Selected work
Research
I work on efficient inference for diffusion and audio language models, as well as robotic ultrasound. Selected projects are grouped below by topic.
Sampling, prompting, and verification methods for faster generation.
Long-context memory and practical KV-cache management.
Navigation and probe guidance for medical robotics.
Diffusion language models
Diffusion LLM inference
Work on sampling, prompting, and self-speculative decoding for faster diffusion language-model inference.
SlowFast Sampling
Switches between exploratory and accelerated stages using token certainty, convergence, and position; 15.63× speedup on LLaDA.
Thinking inside the Mask
Integrates intermediate reasoning into the mask-denoising process through in-place prompting.
Self-Speculative Decoding
Uses self-drafting and parallel verification to accelerate diffusion language-model inference.
Audio and multimodal LLMs
Audio LLM memory
Work on preserving important acoustic and linguistic context while reducing memory use.
AudioKV
Uses semantic-acoustic alignment and Spectral Score Smoothing to retain important KV-cache entries; 60% lower memory overhead.
AudioMarathon
A benchmark for long-context audio understanding and efficiency.
Embodied medical intelligence
Robotic ultrasound
Work on autonomous navigation and real-time probe guidance for ultrasonography.
UltraHiT
A hierarchical Transformer for generalizable internal-carotid-artery robotic ultrasonography.
VA-Adapter
A parameter-efficient vision-action adapter for real-time echocardiography probe guidance.