Weiyang Liu
Assistant Professor
Leads SphereLab, which studies principles of generative AI, foundation models and large language models.
Weiyang Liu's lab at CUHK works on generative & foundation models and large language models. 50 papers since 2023, 30 at top venues, mostly NeurIPS and ICLR. 16 PhD students. 1 alumnus. Applications close .
At a glance
- papers, 2023–2026
- 13/ year
- at T1 venues
- 60%
- PhD students
- 16
- MS students
- 0
- alumni listed
- 1
Papers by yearfrom the lab homepage, 2023–2026
- 16 papers2026 · T1 9 · collecting
- 14 papers2025 · T1 8
- 7 papers2024 · T1 3
- 13 papers2023 · T1 10
T1 venuesOtherVenues: NeurIPS, ICLR, CVPR, ICML, AISTATS
Research focuspaper titles that use each topic's words
- Large language models18 papers
- Generative & foundation models13 papers
- principles0 papers
- foundation models0 papers
- foundation0 papers
Paper list
From the lab homepage.
202616 papers
- Findings of EMNLPBeyond the Sampled Token: Preserving Candidate Support in RLVR (opens in a new tab)
- CVPR (T1)CHIPS: Efficient CLIP Adaptation via Curvature-aware Hybrid Influence-based Data Selection (opens in a new tab)
- EMNLP OralCompositional Machine Design as Program Synthesis with LLMs (opens in a new tab)
- NeurIPS (T1)Denoise First, Orthogonalize Later: Understanding Momentum in Muon via Spectral Filtering (opens in a new tab)
- Findings of EMNLPFormalMATH: Benchmarking Formal Mathematical Reasoning of Large Language Models (opens in a new tab)
- UAI (T1)On the Effect of Sampling Diversity in Scaling LLM Inference (opens in a new tab)
- Open-source ProjectOrbit: Stable and Efficient Reinforcement Learning for Trillion-Parameter LLMs (opens in a new tab)
- ICML (T1)Orthogonal Model Merging (opens in a new tab)
- Findings of EMNLP WorkshopPEFT-Arena: Understanding Parameter-Efficient Finetuning from a Stability-Plasticity Perspective (opens in a new tab)
- NeurIPS (T1)Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation (opens in a new tab)
- ICML OralPOET-X: Memory-efficient LLM Training by Scaling Orthogonal Transformation (opens in a new tab)
- ICLR (T1)Rigidity-Aware Geometric Pretraining for Protein Design and Conformational Ensembles (opens in a new tab)
202514 papers
- ICLR SpotlightCan Large Language Models Understand Symbolic Graphics Programs? (opens in a new tab)
- CVPR (T1)ChatHuman: Chatting about 3D Humans with Tools (opens in a new tab)
- ICLR (T1)Efficient Diversity-Preserving Diffusion Alignment via Gradient-Informed GFlowNets (opens in a new tab)
- TMLR ICLR workshop on World ModelsGenerating Symbolic World Models via Test-time Scaling of Large Language Models (opens in a new tab)
- Harvard Data Science ReviewLarge Language Models Are Zero-Shot Problem Solvers—Just Like Modern Computers (opens in a new tab)
- Proceedings of the National Academy of Sciences ICML workshop on Geometry-grounded Representation Learning and Generative ModelingManifold-Constrained Nucleus-Level Denoising Diffusion Model for Structure-Based Drug Design (opens in a new tab)
- EMNLP (T1)Orthogonal Finetuning Made Scalable (opens in a new tab)
- NeurIPS (T1)Reparameterized LLM Training via Orthogonal Equivalence Transformation (opens in a new tab)
- ICML (T1)SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator (opens in a new tab)
- NeurIPS (T1)Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation (opens in a new tab)
- NeurIPS NeurIPS Workshop on Structured Probabilistic Inference & Generative ModelingValue Gradient Guidance for Flow Matching Alignment (opens in a new tab)
- CVPR (T1)VERA: Explainable Video Anomaly Detection via Verbalized Learning of Vision-Language Models (opens in a new tab)
20247 papers
- NeurIPS (T1)Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision (opens in a new tab)
- ICLR OralGhost on the Shell: An Expressive Representation of General 3D Shapes (opens in a new tab)
- CVPR (T1)GraphDreamer: Compositional 3D Scene Synthesis from Scene Graphs (opens in a new tab)
- Findings of EMNLP WorkshopIn Defense of Structural Sparse Adapters for Concurrent LLM Serving (opens in a new tab)
- ICLR SpotlightMetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models (opens in a new tab)
- ICLR (T1)Parameter-Efficient Orthogonal Finetuning via Butterfly Factorization (opens in a new tab)
- TMLR Selected for presentation at ICLRRe-Thinking Inverse Graphics with Large Language Models (opens in a new tab)
202313 papers
- TMLR (T1)Continual Learning by Modeling Intra-Class Variation (opens in a new tab)
- NeurIPS (T1)Controlling Text-to-Image Diffusion by Orthogonal Finetuning (opens in a new tab)
- TPAMI (T1)Data-Efficient Learning via Minimizing Hyperspherical Energy (opens in a new tab)
- ICLR (T1)Generalizing and Decoupling Neural Collapse via Hyperspherical Uniformity Gap (opens in a new tab)
- UAI Oral & Best HCOMP Poster AwardHuman-in-the-Loop Mixup (opens in a new tab)
- TKDE (T1)Iterative Graph Self-Distillation (opens in a new tab)
- AISTATS (T1)Iterative Teaching by Data Hallucination (opens in a new tab)
- ICLR Notable Top-25% ICCV workshop on AI for 3D Content CreationMeshDiffusion: Score-based Generative 3D Mesh Modeling (opens in a new tab)
- ICML (T1)Nonparametric Iterative Machine Teaching (opens in a new tab)
- NeurIPS (T1)Nonparametric Teaching for Multiple Learners (opens in a new tab)
- ICCV (T1)One-shot Implicit Animatable Avatars with Model-based Priors (opens in a new tab)
- ICCV (T1)Pairwise Similarity Learning is SimPLE (opens in a new tab)
Where alumni went1 listed, 0% to industry
Industry 0Academia 1
Also to
- TUM
Apply to CUHKFall 2027
Computer Science and Engineering MPhil/PhD
- Round
- HKPFS / main round
- Opens
- —
- GRE
- See program page
- English
- Interview in English
- Fee
- See program page
Not yet confirmed for this intake.
Similar labsnearest by research summary
- GBCambridge
Ferenc Huszár
7 T1 - CHETH Zurich
Martin Vechev
- Trustworthy & safe ML
- USCMU
Steven Wu
- Trustworthy & safe ML
- USUSC
Bart Andrew Kosko
0 T1 - USMichigan
Atul Prakash
- Trustworthy & safe ML
- SGNUS
Jiancong Xiao
- Trustworthy & safe ML
Your notes
Save this lab to keep notes.
Data
Checked against the lab homepage.
- Collected
- Verified
- Papers
- The lab homepage
- Members
- Lab homepage
- Alumni
- Lab homepage
Tiers are venue classifications, not lab rankings.
CUHKcloses