RESEARCH MAP · 研究脉络

Follow the questions behind the papers.

A publication list is chronological; research is not. This page groups my work by the recurring questions it tries to answer.

01 Start from the bottleneck

Identify where time, memory, or generalization is actually lost before designing the method.

02 Connect algorithm and system

Prefer improvements that remain visible in measurable end-to-end behavior.

03 Leave reusable artifacts

Pair ideas with code, datasets, benchmarks, or project pages whenever possible.

THREAD 01

DIFFUSION LANGUAGE MODELS

How can parallel generation become genuinely faster?

Diffusion LLMs expose parallelism, but useful acceleration depends on token confidence, denoising dynamics, verification cost, and the way reasoning is represented.

ICLR 2026 · Accepted · CCF-A

SlowFast Sampling

Alternates exploratory and accelerated stages using certainty, convergence, and positional principles; the homepage reports a 15.63× speedup on LLaDA.

ACL 2026 · Submitted

Thinking inside the Mask

Places reasoning directly into the mask-denoising process through in-place prompting.

ACL 2026 · Submitted

Self-Speculative Decoding

Uses internal self-drafting and parallel verification to accelerate diffusion language-model inference.

THREAD 02

AUDIO & MULTIMODAL LLMS

What information deserves to stay in memory?

Long audio contexts make memory allocation a semantic decision: compression must preserve acoustically and linguistically critical evidence rather than treating every token or head equally.

ACM MM 2026 · Accepted · First Author

AudioKV

Prioritizes audio-critical attention heads through semantic-acoustic alignment and Spectral Score Smoothing; the homepage reports 60% lower memory overhead.

ACL 2026 · Submitted

AudioMarathon

A benchmark for evaluating long-context audio understanding together with efficiency.

THREAD 03

EMBODIED MEDICAL INTELLIGENCE

How can models generalize from perception to physical guidance?

Robotic ultrasonography couples visual representation, anatomy-aware generalization, and real-time action guidance under strong safety and data constraints.

ICRA 2026 · Accepted · CCF-B · Equal Contribution

UltraHiT

A hierarchical Transformer architecture for generalizable internal-carotid-artery robotic ultrasonography.

MICCAI 2026 · Accepted · CCF-B

VA-Adapter

A parameter-efficient vision-action adapter for real-time echocardiography probe guidance.

Research has a route as well as a result.

Continue to the journey timeline