State Space Models · Long Contexts · LLM Agents · Efficient Inference
I build sequence models and compact memory for long contexts, so useful information remains accessible without making inference prohibitively expensive.
Seeking Summer 2027 research internships.
Direction
I am a Ph.D. student at
George Mason University, advised by
Prof. Xuan Wang.
The central question in my work is what a model should carry forward from a
long history, and how efficiently it can use that state.
I develop spectral state-space models that adapt what they retain and can be deployed at different compute budgets. I also build small, mergeable states that let language models combine information from many chunks without rereading the full history.
Publications
Dachuan Song, Junyu Yin, Zechen Hu, Xuan Wang
One ES-SSM training run yields standalone models at several compute budgets. Ordered Hankel channels can be truncated directly; adaptive gates and budget dropout train the retained prefixes to remain predictive.
Spectral state space models · Hankel channels · Budgeted inference
Dachuan Song, Chuchu Chen, Xuan Wang
ESSH makes spectral channels selective: input-dependent recurrent updates control what each channel retains. Paired with sliding-window attention, one training run yields compact language models at several sizes with fast decoding.
Selective spectral recurrence · Sliding-window attention · Elastic inference
Dachuan Song, Junyu Yin, Zechen Hu, Xuan Wang
Language models struggle to aggregate counts and set relations over long histories. SketchOps adds a fixed 2 KiB state that merges across chunks; on 3,969 aggregate-then-reason tasks with Gemma 4 (31B), it reached 99.2% accuracy without rereading the full context.
SketchOps · HyperLogLog · Long-context reasoning · Fixed-budget aggregation
Dachuan Song, Li Shen, Duy Duong-Tran, Xuan Wang
A two-timescale state-space model extracts directional brain interactions from fMRI. Modal features identify subjects, while a graph neural network classifies tasks.
ACM BCB · Causal modeling · fMRI · Health AI
Dachuan Song, Li Shen, Duy Duong-Tran, Xuan Wang
Reconstructed directed interactions and fast and slow activity from fMRI, then used those causal signatures to identify subjects and tasks. A brain reachability map visualizes task-specific regional activation.
fMRI · Causal dynamics · Time-series modeling · Health AI
Developed video models that recognize surgical phases and instrument actions, providing procedure-state cues to downstream systems. Fine-tuned a compressed Depth Anything 3 model at 29% of the original size while retaining near-full depth-estimation performance.
Surgical phase recognition · Robot-assisted surgery · Model compression · Monocular depth estimation
Research Areas
Combine input-dependent recurrent memory with local attention to process long sequences at manageable inference cost.
Use ordered spectral channels as a capacity axis: truncate a trained model's channel prefix to export smaller, standalone versions.
Study how agent memory and provenance can keep tool-using systems grounded in current evidence instead of stale context.
ICML 2026 Silver Reviewer
Recognized for review quality evaluated by Area Chairs.
Reviewer for NeurIPS 2026 and AAAI 2027.