research console
research console · online
中文版 Academic view

system.profile / online

Pengcheng Xu Researching how models learn, reason, and transfer.

I develop methods for large language models, reinforcement learning, and multimodal scientific foundation models. I am a Computer Science PhD student at UC Irvine, with an emphasis on reliable evaluation.

Enter research workspace pengcheng.xu@uci.edu
px@uci:~/research

Curated commands only—this terminal does not execute code. Try help.

trace://research-loopillustrative · not live telemetry

model multimodal objective transfer + verify checkpoint reproducible interface sketch · not live telemetry
01 / learn

Post-training

Reinforcement learning and verifiable feedback for language, scientific, and agentic systems.

02 / transfer

Distillation

Turning large general models into smaller, efficient experts without losing what matters.

03 / understand

Models for science

Foundation models and evaluations grounded in genomics, biology, diagrams, and geometry.

profile.yaml education + trajectory

Quick route

stream.01 / latest

News feed

Accepted work, research roles, and recent milestones—without the notification noise.

  1. role.update

    Joined Pandita AI / diagram.ai

    LLM Research Scientist Intern working on verifiable evaluation and post-training for scientific diagrams.

  2. paper.accepted

    MUGO accepted to KDD 2026

    Differentiable combinatorial optimization for causal variant discovery in the non-coding genome.

  3. paper.accepted

    A Very Big Video Reasoning Suite accepted to ICML 2026

    Procedurally generated tasks for long-horizon video reasoning.

  4. paper.oral

    Team XSZ accepted to the ACL 2025 BioNLP Workshop

    Our oral paper for the BioLaySumm 2025 shared task combines section-wise retrieval, LLM generation, and reinforcement learning for biomedical lay summaries.

  5. award.received

    ACM SIGBio Best Paper Award at ACM BCB 2024

    Our work on L2 normalization and geodesic distance in high-dimensional single-cell visualization received the award at the 15th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics (ACM BCB 2024).

workspace.02 / current

Research programs

Methods and systems work, organized by the scientific question rather than the application domain.

grouped view: 8 projects shown
under review · Science Advances

R/01

OmegaGenome

Project lead · model distillation, evaluation, and systems

genomic AIknowledge distillationscientific models

Toward sub-million-parameter expert models distilled from large genomic language models, with controlled evaluation across classification and base-resolution tasks.

  • 18 classification tasks
  • Multi-track sequence prediction
  • End-to-end efficiency analysis
to be submitted · planned for ICLR

R/02

DNAThinker

Project lead · model, data, training, and evaluation

multimodal learninggenomic AILLM methods

A reasoning-grounded, multimodal foundation model for cell-type-conditioned DNA generation, editing, and pair prediction. I lead the model, data, and evaluation effort.

  • Sequence, cell context, and natural-language reasoning
  • Generation, editing, and pair-prediction tasks
  • Unreleased results intentionally omitted
generatereasonverify
ongoing at diagram.ai

R/03

DSV: verifiable rewards for scientific diagrams

LLM Research Scientist Intern · diagram.ai

LLMreinforcement learningmultimodal evaluation

Semantic verifiers turn structured diagram feedback into rewards for reinforcement-learning post-training and iterative repair. The public materials emphasize evaluation design and reproducible experiments; unreleased results remain high level.

renderverifyrewardrepair

Experiment log ↗ TikZ workspace ↗

research in progress

R/04

Scalable, shape-aware distillation

Independent study · controlled empirical research

knowledge distillationLLMrepresentation geometry

Controlled multi-domain studies across NLP and single-cell models, with multi-seed baselines, negative controls, and reproducible cluster-scale experiments.

  • Teacher–student scale trends
  • Representation-shape and objective comparisons
  • Public experiment plan, report, and research goal

Experiments ↗ Report ↗ Research goal ↗

research in progress

R/05

On-policy distillation for long-horizon agents

LLM agents · knowledge distillation · on-policy training

LLMknowledge distillationon-policy distillationagents

Studying stable knowledge transfer across multi-turn agent trajectories, supported by multi-GPU training and evaluation infrastructure built with FSDP, vLLM, Ray, and SLURM.

  • Policy-generated multi-turn trajectories
  • Controlled teacher–student transfer
  • Held-out behavioral evaluation
rolloutdistillevaluate
technical report · 2023

R/06

Pipe-Déjàvu

Technical report · distributed ML systems

ML systemspipeline parallelism

Hardware-aware, latency-predictable differentiable search for faster configuration and convergence of distributed machine-learning pipeline parallelism.

