My research is at the intersection of multimodal learning, large language models, and biomedical AI.
I build interpretable multimodal mixture-of-experts (MoE) systems, study the statistical foundations of large language models, and apply machine learning to drug discovery and healthcare.
* denotes equal contribution
ICML 2025
We introduce a drop-in MoE framework that disentangles redundancy, synergy, and uniqueness interactions between different modalities, achieving superior multimodal fusion performance while exposing interaction weights for user interpretation.
AAAI GenAI4Health Workshop 2025 Best Paper
We impute missing modalities with learnable embeddings and route them through a sparse MoE, outperforming strong baselines on real-world multimodal datasets.
ICML 2025
We propose Adaptive Modality Token Re-BalanCing (AMC), a novel top-down dynamic multi-modal fusion approach.
NeurIPS 2024 Spotlight
We propose Flex-MoE, new MoE framework designed to flexibly incorporate arbitrary modality combinations while maintaining robustness to missing data.
ACL 2026 Findings
We use Smoothed Good-Turing estimator to estimate latent cluster coverage for LLM in-context learning, improving demonstration selection with fixed budget.
arXiv 2026
We propose CurveRL, a principled distribution-aware context reweighting method that assigns weights based on rank and density to reflect the distributional structure of pass rates during LLM reasoning training, consistently outperforming existing RLVR baselines.
EMNLP 2025 Oral
We study language switching in Chinese-English bilingual reasoning models and trained a lightweight probe to predict whether a potential language switch would benefit or harm reasoning, and when used to guide decoding, increases accuracy by up to 6.25 pp.
Under review 2025
We introduce KnowSum, a statistical framework designed to provide a more comprehensive assessment by quantifying the unseen knowledge for a class of evaluation tasks.
ICLR 2023
We formulate an annotation-efficient, two-step framework: selective annotation that chooses a pool of examples to annotate from unlabeled data in advance, followed by prompt retrieval that retrieves task examples from the annotated pool at test time.
arXiv 2026
We develop CAMEL, the first ECG language model for forecasting cardiac events via extended signal analysis, combining a specialized ECG encoder with curriculum learning to achieve state-of-the-art results on multiple benchmarks.
NeurIPS 2025 AI4Science and FM4LS Workshop
We develop a multimodal machine learning framework to predict antibody reformatting success, which incorporates both antibody sequence and structural context, and outperforms protein language models in an evaluation protocol that reflects realistic deployment scenarios.
IEEE TMI 2025
We propose a novel Task-Aware Disentanglement Graph approach that operates on WSI graph representations, effectively identifying and disentangling informative subgraphs to enhance contextual feature extraction.
EMNLP 2024
We show that a simple alternative, Retrieved Sequence Augmentation (RSA), can enhance protein representation learning without the need for alignment and cumbersome preprocessing.
Journal of Chemical Information and Modeling 2023
We extend previous spectrum Transformer methodology for learning to rank chemical formula and adduct assignments given an unannotated tandem MS spectrum.
NeurIPS 2023 Spotlight
We use a new intermediate strategy for predicting mass spectra from molecules by treating mass spectra as sets of molecular formulae, which are themselves multisets of atoms.
Book chapter in Machine Learning, Medical AI and Robotics (IOP) 2023
We survey deep learning-based AI applications of wearable devices.