Yuzhou Chang

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 Β· Archive

Every weekly run is preserved. Click a month to expand the weeks inside. Latest weeks at the top. ← Back to current week

June 2026 48 papers

Week of

Graph Γ— LLM

Vision-Language Models

World Models

Spatial Single-Cell Study

Week of

Graph Γ— LLM

Vision-Language Models

World Models

Kairos: A Native World Model Stack for Physical AI

↑ 35 β˜… 0 Jun 15
  • Problem: World models are transitioning from passive visual generators to foundational, operational infrastructure for Physical AI: they must natively acquire world know
  • Model:
  • Code: not released
  • ⚠ Interested, but agent could not fetch the PDF β€” summary based on abstract only.

Week of

Graph Γ— LLM

APEX: A Network-Native Time-Series Foundation Model for Forecasting and Anomaly Detection for Wireless Edge Operations

↑ 1 πŸ“š 150 β˜… 0 Jun 9
  • Problem: Generic time-series foundation models fail to capture wireless network telemetry characteristicsβ€”bursty, zero-inflated signals with cross-protocol dependencies.
  • Model: APEX: network-native decoder-only transformer for forecasting enterprise AP telemetry and anomaly detection, available in cloud (269M) and edge (10.5M) variants.
  • Code: not released

Detecting Differences Is Not Understanding Structure: Large Language Models Fail at Graph Isomorphism

↑ 0 β˜… 0 Jun 8
  • Problem: LLMs achieve high accuracy on graph isomorphism detection but fail to maintain permutation invariance when nodes are relabeled, suggesting pattern exploitation rather than genuine structural reasoning.
  • Model: approach: diagnostic evaluation protocol testing permutation invariance of LLMs (GPT-4o, Gemini, Llama) on graph isomorphism tasks across multiple serialization formats and prompting strategies.
  • Code: not released

Vision-Language Models

InterleaveThinker: Reinforcing Agentic Interleaved Generation

↑ 77 πŸ“š 42325 β˜… 0 Jun 10
  • Problem: Image generators cannot produce interleaved text-image sequences due to architectural constraints, limiting applications in visual narratives and embodied manipulation.
  • Model: InterleaveThinker: multi-agent framework with planner and critic agents that retrofits existing image generators for interleaved generation using GRPO-based trajectory optimization.
  • Code: zhengdian1/InterleaveThinker

LabVLA: Grounding Vision-Language-Action Models in Scientific Laboratories

↑ 53 πŸ“š 8682 β˜… 0 Jun 11
  • Problem: Existing vision-language-action models lack laboratory-specific training data and cannot accommodate diverse robot embodiments needed for scientific protocol execution.
  • Model: LabVLA: vision-language-action model combining Qwen3-VL-4B-Instruct backbone with FAST action token pretraining and flow matching posttraining via DiT action expert
  • Code: not released

World Models

Avatar V: Scaling Video-Reference Avatar Video Generation

↑ 4 β˜… 0 Jun 10
  • Problem: Existing avatar video generation methods condition on single static images, failing to capture dynamic behavioral patterns and identity nuances required for production-quality talking avatars.
  • Model: Avatar V: production-scale framework for video-reference-conditioned avatar generation using Diffusion Transformer with Sparse Reference Attention, motion representation stream, and identity-aware super-resolution refinement.
  • Code: not released

Scale Buys Interpolation, Structure Buys a Horizon: Certified Predictability for Equivariant World Models

↑ 0 πŸ“š 17151 β˜… 0 Jun 11
  • Problem: World models lack per-prediction trustworthiness certificates and prediction horizons; average error does not indicate whether specific predictions can be trusted or for how long.
  • Model: approach: Equivariant latent world models with computable multi-step certification via Lyapunov spectrum stratification, proving orbit-constant error under equivariance and horizon bounds T_j(Ξ΅)∼log(1/Ξ΅)/Ξ»_j
  • Code: not released

$\texttt{WEAVER}$, Better, Faster, Longer: An Effective World Model for Robotic Manipulation

↑ 2 β˜… 0 Jun 11
  • Problem: Existing robot world models fail to simultaneously achieve high fidelity, long-horizon consistency, and efficient inference for manipulation tasks.
  • Model: WEAVER (World Estimation Across Views for Embodied Reasoning): a multi-view world model combining flow matching, diffusion forcing, pretrained encoders, and latent reward prediction for robot manipulation.
  • Code: Lightning-AI/torchmetrics

