About Me
Jian Chen
I am a Senior AI Researcher at Dolby Laboratories working on audio foundation models, multimodal large language models, and agentic systems. My research spans language, audio, and vision, covering multimodal generation and reasoning, post-training, compression, and distillation.
Research
Speech & Audio AI — 2025 – present
At Dolby Laboratories, my work centers on audio AI — audio foundation models and the agentic systems around them. The research directions closest to my heart include audio agentic systems: LLM-based assistants that interpret natural-language prompts and audio inputs, then orchestrate modular audio skills — source separation, spatial layout, trajectory design — for controllable, object-based spatial audio creation and editing; and efficient voice cloning: transcript-free, non-autoregressive approaches that combine masked generation with accelerated flow-matching rendering to deliver large end-to-end speedups with competitive speaker similarity and intelligibility across English and Chinese. I am also working on full-duplex speech-to-speech models for natural, real-time spoken dialogue.
Efficient LLMs & Compression — 2025 – present
This is a collaboration with friends outside of work: together we have been making LLMs cheaper to serve. Swift-SVD (ICML 2026) extends low-rank KV-cache compression with dynamic rank allocation across layers and shows that incremental-SVD is numerically more stable than classical SVD-based approaches, improving QA and perplexity under compression; KV-Core (NeurIPS 2025 Workshop) provides the data-aware benchmarking and the simple, theoretically optimal core method behind it.
Multimodal LLMs & Document Understanding — 2024 – 2025
A long-standing thread of mine is helping multimodal LLMs truly read. At Adobe I built SV-RAG (ICLR 2025), a framework that lets a 4B multimodal backbone achieve multi-page document understanding competitive with much larger proprietary models, and created VisR-Bench and MMR, benchmarks that exposed systematic weaknesses in structured table reasoning and low-resource languages. I have also contributed to LLaVA-Read and the LaRA/TRINS document-reading models, and explored cross-modal alignment between symbolic music and language with MusiXQA.
Spatial Planning — 2024
In this area I work on layout generation and planning: LACE (ICLR 2024), a continuous diffusion model that uses constrained optimization to unify multiple conditional inputs for layout generation and editing, and TextLap (EMNLP 2024), which customizes language models for text-to-layout planning and reached state-of-the-art on graphical design benchmarks.
Education
- Ph.D. in Computer Science, University at Buffalo, 2018 Fall – 2025 Spring
- M.S. in Electrical Engineering, Drexel University, 2015 Fall – 2017 Spring
- B.S. in Applied Mathematics, Hunan University, 2011 Fall – 2015 Spring
Experience
- Dolby Laboratories, Sound Experience Lab — Senior AI Researcher, 06/2025 – present
- Adobe, Document Intelligence Lab — Research Scientist Intern, 05/2024 – 11/2024
- MBZUAI — Research Assistant, 10/2023 – 02/2024
- University at Buffalo — Research Assistant, 09/2018 – 09/2025
Professional Service
I review for the major AI venues — NeurIPS, ICLR (Notable Reviewer 2025), ICML, AAAI, ACL, ICCV, CVPR, TMLR, IJCAI, ARR — as well as journals such as Genome Biology and TETCI.
Beyond Research
Tennis 🎾
After moving to Atlanta in 2025 I picked up tennis again, and it has grown into a serious pursuit: I take lessons, practice regularly, and work hard in matches with the goal of steadily raising my UTR (Universal Tennis Rating).
Music 🎸
Music has been with me since long before research: I play guitar and arrange music as a hobby, and I share my work on YouTube. In 2025 I completed Berklee Online’s songwriting certificate — writing music for the pure joy of it.