Post-training
Reinforcement learning and verifiable feedback for language, scientific, and agentic systems.
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.
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
Distillation
Turning large general models into smaller, efficient experts without losing what matters.
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.
-
role.update
Joined Pandita AI / diagram.ai
LLM Research Scientist Intern working on verifiable evaluation and post-training for scientific diagrams.
-
paper.accepted
MUGO accepted to KDD 2026
Differentiable combinatorial optimization for causal variant discovery in the non-coding genome.
-
paper.accepted
A Very Big Video Reasoning Suite accepted to ICML 2026
Procedurally generated tasks for long-horizon video reasoning.
-
paper.oral
Team XSZ selected for a BioLaySumm 2025 oral
Section-wise retrieval, LLM generation, and reinforcement learning for biomedical lay summaries.
-
award.received
ACM SIGBio Best Paper Award
For work on L2 normalization and geodesic distance in high-dimensional single-cell visualization.
workspace.02 / current
Research programs
Methods and systems work, organized by the scientific question rather than the application domain.
under review · Science Advances
teacherFoundation
model
model
studentCompact
expert
expert
R/01
OmegaGenome
Project lead · model distillation, evaluation, and systems
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
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
generate→reason→verify
ongoing at diagram.ai
R/03
DSV: verifiable rewards for scientific diagrams
LLM Research Scientist Intern · diagram.ai
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.
render→verify→reward→repair
research in progress
R/04
Scalable, shape-aware distillation
Independent study · controlled empirical research
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
research in progress
R/05
On-policy distillation for long-horizon agents
LLM agents · knowledge distillation · on-policy training
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
rollout→distill→evaluate
technical report · 2023
R/06
Pipe-Déjàvu
Technical report · distributed ML systems
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
public preprint
R/07
MIRACLE
Public preprint · interpretable computational biology
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
open task proposals
R/08
AgenticVBench · game understanding
Open benchmark contributions · Kart and Minecraft proposals
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 Combinatorial Optimization for Causal Variant Discovery in the Non-coding Genome
Efficient differentiable search for causal variants across molecular modalities and tissues.
Read paper
Team XSZ at BioLaySumm 2025
Section-wise retrieval, LLM generation, and reinforcement learning for accessible biomedical summaries.
Read paper
L2 normalization and geodesic distance for single-cell visualization
Project cells to the L2 hypersphere, compare them by angular distance, then use a spherical affinity inside SNE.
wij = exp[κ(cos θij − 1)]
Publication
REINVENT-Transformer
Transformer-based reinforcement learning for oracle-guided molecular de novo design.
Read paper
Adapting Large Language Models for Biomedical Lay Summarization
Adapts language models for accessible biomedical communication, emphasizing readability and factual quality.
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. Move κ to see how fixed angular distances become a sharper or broader probability neighborhood.
ACM BCB 2024 · ACM SIGBio Best Paper Award
L2 normalization and geodesic distance for single-cell visualization
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.
Publication ↗ Author PDF ↗ Method figure ↗ Embedding figure ↗ Evaluation figure ↗
Interactive paper note Connected geometry trace Projection, angular distance, affinity, neighborhood, and embedding stay linked.
The geometry and κ controls stay in view through this method trace; follow the numbered formulas from projection to the low-dimensional neighborhood.
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
Verifiable evaluation and reinforcement-learning post-training for mathematical and scientific diagram generation.
research / scientific ML
Research Scientist Intern · XtalPi
Antibody binding-affinity models, leakage-resistant splits, balanced objectives, and reproducible evaluation.
product systems
Software Engineer Intern · Amazon Web Services
React/TypeScript workflows, Java/AWS backend integrations, validation, and anomaly detection for VMware Cloud on AWS.
open-source ML systems
Open-Source Software Developer · 4Paradigm / OpenMLDB
Built an automated feature-generation and selection workflow with Python and OpenMLDB SQL, contributed it upstream, and presented it to the open-source community.
Contribution ↗ 4Paradigm ↗ OpenMLDB ↗ Meetup talk ↗ Code Camp talk ↗multimodal learning
Multimodal Cognitive Computing Algorithm Intern · Shanghai AI Laboratory
Worked on multimodal target detection with zero-shot depth estimation and multimodal neural architecture search.
Shanghai AI Laboratory ↗model efficiency / open source
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.
Intel Neural Compressor ↗distributed medical AI
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.
Project repository ↗ Shukun Technology ↗ Horovod ↗ Technical talk ↗working stack
PyTorchFSDPvLLMRaySLURM
CUDAC/C++TypeScriptSQL
TEACHING
Discrete structures + computer organization
Teaching Assistant for UCI ICS 6B and ICS 6D, and University of Michigan–Shanghai Jiao Tong University Joint Institute VE370.
SELECTED HONORS
Research, modeling, and building
ACM SIGBio Best Paper Award · SJTU Outstanding Graduate · Microsoft Imagine Cup China third prize · Mathematical Contest in Modeling Meritorious Winner.
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.
live multiplayer / browser-native
BUILD/01
Photon 47
A solo-built, no-install browser RPG with six game modes, including real-time multiplayer battles—spanning WebGL rendering, Deno services, serverless Postgres, and automated browser testing.
BUILD/02 · FIELD GUIDE
The Geometry of Intelligence
Riemannian and information geometry for AI, made visual.
↗
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
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 ↗NAME / ORIGIN
A long journey, built into two characters
“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.”
AUDIO / VIDEO
Music & technical talks
I sing, play guitar, and record occasional technology talks.
- Fly Me to the Moon ↗
- 七里香 · Jay Chou ↗
- Modern Loneliness · Lauv ↗
- OpenMLDB feature engineering talk ↗
- OpenMLDB–GitLink Code Camp ↗
- Distributed ML with Horovod ↗
WORDS / SPECULATIVE
Selected writing
祂 and Father Sun are two forms of the same science-fiction project.
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.