Published in KDD 2025 Machine Learning on Graphs Workshop (Oral) & KDD 2026 (Data & Benchmarks), 2025
We design a temporal ‘obstacle course’ to systematically benchmarks the performance of TGNNs at carefully designed synthetic temporal reasoning tasks.
Published in ICML 2026 Workshop on Combining Theory and Benchmarks: Towards A Virtuous Cycle to Understand and Guarantee Foundation Model Performance (T&B), 2026
We analyze the feedforward mixing process in neural networks and establish backward mixing time as a key metric for evaluating information propagation.
Published in NeurIPS Symmetry and Geometry in Neural Representations Workshop 2024 (Oral), 2024
We show that adding source-target identification to Temporal Graph Networks (TGNs) strictly increases its expressivity and demonstrate that the resulting architecture outperforms all TG methods in dynamic node affinity prediction on all TGB datasets.
Published in NeurIPS Safe Generative AI Workshop 2024, 2024
We fine-tune LLMs using semantic entropy to make them abstain from answering uncertain questions.
Published in arXiv, 2021
We conducted experiments using graph neural network ensembles to predict molecular HOMO-LUMO gaps.