Spatial Single-Cell Study

OCOO-T : A Simple and Scalable Virtual Cell Model for Transcriptional Perturbation Response Prediction

↑ 0 πŸ“š 329 β˜… 0 Jun 11
  • Problem: Predicting single-cell transcriptional responses to perturbations requires models that capture population-level distribution shifts without relying on complex auxiliary encoders or specialized latent spaces.
  • Model: OCOO-T: flow-matching-based Transformer model that directly denoises continuous gene expression profiles conditioned on perturbation embeddings, dosage, and cell covariates via adaptive layer normalization.
  • Code: not released

Adaptive spatial blocking for scalable clustering inference with applications to high-throughput spatial proteomics

↑ 0 β˜… 0 Jun 10
  • Problem: Existing Ripley's K-function methods for spatial clustering are computationally prohibitive for large-scale spatial proteomics data due to O(nΒ²) complexity.
  • Model: "B-KAMP" (block-based KAMP): adaptive spatial blocking algorithm aggregating clustering evidence across disjoint rectangular blocks with asymptotic normal inference.
  • Code: mingyugo/B_KAMP

Week of

Graph Γ— LLM

When Graph Tokens Sink: A Mechanistic Analysis of Graph Language Models

↑ 2 πŸ“š 1201 β˜… 0 Jun 2
  • Problem: Graph Language Models may develop internal pathologies where graph tokens become activation outliers without meaningfully representing graph structure.
  • Model: approach: mechanistic interpretability analysis of Graph Language Models (LLaGA, TEA-GLM) through graph sink token detection and intervention experiments
  • Code: not released

The Post-GCN Decade Revisited: Curvature-Stratified Evaluation of Relational Learning

↑ 0 πŸ“š 6199 β˜… 0 Jun 4
  • Problem: Standard benchmarks average model performance across heterogeneous datasets, obscuring geometry-dependent performance variations and misleading conclusions about generalization.
  • Model: "CurvBench": curvature-stratified evaluation framework partitioning datasets by intrinsic geometry (positive, negative, near-zero curvature) to reveal geometry-dependent model performance trade-offs.
  • Code: https://sirbabbage.github.io/CurvBench_HOME/

A Graph Foundation Model with Spectral Parsing and Prototype-Guided Spatial Propagation

↑ 0 πŸ“š 12963 β˜… 0 Jun 2
  • Problem: Graph foundation models struggle with cross-graph transfer due to diverse graph structures and entangled spectral components requiring different propagation behaviors.
  • Model: SPG: graph foundation model combining learnable Chebyshev spectral filters for feature decomposition with Gromov-Wasserstein prototype geometry for transferable structural knowledge.
  • Code: not released

Vision-Language Models

LoomVideo: Unifying Multimodal Inputs into Video Generation and Editing

↑ 21 πŸ“š 27459 β˜… 0 Jun 3
  • Problem: Existing unified video generation and editing models are computationally expensive, relying on massive parameters and token concatenation that quadruples self-attention complexity.
  • Model: LoomVideo: 5B-parameter unified video generation and editing architecture using MLLM encoder, Deepstack injection, and zero-overhead Scale-and-Add conditioning.
  • Code: MSALab-PKU/LoomVideo

Personal AI Agent for Camera Roll VQA

↑ 19 πŸ“š 7113 β˜… 0 Jun 2
  • Problem: AI assistants cannot efficiently answer personalized questions over thousands of personal camera roll photos spanning years.
  • Model: camroll-agent: conversational AI agent with hierarchical memory and tools for efficient navigation over large personalized visual memory
  • Code: not released

Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators

↑ 9 πŸ“š 3973 β˜… 0 Jun 4
  • Problem: Vision-Language Models struggle with spatial reasoning beyond observed images, failing to infer unobserved layouts and reason from alternative viewpoints with limited egocentric observations.
  • Model: "Astra: agentic spatial reasoning framework coupling Astra-VL (RL-trained Qwen3-VL policy) with Astra-WM (Bagel-based world simulator for action-conditioned novel-view generation with view consistency tuning)"
  • Code: not released

World Models

Dream.exe: Can Video Generation Models Dream Executable Robot Manipulation?