  • Pipeline configuration as a differentiable search problem
  • Hardware-aware latency prediction

Research report ↗ Project visual ↗

public preprint

R/07

MIRACLE

Public preprint · interpretable computational biology

scientific modelsinterpretable MLgenomics

Interpretable multi-task learning for shared epigenetic regulation across autoimmune diseases, with site–gene–pathway structure built into the analysis.

  • Shared learning across related disease tasks
  • Biological interpretation from site to pathway

Preprint ↗ Repository ↗ Project visual ↗

open task proposals

R/08

AgenticVBench · game understanding

Open benchmark contributions · Kart and Minecraft proposals

agentsmultimodal evaluationvideo reasoning

I designed long-horizon video-understanding proposals for race-telemetry reconstruction in SuperTuxKart and action-ledger reconstruction in Minecraft, with seeded generators, machine-exact ground truth, deterministic graders, and anti-shortcut calibration.

  • Procedurally generated, reproducible episodes
  • Machine-exact outputs and deterministic graders
  • Difficulty calibration against shortcut strategies

Kart issue #73 ↗ Minecraft issue #74 ↗ Kart branch ↗ Minecraft branch ↗

method.lab / transfer

Distillation, from objective to rollout.

Two inspectable method notes: the manuscript-derived OmegaGenome loss anatomy and established on-policy curricula for multi-turn agents. Unreleased results are intentionally omitted.

archive.03 / selected

Published work

Selected peer-reviewed work across genomic discovery, video reasoning, language models, and single-cell geometry.

MUGO differentiable causal variant discovery framework

KDD 2026 · accepted

MUGO: Differentiable Combinatorial Optimization for Causal Variant Discovery in the Non-coding Genome

Efficient differentiable search for causal variants across molecular modalities and tissues.

Read paper

ICML 2026 · accepted

A Very Big Video Reasoning Suite

Procedurally generated tasks and evaluation infrastructure for long-horizon video reasoning.

Paper ↗ Project ↗ ICML ↗
Team XSZ BioLaySumm retrieval, generation, and reinforcement-learning pipeline

ACL 2025 BioNLP Workshop · BioLaySumm Shared Task · oral

Team XSZ at BioLaySumm 2025

Section-wise retrieval, LLM generation, and reinforcement learning for accessible biomedical summaries.

Read paper
t-SNE and hyperspherical SNE embeddings across three single-cell datasets

ACM BCB 2024 · SIGBio Best Paper Award

L2 Normalization and Geodesic Distance for Enhanced Information Preservation in Visualizing High-dimensional Single-cell Sequencing Data

Project cells to the L2 hypersphere, compare them by angular distance, then use a spherical affinity inside SNE.

wij = exp[κ(cos θij − 1)]
Official PDF
REINVENT-Transformer molecular generation and reinforcement-learning workflow

KDD AIDSH 2024 · oral

REINVENT-Transformer

Transformer-based reinforcement learning for oracle-guided molecular de novo design.

Read paper
Evaluation results for biomedical lay summarization systems

ACL BioNLP 2024

Adapting Large Language Models for Biomedical Lay Summarization

Adapts language models for accessible biomedical communication, emphasizing readability and factual quality.

Publication ↗ PDF ↗

paper.controls / inspect

Published methods, under direct control.

Move a published optimization parameter and inspect an exact reported ablation. Illustrative quantities remain separate from experimental measurements.

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 RECORD

Author PDFs, full author lists, and figure notes Open the academic publication section →

paper.lab / interactive reconstruction

From geometry to neighborhood.

An interactive reading of our ACM BCB 2024 paper. Follow the same feature directions from a flat simplex to a curved hypersphere, then inspect how angular distance becomes a probability neighborhood.

ACM BCB 2024 · ACM SIGBio Best Paper Award

L2 Normalization and Geodesic Distance for Enhanced Information Preservation in Visualizing High-dimensional Single-cell Sequencing Data

The method projects each cell to the unit hypersphere, measures angular distance, transforms angles into spherical affinities, and normalizes those affinities before optimizing the low-dimensional embedding. The interaction below makes that causal chain inspectable.

Official PDF ↗ DOI record ↗ Method figure ↗ Embedding figure ↗ Evaluation figure ↗

Interactive paper note Connected geometry trace Projection, angular distance, affinity, neighborhood, and embedding stay linked.

