Yuzhou Chang

His/Him

I am a computational biologist interested in spatial omics, graph signal processing, foundation models, and immuno-oncology. Welcome to connect.

Paper Tracking

An automated weekly digest. Every Monday an agent scans arxiv and HuggingFace Daily Papers, downloads the PDFs, and summarizes the top 3 picks per topic.

Week of October 5, 2026 · View archive →

Graph × LLM

Graph-structured reasoning, graph foundation models, and LLM-augmented GNNs.

HiTS-CL: A Continual Learning Framework for Long-Horizon Temporal Knowledge Graph Extrapolation

↑ 0 📚 10198 ★ 0 Sep 29

Problem: Extrapolative temporal knowledge graph reasoning (TKGR) predicts future facts from historical snapshots. Most existing methods train once on an early prefix of
Model:
Code: liuyansong98/HiTS-CL

GraphMAS: A Systematic Benchmark of Multi-Agent Coordination for Graph Learning

↑ 0 📚 1557 ★ 0 Sep 30

Problem: Understanding when and how multi-agent LLM coordination improves graph learning compared to single-agent approaches.
Model: GraphMAS: benchmark of multi-agent coordination for graph learning with seven coordination methods across four paradigms (Static/Dynamic × Independent/Interactive).
Code: not released

GraphVQ: Structure-Aware Autoregressive Decoding over Context-Quantized Graph Tokens

↑ 0 📚 1276 ★ 0 Sep 29

Problem: Autoregressive graph generation fails to capture long-range edges and cycles because serialization windows limit edge emission; global conditioning cannot distinguish candidate edges.
Model: GraphVQ: two-stage autoregressive decoder with VQ-VAE tokenization of node contexts and structure-aware pair-conditioned edge generation
Code: not released

Vision-Language Models

Multimodal VLMs, vision encoders, and image-text foundation models.

Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer Tokens

↑ 45 📚 10430 ★ 0 Sep 30

Problem: VLM-based robot agents incur substantial token overhead from repeated model invocations and redundant visual observations during task execution.
Model: PyRUA-Lean: interactive code-execution framework coupling feedback-driven primitive composition with selective observation for robot manipulation.
Code: DAGroup-PKU/PyRUA-Lean

Multimodal Flow: Unified Flow Modeling of Language and Vision in Embedding Spaces

↑ 27 📚 14588 ★ 0 Sep 29

Problem: Unified multimodal models either quantize images to discrete tokens or use separate objectives for language and vision; a fully continuous shared generative process remains underexplored.
Model: "Multimodal Flow (MF-1): fully continuous flow-based generative model using ordered hyperchunks and chunk-causal Flow Matching backbone for joint language-vision modeling in embedding spaces.
Code: hustvl/Multimodal-Flow

PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop

↑ 23 📚 10960 ★ 0 Sep 29

Problem: Vision-Language Models lack holistic evaluation of physical understanding across perception, reasoning, and assessment stages with real-world and generated videos.
Model: "PhysVista: benchmark for evaluating physical intelligence in VLMs through a closed-loop perception-reasoning-assessment framework"
Code: Helen1p/PhysVista

World Models

Predictive world models, JEPA-style learning, video generation, and embodied simulators.

World Observer: Joint Actor-Observer Generation for Persistent World Modeling

↑ 77 📚 3981 ★ 0 Oct 1

Problem: Existing video world models fail to maintain object states and dynamics after they leave the agent's view, limiting persistent world modeling.
Model: World Observer: decoupled actor-observer generation using a pretrained video Diffusion Transformer that jointly generates perspective actor and panoramic observer streams sharing a single world state.
Code: not released

FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation

↑ 19 📚 86589 ★ 0 Sep 29

Problem: Long-horizon video generation loses coherence when early visual details fall outside the model's input window; existing selection methods rely on current content rather than future information needs.
Model: FrameMorrow: prospective frame selector predicting compact prospective tokens representing future information needs to guide selection of relevant historical frames across diverse generators.
Code: not released

Ego2Act: Evaluating Goal-Directed Manipulation in Egocentric Video Generation

↑ 25 📚 2162 ★ 0 Oct 1

Problem: Existing benchmarks underexplore multi-step physical reasoning in egocentric video generation for goal-directed manipulation tasks.
Model: approach: Ego2Act benchmark with Ego2ActJudge, a reference-free evaluation pipeline assessing task completion and physics plausibility in egocentric video generation
Code: Comfy-Org/ComfyUI

Spatial Single-Cell Study

Single-cell + spatial omics foundation models, virtual cells, biological world models, and biomedical digital twins.

GATE-ST: Gene-Aware Text-image Encoder for Spatial Transcriptomics

↑ 0 📚 23935 ★ 0 Sep 30

Problem: Spatial gene expression prediction from histopathology images relies on expensive tests; text-based optimization for incorporating gene semantic information remains underexplored.
Model: GATE-ST: Gene-Aware Text-image Encoder for Spatial Transcriptomics—combines H&E image patches with gene text summaries via cross-attention layers to predict spatial gene expression.
Code: not released

CellMSA: Context Modeling for Single-Cell Representation Learning

↑ 0 📚 3818 ★ 0 Sep 30

Problem: Single-cell foundation models underutilize cross-batch and cross-cell-type information, limiting their ability to capture fine-grained gene dependencies.
Model: CellMSA: MSA-inspired framework that retrieves contextual cells across batches and related cell types to learn gene-pair representations for enhanced single-cell encoding.
Code: PharMolix/CellMSA

P2P: Cross-View Population Denoising for Unpaired Single-Cell Perturbation Response Prediction

↑ 0 📚 7476 ★ 0 Sep 28

Problem: Predicting population-level transcriptional responses to genetic perturbations from unpaired, noisy single-cell sequencing data.
Model: P2P: permutation-invariant set encoder with structured perturbation representation, memory-gated effect head, and cross-view population denoising supervision.
Code: not released

Autonomous Labs & Lab-in-the-Loop

Self-driving labs, closed-loop ML-guided experimentation, and agents that run real experimental campaigns.

Experimental Experience Modeling for Autonomous Research

↑ 0 📚 21750 ★ 0 Sep 30

Problem: Autonomous research agents lack systematic methods to decide which experiments warrant running when evidence is insufficient.
Model: Experimental Experience Modeling (EEM): framework for acquiring, reusing, and accumulating experimental experience to guide experimental decisions through historical retrieval, sufficiency assessment, targeted pilots, and iterative library growth.
Code: not released

esQueranto: Differentiable Structured Quantum Light for Automated Scientific Discovery

↑ 0 📚 1470 ★ 0 Sep 30

Problem: Existing quantum-optical simulators lack a unified framework combining photon-number quantum optics with structured-light propagation for automated experimental design.
Model: esQueranto: differentiable JAX-based simulator integrating Fock-state quantum optics, spatial beam geometry, and light propagation with automatic differentiation for optimization.
Code: artificial-scientist-lab/esQueranto

A roadmap for polymer informatics super-intelligence

↑ 0 📚 9 ★ 0 Sep 28

Problem: Polymer informatics lacks integrated systems for inverse design, causal reasoning across chemistry-processing-performance, and autonomous closed-loop experimentation.
Model: approach: modular agent-directed architecture with conversational LLM layer coordinating domain-specialized tools for neat polymers, composites, solvents, synthesis, and processing across molecular-to-product design hierarchy
Code: not released