Hi, I’m Yu Mao. Feel free to call me Audrey! I’m a Research Scientist at ByteDance, working on AI infrastructure and LLMs. My research interests lie at the intersection of machine learning and computer systems, with a focus on model capabilities, efficient training, and inference. I received my Ph.D. in Computer Science from City University of Hong Kong, where I worked on compression and computer systems.
Research interests
Inference Systems Compression & Representation LLM Capability & Agent Reliability
🔬 Current Work
RL as a Probe of LLM Capability
Current accuracy does not tell us how a model will respond to further training. My first-author paper uses random-reward RL to probe that response without correctness information in the reward.
Controlled comparisons show that the learner’s starting state changes what the same training signal can unlock. The paper and code are available; a reusable capability evaluation framework is in development.
Random-reward RL probes model capabilities by measuring how models respond to further training.
Best terminal GSM accuracy across seeds (correctness reward: 4; random reward: 8).
Omni Profiler: Continuous Profiling for RL Training
High utilization or growing memory alone does not identify the responsible code. I co-developed Omni Profiler to connect CPU/GPU memory and compute to Python and native stacks.
Seed RL uses the profiler to investigate training bottlenecks and memory growth, connecting resource measurements to code that engineers can inspect.
Resource-attribution schematic. CPU time, GPU duration and live memory are distinct measurements; this is not a measured performance result.
AIBrix Hybrid KV Cache
Hybrid-model prefixes need recurrent state and attention history to survive offloading together. I built and open-sourced hybrid KV offloading with bounded storage and reader-safe eviction, now used by TikTok Infra.
The open draft AIBrix PR proposes the byte-store contract and has not been merged. It excludes the hybrid serving connector and codecs. The public CPU quickstart verifies byte-store round trips; whole-model output equivalence remains open.
Storage mechanism: encode hybrid state and KV spans, then restore and check the bytes. This does not establish whole-model output equivalence.
More research
KVBytePack: Byte-Exact KV-State Compression
A reproducible artifact for measuring when smaller KV records pay for their codec cost, using public records and complete-record CPU and memory measurements. This is a separate artifact from the hybrid serving integration and the earlier KV cache benchmark.
Bitwise-Reproducible LLM Inference Across Hardware
Historical runs recorded matching logits and tokens across selected GPU and CPU configurations, including int8 vLLM batching and cache experiments. The current source adds arithmetic guards and CPU checks; these do not revalidate the historical whole-model or CUDA results.
Agent Reliability & Semantic Profiling
I test whether agents actually execute the actions they claim. I also coauthored AgentPProf for semantic resource profiling, winner of the AgenticOS @ SOSP 2026 Best Paper Award.
MY LITTLE TOOL ✦
A free, lightweight speech-to-text tool for Apple Silicon Macs. Build from source.
🔥 News
- 2026.10: Released an experimental source preview of AIBrix Hybrid KV Cache.
- 2026.10: Our paper, Probe with Participation Trophies: Random-Reward RL as a Probe of LLM Capability, is now available on arXiv.
- 2026.09: Our paper, AgentPProf: Semantic Profiler for AI Agents, received the Best Paper Award at AgenticOS @ SOSP 2026.
- 2026.01: BAHOP and ECM accepted at ICASSP 2026.
Earlier news
- 2025.12: Hitcher accepted at KDD 2026, Parallel-SA at DATE 2026.
- 2025.05: WISE accepted at CVPR 2025, PMR at USENIX ATC 2025, Easz and DAWN at DAC 2025.
📝 Selected Publications
🎖 Honors and Awards
- 2026.09 Best Paper Award, AgenticOS @ SOSP 2026: AgentPProf: Semantic Profiler for AI Agents
- 2025.11 DAAD Interpretable AI Postdoctoral Fellowship
- 2024.12 NeurIPS Outstanding Reviewer
- 2023.10 TinyML Contest ICCAD 2023, Second Place, San Francisco, USA
- 2023.09 Outstanding Academic Performance Award, City University of Hong Kong
- 2023.07 DAC Young Research Fellow, San Francisco, USA
- 2022.10 EDAthon 2022, Second Place, Hong Kong
🤝 Services
- Conference reviewing: NeurIPS 2024–2026; ICLR 2025–2026; ICML 2025; CVPR 2024–2026; ICCV 2025; BMVC 2026; ACM MM 2023–2024; ARR
- Program committees: AAAI 2026; DAC 2026; GLSVLSI 2025–2026; RTCSA 2025
- External review committee: USENIX ATC 2024
- Artifact evaluation committees: EuroSys 2025; DSN 2024
- Journal reviewing: ACM TECS; TMLR; IEEE TKDE
👩🏫 Supervision
- Shashwat Jaiswal, summer intern (Ph.D. student at UIUC), 2026
- Yusheng Zheng, summer intern (Ph.D. student at UCSC), 2026
- Jun Wang, Ph.D. student with Prof. Jason Xue (2024–2025)
- Weilan Wang, Ph.D. student with Prof. Jason Xue (2024–2025)
- Dongdong Tang, Ph.D. student with Prof. Jason Xue (2024–2025)