The geometry control stays in view through this method trace; open Deeper math for κ, affinities, and the numbered path to the low-dimensional neighborhood.

Paper method figure comparing L1 hypersimplex and L2 hypersphere normalization, conditional probabilities, and the resulting embeddings
Method figures from the paper: normalization changes the geometric support, which changes neighbor probabilities before the two-dimensional optimization.
t-SNE and hyperspherical SNE embeddings across three single-cell datasets
Embedding comparisons from the paper’s reported experiments.
Distance separation and nearest-neighbor accuracy on the lung-cell dataset
Distance separation and nearest-neighbor accuracy from Figure 6. The interactive toy neighborhood above is an explanatory reconstruction, not a reported experiment.

systems.04 / field log

Research meets engineering

I move between objectives, datasets, distributed systems, product interfaces, and evaluation infrastructure.

research / post-training

LLM Research Scientist Intern · Pandita AI / diagram.ai

Built DSV, which evaluates mathematical and scientific diagrams as typed relational graphs and returns failure witnesses tied to source locations and rendered regions; developing verifier-guided reinforcement-learning post-training for generation and repair.

research / scientific ML

Research Scientist Intern · XtalPi

Developed antibody-binding ΔΔG models with binding-ddg-predictor, a CarbonDesign encoder, and GearBind; created leakage-resistant complex-level SKEMPIv2 splits and balanced alanine/non-alanine sampling, improving Pearson correlation by ~10% and Spearman by ~5%.

product systems

Software Engineer Intern · Amazon Web Services

Built a React/TypeScript rate-card workflow for VMware Cloud on AWS with Java, API Gateway, and Lambda integrations, including CSV validation and pricing-error previews; added Athena/S3/DynamoDB anomaly detection with CloudWatch alerts for usage and subscription anomalies.

open-source ML systems

Open-Source Software Developer · 4Paradigm / OpenMLDB

Built and upstreamed a 600+ line AutoX/Python/OpenMLDB SQL pipeline that generated time-series and statistical features, then selected top features through adversarial validation, GRN, or reinforcement learning; presented the system to the open-source community.

Contribution ↗ 4Paradigm ↗ OpenMLDB ↗ Meetup talk ↗ Code Camp talk ↗
Watch the OpenMLDB Meetup talk click to unfold

multimodal learning

Multimodal Cognitive Computing Algorithm Intern · Shanghai AI Laboratory

Developed multimodal target detection that consumes images and two audio channels to predict behavior and distance, integrating MiDaS zero-shot monocular depth estimation; also investigated multimodal neural architecture search.

Shanghai AI Laboratory ↗

model efficiency / open source

Deep Learning Software Engineer Intern · Intel Corporation

Implemented Cross-Layer Equalization for data-free FP32→INT8 quantization in Intel Neural Compressor; studied NVIDIA Triton and AI Model Efficiency Tool, presented their designs to ~100 colleagues, and helped build C++ multi-framework inference tooling for CPU/GPU deployment.

Intel Neural Compressor ↗

distributed medical AI

Algorithm Intern · Shukun Technology

Built Horovod multi-node, multi-GPU 3D U-Net training with NVIDIA Clara, OpenMPI, and NCCL2, reaching 2.5× speedup on four GPUs across two nodes; benchmarked configurations and added Java backend support for distributed medical-imaging jobs.

Project repository ↗ Shukun Technology ↗ Horovod ↗ Technical talk ↗

working stack

PyTorchFSDPvLLMRaySLURM CUDAC/C++TypeScriptSQL

TEACHING

Discrete structures + computer organization

Teaching Assistant for UCI ICS 6B (Boolean Logic & Discrete Structures, Winter 2025), UCI ICS 6D (Discrete Mathematics, Spring 2025), and the University of Michigan–Shanghai Jiao Tong University Joint Institute VE370 (Computer Organization, Fall 2021).

SELECTED HONORS

Research, modeling, and building

  • ACM BCB 2024 · ACM SIGBio Best Paper Award
  • Shanghai Jiao Tong University Outstanding Graduate
  • Microsoft Imagine Cup, third prize in China
  • CCVR “Jidong Cup,” product creative group, second prize
  • Mathematical Contest in Modeling, Meritorious Winner
  • Undergraduate Excellence Scholarship
MCM paper ↗

SYSTEMS MODE

From objective to reproducible run

Python, C/C++, Java, TypeScript, SQL, shell, CUDA, PyTorch, TensorFlow, Horovod, distributed training, evaluation, and benchmark design.

playground.05 / shipped

Builds with a pulse

Browser worlds, interactive mathematical essays, and tools for thinking with models.

