Deep Learning · State Space Models · Long Contexts · LLM Agents

Dachuan Song 宋大川

I study how learning systems can process long sequences efficiently and use accumulated information reliably in complex tasks.

GitHub Scholar LinkedIn

Direction

Deep Learning,
Sequence Modeling.

I am a Ph.D. student at George Mason University, advised by Prof. Xuan Wang.
I study learning systems that must process long and evolving information under practical constraints on computation, memory, and latency.

My research focuses on state space models, long-sequence modeling, and LLM-based agents, with an emphasis on efficient and reliable learning systems.

Publications

Published papers

Preprint arXiv · 2026

Mergeable Model-Side Aggregation States for Long-Context Language Models

Dachuan Song, Junyu Yin, Zechen Hu, Xuan Wang

Introduced SketchOps, a model-side aggregation interface that keeps compact, mergeable sketch states alongside frozen language models for fixed-budget long-context aggregation.

arXiv:2607.26448 Code

SketchOps · HyperLogLog · Long-context reasoning · Fixed-budget aggregation

SketchOps workflow and main experimental results

Preprint arXiv · 2026

Elastic Spectral State Space Models for Budgeted Inference

Dachuan Song, Xuan Wang

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

arXiv:2601.22488 Code

PyTorch · CUDA · FFT convolution · Long-context benchmarks

Journal Article Health Information Science and Systems · 2025

Reconstructing brain causal dynamics for subject and task fingerprints using fMRI time-series data

Dachuan Song, Li Shen, Duy Duong-Tran, Xuan Wang

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

DOI: 10.1007/s13755-025-00388-w

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

Journal Figures 1 and 3: two-timescale sampling and five-layer GNN architecture

Conference Paper ACM BCB · 2024 Oral Presentation

Causality-based Subject and Task Fingerprints using fMRI Time-series Data

Dachuan Song, Li Shen, Duy Duong-Tran, Xuan Wang

Studied subject and task fingerprints from fMRI time-series data through causal modeling, with an oral presentation at ACM BCB 2024.

DOI: 10.1145/3698587.3701342

ACM BCB · Causal modeling · fMRI · Health AI

Conference Figures 1 and 2: causal dynamics reconstruction and five-layer GNN

Internship Optosurgical · Summer 2026

Surgical Video Understanding and Efficient Depth Estimation (Machine Learning Engineer Intern)

Built video models to recognize surgical phases and instrument actions for robot-assisted procedures. Compressed and fine-tuned Depth Anything 3 to 29% of its original size while retaining nearly the full model's performance.

Surgical phase recognition · Robot-assisted surgery · Model compression · Monocular depth estimation

Combined non-graphic surgical grasper and clip-applier illustration beside a masked small-model depth prediction

Research Areas

Efficient Long-Sequence Modeling

Efficient state-space and hybrid sequence architectures that use attention selectively and adapt computation to each input, reducing inference cost while preserving model quality.

Spectral State Space Models

State-space sequence models with spectral structure and elastic capacity for efficient, budget-aware inference.

Reliable LLM Agents

Agent memory, provenance tracking, and tool-use mechanisms designed to reduce stale information use and improve workflow reliability.

Core technical areas.

Deep Learning

Transformers and attention
Foundation model adaptation
Optimization and representation learning

Long-Sequence Modeling

Hybrid sequence architectures
Length generalization
Budget-aware inference

State Space Models

Spectral state-space models
Linear dynamical systems
Efficient sequence operators

LLM Agents

Tool-using agents
Agent memory
Workflow reliability

Education

Academic Background.

Ph.D.: George Mason University Electrical and Computer Engineering
M.Sc.: University of Southampton Computer Science
B.Eng.: Xinjiang University Software Engineering

ICML 2026 Silver Reviewer

Recognized for review quality evaluated by Area Chairs.