Jiawei Li (David Lee)
Independent researcher and AI product builder. I study how fast decision models, tool use, and agent harnesses behave under real reliability, robustness, and cost constraints, most recently in Fast Models, Slow Evidence (arXiv, 2026). Earlier: co-author of the KDD 2018 Research Track Best Student Paper and co-inventor on 7 patent families spanning generative music, VR live systems, spatial interaction, and immersive commerce.
Current research interests
Interests, not a publication list: not every area has a paper yet.
Publications
Fast Models, Slow Evidence: A Paired and Self-Audited Evaluation of System-1 Decision Models for LLM Agent Harnesses
Agent harnesses make many small, typed decisions per task: which model to call, which tool to use, whether retrieved text is relevant, whether an input is a prompt injection. Can fast, single-forward-pass System-1 decision models make or gate those calls reliably, instead of a full LLM call each time?
Two System-1 channels, Laya (open-weight) and Jev (hosted), compared on byte-identical paired inputs with paired statistical tests, cross-hardware and cross-day reproducibility checks, and an explicit self-audit of the evaluation pipeline.
- Jev was significantly more accurate on 9 of 11 decision points, by +10.8 to +46.0 percentage points.
- Neither model beat chance on zero-shot model routing; the two tied on RAG relevance gating.
- Laya changed about 30% of its answers when option order was reversed, and on items with one correct tool it reached 31% accuracy against 50 nearest-neighbour candidates, versus 98% for Jev.
- The self-audit overturned attractive deployment claims: counting an omitted pre-screen cost cut a reported 23.9% saving to 4.3%, and thresholds tuned for 5% misses in-sample reached up to 17% on held-out data.
XiaoIce Band: A Melody and Arrangement Generation Framework for Pop Music
Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 2837–2846. London, United Kingdom.
End-to-end pop music generation that handles melody and multi-track arrangement together: CRMCG generates melody conditioned on chord progression and rhythm, and MICA arranges multiple instruments with multi-task learning.
Microsoft built XiaoIce's machine-composition feature on this model and demonstrated it on CCTV's 机智过人, per USTC News.
Per the paper, completed while the first and fifth authors (Jiawei Li is fifth) were interns at Microsoft.
Patents & inventions
Named co-inventor on seven distinct invention families. Filings of one invention in different countries count once.
Generative music
Microsoft · filed 2019Song generation from a text input
priority 2019-06-14Official title: Song generation based on a text input
Extracts topic and emotion from a text input, determines a melody conditioned on them, writes lyrics from the melody and the text, and produces a song.
Inventors: Guang Zhou, Wei Wang, Jiawei Li, Peijun Xia, Xianchao Wu, Lu Yang, Yuanchun Xu
VR live systems
ByteDance · filed 2022–2023Key-point-driven effect rendering for VR livestreams
priority 2022-09-06Official title: Live special effect rendering method, device, equipment, readable storage medium and product
Locates target key points in ultra-high-resolution VR livestream frames and concentrates effect rendering on the related regions, cutting computation and latency.
Inventors: Zhang Yi, Jiawei Li, Chen Siying
Region-based processing of 8K VR livestream frames
priority 2022-09-06Official title: Live image frame processing method, device, equipment, readable storage medium and product
Selects only the regions of a binocular, ultra-high-resolution VR frame that need processing, such as beautification, so 8K livestreams stay real-time.
Inventors: Chen Yongzhan, Jiang Linjun, Luo Jiajia, Jiawei Li, Chen Siying
Cloud / headset split rendering for VR livestream effects
priority 2022-11-17Official title: Live video stream rendering method, device, equipment, storage medium and product
Chooses where each effect runs based on how real-time it must be: rendered entirely in the cloud, or detected in the cloud and rendered on the VR headset.
Inventors: Wang Can, You Dong, Jiawei Li
Distortion-aware beautification for binocular VR livestreams
priority 2023-03-07Official title: Image processing method and apparatus, and device, computer-readable storage medium and product
Undistorts the target region of a binocular VR frame, applies beautification or filters, then maps the result back into the original projection, improving quality while reducing computation.
Inventors: Jiang Linjun, Deng Qili, Jiawei Li
Spatial interaction & commerce
ByteDance · filed 2023Perspective windows across virtual scenes
priority 2023-02-15Official title: Video processing method, device, electronic equipment and storage medium
Opens a portal-like window inside the current VR scene that shows a view into another pre-built scene, computed from the viewer's observation point.
Inventors: Liang Xinglun, Jiawei Li, Zhang Fan
Compositing real objects with virtual goods in VR / AR
priority 2023-02-23Official title: Method and apparatus for interaction in virtual environment, device and storage medium
Shows a real object, such as the user's foot, alongside virtual goods; the selected item is composited onto it, as in a virtual try-on, and can then be associated with the user.
Inventors: Chao Zhou, Dong You, Jiawei Li, Xinglun Liang, Fan Zhang, Yepeng Chen, Zhiwei Su, Lei Gao
Status as checked on 2026-09-30. Legal status changes; the official registers (USPTO, CNIPA, EPO, WIPO) are authoritative. Being a named inventor does not mean owning the patent: the assignees are Microsoft and ByteDance.
Where the research went
The research and the products share their problems: generative music at XiaoIce, VR live systems at PICO, and agent evaluation in the agents I ship today.