↑ 16 πŸ“š 4009 β˜… 0 Jun 3
  • Problem: Standard video generation benchmarks measure visual quality, not whether generated robot manipulation videos produce executable physical behavior.
  • Model: "Dream.exe: evaluation framework for video-to-execution grounding in robotic manipulation, combining video assessment, trajectory extraction, and physics simulator execution.
  • Code: not released

Towards World Models in Biomedical Research

↑ 0 πŸ“š 38743 β˜… 0 Jun 4
  • Problem: Current biomedical AI focuses on static pattern recognition rather than simulating how biological systems evolve under interventions and perturbations.
  • Model: approach: biomedical world models that learn latent representations of biological states and intervention-conditioned dynamics to simulate future trajectories
  • Code: not released

PiL-World: A Chunk-Wise World Model for VLA Policy-in-the-Loop Evaluation

↑ 0 πŸ“š 5409 β˜… 0 Jun 4
  • Problem: Existing world models for robot evaluation use open-loop prediction, but VLA policies operate in closed-loop feedback; a world model for policy-in-the-loop evaluation is needed.
  • Model: "PiL-World": chunk-wise world model for closed-loop VLA evaluation using action-derived visual control, latent multi-view history conditioning, and joint multi-view prediction.
  • Code: not released

Spatial Single-Cell Study

Do Foundation Models See Biology? Evaluating Attention Coherence with Spatial Transcriptomics in Glioblastoma

↑ 0 πŸ“š 3679 β˜… 0 Jun 3
  • Problem: Whether attention maps from pathology foundation models capture genuine biological signals remains unknown, hindering clinical trust and regulatory approval.
  • Model: approach: spatial transcriptomics-based framework for objective evaluation of attention coherence in pathology foundation models using attention-based multiple instance learning
  • Code: not released

GC-MoE: Genomics-Guided Cell-Type-Specific Mixture of Experts for Histology-Based Single-Cell Spatial Transcriptomics

↑ 0 πŸ“š 1102 β˜… 0 Jun 1
  • Problem: Predict gene expression for individual cells from histology images and cell locations, accounting for cell-type-dependent expression variability.
  • Model: GC-MoE: Genomics-Guided Cell-Type-Specific Mixture of Experts that routes cells to type-specific experts and incorporates cell-type co-expression priors from scRNA-seq data.
  • Code: not released

Week of

Graph Γ— LLM

Vision-Language Models

Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments

↑ 117 πŸ“š 674 β˜… 0 May 27
  • Problem: Embodied AI systems are fragmented across task families and robot embodiments, limiting generalization across manipulation, navigation, and diverse platforms.
  • Model: Qwen-VLA: unified vision-language-action model extending Qwen3.5-4B with DiT-based flow-matching action decoder for cross-task, cross-embodiment embodied control.
  • Code: QwenLM/Qwen-VLA

Why Far Looks Up: Probing Spatial Representation in Vision-Language Models

↑ 45 πŸ“š 9553 β˜… 0 May 27
  • Problem: Vision-language models achieve high spatial reasoning benchmark scores but may rely on statistical shortcuts rather than structured 3D understanding.
  • Model: approach: Representation-level analysis framework using minimal contrastive pairs to measure spatial axis organization and disentanglement in VLM embeddings, plus SpatialTunnel synthetic benchmark.
  • Code: not released

World Models

minWM: A Full-Stack Open-Source Framework for Real-Time Interactive Video World Models

↑ 50 πŸ“š 36656 β˜… 0 May 27
  • Problem: Converting video diffusion foundation models into real-time interactive world models requires scattered techniques across data, training, and inference pipelines.
  • Model: minWM: full-stack framework converting T2V/TI2V bidirectional diffusion models into camera-controllable few-step autoregressive video world models via Causal Forcing/Forcing++ distillation
  • Code: shengshu-ai/minWM

YoCausal: How Far is Video Generation from World Model? A Causality Perspective

↑ 42 πŸ“š 16919 β˜… 0 May 27
  • Problem: Video diffusion models may perceive temporal direction without understanding causality; existing benchmarks rely on synthetic data with limited real-world generalization.
  • Model: YoCausal: two-level benchmark with Reverse Surprise Index (RSI) and Causality Cognition Index (CCI) metrics for evaluating causal cognition in video diffusion models.
  • Code: genmoai/models

