apr 2021
Algorithm Intern · Shukun Technology
Built multi-node, multi-GPU 3D U-Net training with Horovod, OpenMPI/NCCL, and NVIDIA Clara, and added distributed-training backend support for medical-imaging workloads.
Teaching
Teaching Assistant for UCI ICS 6B (Boolean Logic & Discrete Structures), UCI ICS 6D (Discrete Mathematics), and University of Michigan–Shanghai Jiao Tong University Joint Institute VE370 (Computer Organization).
Technical toolkit
Python, C/C++, TypeScript/JavaScript, Java, SQL, shell, CUDA, MATLAB, Verilog, and R · PyTorch, TensorFlow, Keras, scikit-learn, Pandas, Horovod, FSDP, vLLM, Ray, and SLURM · distributed training, evaluation, and benchmark design.
Selected honors
- ACM SIGBio Best Paper Award · 2024
- Shanghai Jiao Tong University Outstanding Graduate · 2022
- Microsoft Imagine Cup, third prize in China · 2021
- Mathematical Contest in Modeling, Meritorious Winner · 2020 PDF
- CCVR “Jidong Cup,” product creative group, second prize · 2020
- Undergraduate Excellence Scholarship · 2018–2019 and 2020–2021
Outside the paper
Interactive builds
Things I ship to learn, explain, and play.
01
Game · full-stack systems
Photon 47
A solo-built, no-install browser RPG with six game modes, including real-time multiplayer battles. It combines WebGL rendering, game systems, Deno services, serverless Postgres, and automated browser testing.
Interactive field guide
The Geometry of Intelligence
A bilingual, visual guide to Riemannian and information geometry for AI, with interactive diagrams and intuition-first explanations.
Explore the guide ↗Interactive mathematical essay
Particles → Probability
An interactive tour from interacting particles to probability, built around the mathematics of Fields Medalist Yu Deng.
Read the essay ↗Developer tool
File → Prompt
A private-by-design browser tool that builds a file tree, orders project files by folder, dependency, or name, and turns them into one structured prompt. Everything stays local.
Open the tool ↗Conversational experiment
PengchengGPT
A research-focused portfolio assistant with current UCI/CV context, offline profile commands, and a free-first model route with an explicit visitor-key fallback.
Small worlds, shipped
Four browser games on itch.io
A little more human
Beyond research
Music, writing, and the story behind my name.
徐鹏程
A name about a long journey
“Pengcheng” pairs peng—the giant bird of Chinese mythology—with cheng, a journey. My parents chose it as a hope for an ambitious path and a wide horizon. In English, I pronounce my surname like “Hsu.”
Music & talks
I sing, play guitar, and record occasional technology talks. Away from a screen, I enjoy working out, basketball, tennis, table tennis, swimming, reading, science fiction, and any excuse to learn how something works. Richard Feynman and Tsung-Dao Lee are enduring inspirations.
Watch on Bilibili ↗Selected writing
祂 is a gender-neutral Chinese pronoun sometimes used when referring to God; Father Sun and the original Chinese version are two forms of the same science-fiction project.
Video
A compact, click-to-load player; no third-party video request is made until you choose it.
Watch on Bilibili
PDF · selected writing
Writing
Good “New” Days
A first-page preview keeps the writing visible without loading a full PDF viewer.
Read PDFPersonal note retained from the original site
Fun fact: my MBTI result has moved between INTJ and ENTJ over time. More importantly, I try to make work that contributes something positive and outlasts the moment in which it was made.
有的时候会惧怕死亡，每个人终有一死，无一例外，如何让自己对于死亡的恐惧少一些呢？于我而言就是为社会作出尽可能多的正向贡献。 这样即使死去了，这些工作依然会留在人间，为社会上每个人的生活带来点滴的进步。就是这么一点一滴的工作才汇成了时代进步的滚滚洪流。 这些工作会代替我们留在人间。于是每能做出一些好的工作我就会对于死亡的恐惧少一点。因为我的存在或许能让世界变好那么一点点。 好像我的存在就能超脱于我的肉体而存在更久了。希望这些工作能表达我对于人类的爱，做出的这些工作就像是把对人类每一个人的爱， 写成一首情诗，嵌入在了社会大厦的一个小小砖石中。虽然路还很远自己目前做的工作根本不算什么。但我希望有一天我即使面对死亡也能 安心地说：我已经把对世界上每一个人的爱，写在了我的工作里，即使死亡也不能带走我已经有的幸福了，此生无憾矣。
Get in touch
Let’s build something rigorous—and useful.
I welcome conversations about research collaborations, reliable AI evaluation, genomics, distillation, and scientific machine learning.