Dachuan Song

AI Research · Intelligent Agents · Long-Sequence Modeling · Health AI

Dachuan Song

I work across AI, intelligent agents, memory mechanisms, long-sequence modeling, and health AI, with an emphasis on clear research questions, reproducible experiments, and rigorous empirical evaluation.

Email GitHub Scholar LinkedIn

Direction

AI, Agents, and Sequence Modeling.

I am a Ph.D. student at George Mason University, advised by Prof. Xuan Wang. My current interests include intelligent agents, memory mechanisms, long-context reasoning, state-space and sequence modeling.

I focus on research problems where modeling ideas, implementation details, and empirical evaluation need to support each other.

Publications

Published papers

Research Areas

Intelligent Agents

Memory, tool-use reliability, long-context reasoning, and evaluation methods for models that need to plan, call tools, and avoid avoidable errors.

Long-Sequence Modeling

State-space models, spectral sequence modeling, budgeted inference, Mamba-style ideas, and mechanisms for keeping useful information over long inputs.

AI for Health

Causal and temporal modeling for fMRI, EEG, and other health data, with emphasis on interpretable signals and robust evaluation.

Skills

Technical toolkit.

Machine Learning

  • Linear and logistic regression
  • SVM, trees, and ensembles
  • PCA and k-means
  • Cross-validation and regularization

Deep Learning

  • CNNs and residual blocks
  • RNN, LSTM, and attention
  • Transformers and GNNs
  • Contrastive and triplet losses

Sequence & Control

  • State-space models
  • Kalman filtering and smoothing
  • AR and VAR modeling
  • LQR and MPC concepts

Agents & Evaluation

  • Tool-use workflows
  • Long-context memory
  • Error analysis
  • Benchmark design

Research Engineering

  • PyTorch experiments
  • CUDA-aware workflows
  • Reproducible ablations
  • Remote GPU execution

Programming & Tools

  • Python
  • JavaScript and TypeScript
  • Swift and iOS prototypes
  • Git, Linux, and data pipelines

Selected Work

Research projects with measurable technical claims.

2026 · Preprint

Elastic Spectral State Space Models for Budgeted Inference

Developed ES-SSM to train once at full spectral capacity and deploy elastically by activating ordered spectral channels through masked normalization and input-adaptive spectral gating.

PyTorch · CUDA · FFT convolution · Long-context benchmarks

2024–2025 · Publication line

Causal Fingerprinting from fMRI Time-Series Data

Reconstructed brain causal dynamics for subject and task fingerprints, connecting causal time-series modeling with interpretable neural signatures.

fMRI · Causal dynamics · Time-series modeling · Health AI

2023 · Master project

Machine Learning for EEG Brain Activity Patterns

Built a deep learning pipeline combining CNN spatial extraction, LSTM temporal modeling, attention, autoencoder pretraining, and transfer learning.

CNN · LSTM · Attention · Autoencoder · ResNet50

Open Code

Open research code.

ES-SSM

Budget-aware state-space modeling with elastic spectral capacity for efficient inference.

Research code

HealthPlanner

Local-only iOS health and training planner using personal baselines, recovery state, and workout history.

Project page

StoryVocab

English learning product concept that turns vocabulary into story-based daily practice.

Project page

Education & Service

Academic Background.

George Mason University

Ph.D. in Electrical and Computer Engineering · 2024–Present

Linear systems, random processes, Kalman filtering, distributed control and optimization, system identification.

University of Southampton

Master of Science (MSc)

Advanced machine learning, deep learning, computer vision, Bayesian learning, active and reinforcement learning.

Academic Service

ICML 2026 Silver Reviewer

Recognized for review quality evaluated by Area Chairs.