Selected work and experience
Publications and projects
Selected projects in algorithms, optimization, machine learning theory, Transformer foundations,
generative modeling, and decision-focused learning.
arXiv · Transformer theory and automata
Partial Answer of How Transformers Learn Automata
Tiantian Zhang
A theoretical project motivated by the open questions in Transformers Learn Shortcuts to
Automata. The work develops representation-theoretic constructions for simulating richer
finite automata with Transformers, using semidirect-product structure, group actions, and
Fourier modules to study when automaton dynamics admit efficient parallelizable realizations against KDA
Neurips Main· 2026 · generative modeling theory
Bridge Graphical Models: Markov Compression as a Design Principle for Continuous-Time Generative Modeling
Tiantian Zhang
A path-space graphical framework for continuous-time generative models, including diffusion models, flow matching, rectified flows, Schrödinger bridges,
and field-based samplers. We formulate these models as endpoint-conditioned latent path processes and study the fundamental compression from non-causal
bridge information to causal sampler representations. The Markovization gap quantifies the irreducible uncertainty of this projection and provides a
pre-training diagnostic of the difficulty of realizing a chosen bridge design with a Markov decoder.
arXiv · 2025 · online learning theory
Rank-Induced PL Mirror Descent: A Rank-Faithful Second-Order Algorithm for Sleeping Experts
Tiantian Zhang · Technical Feedback from Prof.Daniel Hsu
A mirror-descent algorithm for sleeping experts under partial availability, using a
rank-induced Plackett–Luce parameterization. The project studies how online learners can
preserve rank structure while adapting to variance in sequential decision problems.
Ongoing · decision-focused optimization
Deep Decision Optimization: Regularized FY/SPO+ Surrogates and Compressed Moment Oracles
Tiantian Zhang · technical feedback from Prof. Adam Elmachtoub
An active predict-then-optimize direction that views SPO+ through Fenchel–Young/support-function
geometry, then replaces hard exposed-face subgradients with smooth support moments. Current
work studies compressed moment oracles, including PL, soft dynamic programming, Sinkhorn,
perturbation pools, and geometry-controlled neural oracles.
Industry research · portfolio optimization
Dynamic Prediction + Decision Optimization for Portfolio Decisions
Boke Technology · manuscript in preparation
A dynamic prediction-to-decision project studying multi-horizon forecasting and downstream
portfolio optimization. Current diagnostics emphasize fair comparison to simple baselines,
conservative decision learning, and the gap between predictive accuracy and realized portfolio
decisions.
Independent study · combinatorial optimization
Extension Complexity of the Stable-Set Polytope
Advisor: Prof. Yuri Faenza
A combinatorial optimization project studying extension-complexity lower bounds through
polynomial and geometric methods over finite fields, connecting slack-matrix structure with
communication-complexity style arguments.
manuscript in preparation
Earlier research · robust and generative learning
Causal DRO and Rectified Flow Models
Research training with Prof. Kede Ma and Prof. Qiang Liu
Earlier projects studied causal constraints for robust generalization and rectified-flow ideas
for generative learning and transport-based sampling. This work helped motivate the later
Bridge Graphical Models direction.
paper links pending
