LLM post-training & RL
Reliable reinforcement learning, reasoning, and optimization for large language models.
Ph.D. Student · Computer Science · UIUC
Yifan Hao
LLM post-training,
grounded in theory.
I study how to make machine learning systems learn more efficiently and generalize more reliably. I’m a Ph.D. student at the University of Illinois Urbana-Champaign, advised by Prof. Tong Zhang, working at the intersection of machine learning and statistics.
01Research interests
Efficient learning
Algorithms that improve sample efficiency and reduce the computational cost of adaptation.
Statistical foundations
Generalization, robustness, and the theory behind modern machine learning systems.
02Recent work
Understanding how RLVR learns
A practical Muon training recipe, together with spectral diagnostics and checkpoint interventions to investigate optimization and learned behavior.
Muon for RLVR · PreprintMore stable off-policy learning
Adaptive Layerwise Perturbation unifies off-policy corrections for LLM RL, addressing policy staleness and training–inference mismatch.
Read the ALP paper03Research experience
Full background ↗May – Aug 2026
Applied Scientist Intern · Amazon Rufus
Hosts: Zoey Li and Zhenyu Shi
- Developed a Muon recipe for RLVR with no momentum, achieving 29.17% macro-averaged Pass@1 on Qwen3-4B-Base compared with 24.47% for AdamW across three math benchmarks.
- Used spectral diagnostics and checkpoint interventions to study optimization stability and how learned behavior is organized across singular directions and attention projections.
- Investigated RLVR trajectory predictability with random-walk controls. Read the research write-up ↗
Sep 2024 – present
Ph.D. Student · University of Illinois Urbana-Champaign
Computer Science · Advisor: Prof. Tong Zhang
- Developed Adaptive Layerwise Perturbation (ALP) to stabilize off-policy LLM reinforcement learning under policy staleness and training–inference mismatch.
- Proposed DIPNeT for generalization across reasoning and adversarial robustness settings.
- Developed GVM-RAFT, a prompt-specific sample allocation strategy, achieving a 2–4× speedup and improved accuracy over vanilla RAFT.
- Studied feature decoupling in Transformers and their expressiveness for hidden Markov processes, and explained how ensemble methods reduce overadaptation in supervised fine-tuning.
04Selected publications
All publications ↗Muon for RLVR: What Works, Why, and What It Learns
* indicates alphabetical author order or equal contribution.
Let’s talk research.
I welcome conversations about LLM post-training, reinforcement learning, AI agents, and full-time or internship opportunities.