I am a senior undergraduate in Computer Science at Shanghai Jiao Tong University (ACM Honors Class, Zhiyuan College). I am currently a research intern at HuLab, University of Wisconsin–Madison, advised by Prof. Junjie Hu. At SJTU, I work in the Advanced Network Laboratory, advised by Prof. Shengzhong Liu and Prof. Fan Wu. Previously, I was a research intern at Audiocc, advised by Prof. Chenda Li and Prof. Yanmin Qian. Friends call me Hugo.
My research centers on test-time computation and efficient reasoning in large language models: RL-based post-training, inference-time reasoning control, and test-time training and adaptation. More broadly, I want to understand the principles of computation and inference that lead to efficient, generalizable intelligence across modalities.
I am always open to collaboration and academic discussion. Feel free to reach me at hugo0713@sjtu.edu.cn or qhu99@wisc.edu, or on WeChat (ID: HUGO--2025).
I am applying to PhD programs for Fall 2027.
Shares test-time-training fast weights across the whole depth stack instead of keeping one private state per layer, which turns the model into a single recurrence over (chunk, layer) and improves long-context retrieval at equal state size.
Reads a fast-weight memory by generating from it: a key-conditioned denoising readout samples a value instead of regressing to the conditional mean, so a compressed memory can recall competing values instead of their average, improving long-context retrieval and novel view synthesis at the same state size.
Lets every layer read only the last layer's KV cache of earlier chunks, turning the Transformer into a chunk recurrence across depth and time: long-term KV storage shrinks toward 1/L while RULER retrieval stays the best among the compared methods at 124M and 760M.
A sparse Mixture-of-Experts framework that routes experts over time frames and mel bands, raising the capacity of speech-separation models with almost no extra inference cost (+3.8 dB SDR over BSRNN on Libri2Mix at a comparable 4.1 GMACs/s).
GRPO-based efficient reasoning with progressive chain-of-thought length calibration: it estimates the optimal reasoning length for each question and adapts the length-reward coefficient, cutting reasoning tokens by up to 52% while improving accuracy on benchmarks such as AIME25.
Three projects on long-context memory, all under review at ICLR 2027. Universal Test-Time Training (co-first author): shared fast-weight memory across time and depth—one state pool for the whole stack instead of a private state per layer; at matched state and active compute, uTTT-MoE gains 2.6 / 2.1 RULER points over layer-private TTT-MoE at 124M / 760M. Test-Time Generation (second author): generate each memory read from fresh noise, guided by the query, so a compressed memory can recover alternative values instead of their average. LastKV: let every layer read a shared last-layer KV history, feeding deep past representations back into future computation and cutting long-term KV storage toward 1/L of a standard Transformer.
Reinforcement learning and post-training for efficient LLM reasoning, including chain-of-thought compression and on-policy distillation.
Co-built JEDI, an AI participant that joins Jitsi meetings and decides when to speak; owned its Kubernetes-native data layer (Ingress → PostgREST → PostgreSQL).
Making speech-separation models cheaper to run without shrinking their capacity, via sparse MoE routing over temporal frames and mel-frequency bands.
Course notes and research notes. English translations of the Chinese originals.