Viet-Hoang Tran

I'm a final-year Ph.D. student in the Department of Mathematics at the National University of Singapore (expected to graduate this October). I began my Ph.D. in August 2022 in Mathematics. In September 2023, I shifted my research focus to Machine Learning and started working under the supervision of Professor Tan Minh Nguyen. I completed an Honors degree in Mathematics at the Hanoi University of Science (2018-2022). email: hoang.tranviet [at] u.nus.edu | Google Scholar | CV

Research Interests

• I study symmetries in neural networks, which boils down to one question: Given a model (f), if two parameter settings (theta) and (bar{theta}) represent the same function, i.e., (f_{theta} equiv f_{bar{theta}}), how are (theta) and (bar{theta}) related? These symmetries have all sorts of applications, including understanding implicit bias through neural conservation laws, symmetry-aware optimization and quantization, equivariant metanetworks, mode connectivity up to symmetry, and more.

• Recently, I've also become interested in using AI to tackle math problems, mostly related to commutative algebra and algebraic geometry. Partly for fun and curiosity, but more importantly, as a way to better understand what LLMs can and can't do in mathematical reasoning. I share my thoughts and progress here.

News

• Aug 2026. Amid the Math-AI hype, with help from models available through ChatGPT and Claude Code and an agentic harness we built around them, I started exploring a few maths problems and got some decent results. The goal is to better understand what LLMs can and can't do in mathematical reasoning, rather than simply to get the problems solved. I share my progress on this project page.

• May 2026. 3 papers accepted at ICML 2026:

  • Conservation Laws for Modern Neural Architectures (spotlight)

  • Functional Equivalence in Attention: A Comprehensive Study with Applications to Linear Mode Connectivity

  • Geometric and Stochastic Analysis of Discontinuities in Sparse Mixture-of-Experts (spotlight)

• Jan 2026. 4 papers accepted at ICLR 2026:

  • Quasi-Equivariant Metanetworks

  • Mixed-Curvature Tree-Sliced Wasserstein Distance

  • Tree-sliced Sobolev IPM

  • Revisiting Tree-Sliced Wasserstein Distance Through the Lens of the Fermat-Weber Problem

> older news

• Oct 2025. 1 paper accepted at AAAI 2026 Workshop MATH4AI:

  • Modeling Expert Interactions in Sparse Mixture of Experts via Graph Structures

Sep 2025. 3 papers accepted at NeurIPS 2025:

• On Linear Mode Connectivity of Mixture-of-Experts Architectures (oral)

• Tree-Sliced Entropy Partial Transport

• Dynamical Properties of Tokens in Self-Attention and Effects of Positional Encoding

May 2025. 3 papers accepted at ICML 2025:

• Equivariant Polynomial Functional Networks

• Tree-Sliced Wasserstein Distance: A Geometric Perspective

• Tree-Sliced Wasserstein Distance with Nonlinear Projection

Feb 2025. 1 paper accepted at ICLR 2025 Workshop Weight Space Learning:

• Equivariant Neural Functional Networks for Transformers (spotlight)

Jan 2025. 3 papers accepted at ICLR 2025:

• Equivariant Neural Functional Networks for Transformers

• Distance-Based Tree-Sliced Wasserstein Distance

• Spherical Tree-Sliced Wasserstein Distance

Sep 2024. 1 paper accepted at NeurIPS 2024:

• Monomial Matrix Group Equivariant Neural Functional Networks

Selected Publications

• Conservation Laws for Modern Neural Architectures Viet-Hoang Tran¹, Vinh Khanh Bui¹, Tan Lai Ngoc, Nam Nguyen, Tuan Dam, Tan M. Nguyen² ICML 2026 (spotlight)

• On Linear Mode Connectivity of Mixture-of-Experts Architectures Viet-Hoang Tran¹, Van Hoan Trinh¹, Khanh Vinh Bui¹, Tan M. Nguyen² NeurIPS 2025 (oral)

• Equivariant Neural Functional Networks for Transformers Viet-Hoang Tran¹, Thieu N. Vo¹, An Nguyen The¹, Tho Tran Huu, Minh-Khoi Nguyen-Nhat, Thanh Tran, Duy-Tung Pham, Tan Minh Nguyen² ICLR 2025

• Distance-Based Tree-Sliced Wasserstein Distance Viet-Hoang Tran¹, Khoi N.M. Nguyen¹, Trang Pham, Thanh T. Chu, Tam Le², Tan M. Nguyen² ICLR 2025

> all publications