Shanghai Jiao Tong University

Efficient inference at EPIC Lab

I was an intern at EPIC Lab at Shanghai Jiao Tong University from August 2025 to April 2026. My research focused on diffusion LLM inference and KV cache eviction for audio language models.

Diffusion language models

Related research includes in-place prompting and self speculative decoding for diffusion language models. In-place prompting embeds reasoning in masked positions and uses confidence-aware early exit. Self speculative decoding combines self-drafting with hierarchical verification to accept multiple tokens in an iteration.

Audio language models

AudioKV studies KV cache eviction in large audio language models, using semantic-acoustic alignment and Spectral Score Smoothing to preserve audio-critical cache entries. I am the first author of this ACM MM 2026 paper.

Another related work, AudioMarathon, provides a benchmark for long-context audio understanding and efficiency across audio language models.