AdaState: Self-Evolving Anchors for Streaming Video Generation

↑ 6 πŸ“š 1390 β˜… 0 May 27
  • Problem: Autoregressive video diffusion models anchor to static first frames, suppressing dynamics and locking scene composition despite natural evolution during generation.
  • Model: "AdaState": replaces static first-frame anchor with adaptive latent state that denoises alongside content at each chunk, evolving with generated scenes via relative time formulation.
  • Code: not released
May 2026 22 papers

Week of

Graph Γ— LLM

S2Aligner: Pair-Efficient and Transferable Pre-Training for Sparse Text-Attributed Graphs

↑ 0 πŸ“š 2080 β˜… 0 May 18
  • Problem: Graph foundation models struggle with sparse text-attributed graphs where node texts are missing, noisy, or uneven, causing unreliable structure-semantics alignment and transfer bias.
  • Model: S2Aligner: sparsity-aware and structure-enhanced LLM-as-Aligner framework that decouples semantic alignment from structural modeling via content-structure factorization and sparsity-aware cross-domain risk balancing.
  • Code: not released

Deep Neural Sheaf Diffusion

↑ 0 πŸ“š 1013 β˜… 0 May 18
  • Problem: Scaling Graph Neural Networks to depth is hindered by representation collapse and vanishing signals in existing sheaf diffusion methods.
  • Model: Deep Neural Sheaf Diffusion (DNSD): sheaf-based GNN replacing sheaf Laplacian with adjacency operator, adding normalization, odd nonlinearities, and gating to maintain informative signals across layers.
  • Code: not released

Vision-Language Models

LatentOmni: Rethinking Omni-Modal Understanding via Unified Audio-Visual Latent Reasoning

↑ 42 πŸ“š 15499 β˜… 0 May 20
  • Problem: Current multimodal LLMs struggle with audio-visual reasoning because text-based chain-of-thought compresses continuous signals into discrete tokens, losing temporal grounding.
  • Model: LatentOmni: cross-modal reasoning framework interleaving textual reasoning with audio-visual latent states, using feature-level supervision and Omni-Sync Position Embedding for temporal alignment.
  • Code: not released

RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution

↑ 9 πŸ“š 19747 β˜… 0 May 19
  • Problem: Discrete autoregressive text-to-image models suffer from latent covariate shift when optimizing only the policy with a frozen decoder, causing alignment-fidelity trade-offs.
  • Model: RankE: End-to-end post-training framework for discrete text-to-image generation that co-evolves the AR policy and VQ decoder through alternating ranking-based optimization.
  • Code: not released

World Models

Q-ARVD: Quantizing Autoregressive Video Diffusion Models

↑ 19 πŸ“š 13800 β˜… 0 May 20
  • Problem: Quantizing autoregressive video diffusion models is unexplored; standard quantization schemes designed for bidirectional diffusion transformers perform suboptimally on ARVDs.
  • Model: Q-ARVD: quantization framework for autoregressive video diffusion models using final-quality-guided frame-weighting and outlier-aware adaptive dual-scale quantization
  • Code: not released

Efficient Agentic Reasoning Through Self-Regulated Simulative Planning

↑ 7 πŸ“š 41260 β˜… 0 May 21
  • Problem: Current agentic LLMs lack control over when and how to plan, causing inefficient token use without reliable accuracy gains.
  • Model: "SRΒ²AM" (Self-Regulated Simulative Reasoning Agentic LLM): decomposes decision-making into three systemsβ€”simulative reasoning via world model, self-regulation via learned configurator, and reactive executionβ€”implemented as distinct chain-of-thought stages within an LLM.
  • Code: sailing-lab/sr2am

Spatial Single-Cell Study

AgentCo-op: Retrieval-Based Synthesis of Interoperable Multi-Agent Workflows

↑ 0 πŸ“š 20431 β˜… 0 May 19
  • Problem: Multi-agent workflow design in open-ended scientific settings lacks curated training sets, reliable metrics, and standardized interfaces between tools and agents.
  • Model: AgentCo-op: retrieval-based synthesis framework that composes reusable skills, tools, and external agents into executable workflows through typed artifact handoffs and bounded evidence-guided local repair.
  • Code: ma-compbio-lab/AgentCo-Op