Interactive geometric manifold visualization BUILD/02 · FIELD GUIDE The Geometry of Intelligence Riemannian and information geometry for AI, made visual. Particle field from the Particles to Probability interactive essay BUILD/03 · INTERACTIVE ESSAY Particles → Probability From interacting particles to the mathematics of Fields Medalist Yu Deng. BUILD/04 · DEVELOPER TOOL File → Prompt A private-by-design browser tool that orders project files and turns them into one structured prompt. Everything stays local. BUILD/05 · CONVERSATIONAL WEB PengchengGPT A conversational interface to my work, background, and public projects.

inspect source: PengchengGPT repository ↗

side.quest / itch.io

Four small worlds

View profile ↗
Photon 47Multiplayer RPG Wacky BrawlersCreature battles Rusted WarBrowser RTS OverwatchFPS experiment

offscreen.06 / human context

Beyond the lab

Music, speculative writing, browser experiments, and the story encoded in a name.

SELECTED PERSONAL MEDIA

Another kind of improvisation.

A guitar recording away from the research loop. The embedded player loads only after you choose it, so Bilibili receives no request on the initial page load.

Watch on Bilibili ↗
鹏, the mythic giant bird, begins the journey 程 The character first separates into its phonetic component 朋 and semantic component 鸟. A large 朋 contour, with its two 月 forms side by side, becomes the feather framework of one huge open left wing. A conventional bird silhouette forms the body while a lighter, semi-transparent 鸟 contour remains visible within it; 鸟's inner dot is highlighted as the eye before 鹏 faces 程. 01 · THE CHARACTER 鹏 02 · DECOMPOSE 朋 PHONETIC · SOUND 鸟 SEMANTIC · BIRD 03 · 朋 → GREAT LEFT WING · 鸟 → BODY TRACE + EYE 朋 鸟 鸟 · INNER DOT → EYE 程 THE JOURNEY 鹏程万里 A ROC'S VAST HORIZON
朋 glyph · 月月 → huge open left wing 鸟 character → body · inner dot → eye 鹏 mythic giant bird 程 journey

NAME / ORIGIN

鹏程万里: a vast horizon, built into my name

In 鹏, 鸟 gives the bird meaning while 朋 supplies the sound. In the animation, the two 月 forms remain side by side within a large 朋 contour, roughly matching 鸟 in height, and become the feather framework of one huge open left wing. A recognizable bird silhouette forms the body while a lighter, semi-transparent 鸟 remains visible inside it; the character’s inner dot is highlighted as the eye. Together they form a visual metaphor for 鹏, the enormous bird of Chinese mythology, setting out on a ten-thousand-li journey. My parents chose Pengcheng as a hope for an ambitious path and a wide horizon. In English, I pronounce Xu like “Hsu.”

OFFLINE / INTERESTS

Keeping more than one model of the world

Basketball, tennis, table tennis, swimming, reading, science fiction, and learning how things work. Richard Feynman and Tsung-Dao Lee remain enduring inspirations.

Why make things?

I try to make work that contributes something positive and outlasts the moment in which it was made. The full personal note remains in the academic archive.

channel.07 / open

Let’s investigate something difficult.

I welcome conversations about large language models, reinforcement learning, reliable AI evaluation, genomics, distillation, scientific machine learning, and unusually ambitious web experiments.

emailpengcheng.xu@uci.edu codegithub.com/explcre networkLinkedIn ↗ papersGoogle Scholar ↗ short formX / Twitter · @Ryan_xpc ↗ photosInstagram · @xpc_1025 ↗ videoBilibili channel ↗
Your message goes directly to my inbox.
calendar.open / conversation Schedule a meeting 30 or 60 min · Google Calendar · Local time

Choose 30 minutes for a focused question, or one hour for a deeper research, paper, or project conversation. The hour-long calendar also includes weekends.

Duration
30 or 60 minutes
Availability
30 min weekdays · 60 min daily
Time display
Your local time
Confirmation
Calendar event + email
Request a time by email

Live booking is handled by Google Calendar; the email fallback goes to my UC Irvine inbox. The calendar loads only after you open this panel.

Live Google Calendar

Open this panel to see current weekday and weekend availability.