Week of

Graph Γ— LLM

Vision-Language Models

MMSkills: Towards Multimodal Skills for General Visual Agents

↑ 99 πŸ“š 30599 β˜… 0 May 13
  • Problem: Visual agents need reusable multimodal procedural knowledge that binds actions to visual state recognition and decision-making, beyond text-only skills.
  • Model: "MMSkills: framework for representing, generating, and utilizing reusable multimodal procedures for visual agents. Each skill couples textual procedures with runtime state cards and multi-view keyframes, generated via trajectory-to-skill generator and consulted via branch loading.
  • Code: DeepExperience/MMSkills

MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models

↑ 71 πŸ“š 3089 β˜… 0 May 13
  • Problem: No benchmark systematically compares long-context LVLMs and memory-augmented agents on multimodal multi-session conversations requiring visual evidence.
  • Model: "MemLens": benchmark with 789 questions across five memory abilities (information extraction, multi-session reasoning, temporal reasoning, knowledge update, answer refusal) at four context lengths (32K-256K tokens)
  • Code: xrenaf/MEMLENS

MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory

↑ 58 πŸ“š 1421 β˜… 0 May 13
  • Problem: Existing multimodal agent memory evaluations fail to assess whether agents preserve fine-grained visual evidence needed for reasoning over time.
  • Model: "MemEye: a visual-centric evaluation framework measuring visual evidence granularity and reasoning complexity in multimodal agent memory"
  • Code: not released
  • ⚠ Interested, but agent could not fetch the PDF β€” summary based on abstract only.

World Models

Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation

↑ 87 πŸ“š 36802 β˜… 0 May 14
  • Problem: Existing AR diffusion distillation methods require 4+ sampling steps; frame-wise 1–2 step generation needs efficient, scalable student initialization.
  • Model: "Causal Forcing++": AR diffusion distillation pipeline using causal consistency distillation for few-step student initialization, avoiding expensive full PF-ODE trajectory precomputation.
  • Code: thu-ml/Causal-Forcing

Warp-as-History: Generalizable Camera-Controlled Video Generation from One Training Video

↑ 38 πŸ“š 18942 β˜… 0 May 13
  • Problem: Existing camera-controlled video generation methods require large-scale camera-annotated data or expensive test-time optimization; no simple way to leverage pretrained video models' latent camera-control capability.
  • Model: "Warp-as-History": converts camera-induced geometric warps into camera-warped pseudo-history fed through pretrained video models' native history pathway, with target-frame positional alignment and visible-token selection.
  • Code: yyfz/Warp-as-History

PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation

↑ 7 πŸ“š 19915 β˜… 0 May 13
  • Problem: Existing video rewards fail to reliably score human motion realism because they operate in 2D pixel space without explicitly modeling 3D body physics and constraints.
  • Model: "PhyMotion": physics-grounded motion reward that recovers SMPL meshes, retargets to MuJoCo simulator, and evaluates motion via three axes (kinematic plausibility, contact/balance, dynamic feasibility)
  • Code: not released

Spatial Single-Cell Study

DUET: Dual-Paradigm Adaptive Expert Triage with Single-cell Inductive Prior for Spatial Transcriptomics Prediction

↑ 0 πŸ“š 5699 β˜… 0 May 13
  • Problem: Existing methods for inferring spatial gene expression from histology images oversimplify morphology-to-expression mapping and underutilize large-scale single-cell data as biological constraints.
  • Model: DUET: dual-paradigm framework synergizing parametric regression and memory-based retrieval with cellular inductive priors and adaptive expert triage for spatial transcriptomics prediction.
  • Code: Junchao-Zhu/DUET

StateXDiff: Cell State-Contextualized Multimodal Diffusion for Single-Cell Perturbation Prediction

↑ 0 β˜… 0 May 15
  • Problem: Predicting drug-induced cellular state changes at single-cell resolution under out-of-distribution conditions with limited multimodal information.
  • Model: StateXDiff: cell State-contextualized multimodal Diffusion framework integrating transcriptomic and pseudo-protein representations with mechanism-aware drug templates via latent conditional diffusion.
  • Code: not released