Selected work

Representative Papers

Graph Neural Networks with Triangle-Based Messages for the Multicut Problem

May 13, 2026

This work addresses the NP-hard Multicut problem, which finds broad applications in bioinformatics, data mining, and computer vision, by proposing a specialized graph neural network architecture. The method models features exclusively on edges and introduces, for the first time, a message-passing mechanism that leverages triangular structures inherent in the input graph to align with the objective function and constraints of the Multicut problem. Experimental results demonstrate that the proposed approach outperforms state-of-the-art heuristic solvers in solution quality on both synthetic and real-world instances with up to 200 nodes. Notably, it obtains optimal solutions within seconds on certain instances where exact solvers require several hours.

0 citationsRead paper

Predicting Microbial Interactions Using Graph Neural Networks

Nov 03, 2025

Predicting interspecies microbial interactions is a fundamental challenge in deciphering microbial community structure and function. This paper introduces the first graph neural network (GNN) framework tailored for large-scale microbial interaction prediction, where species pairs are modeled as edges and co-culture experiments as nodes. The framework integrates multi-source features—including monoculture growth phenotypes, phylogenetic distances, and prior knowledge of known interactions—to enable fine-grained, directional prediction of interaction types (e.g., mutualism, competition, parasitism). Innovatively, it leverages edge-centric graph structures to capture cross-experiment shared information, overcoming limitations of conventional classifiers in modeling interaction directionality and type specificity. Evaluated on a benchmark dataset comprising over 7,500 experimentally validated interactions, our method achieves an F1-score of 80.44%, significantly outperforming XGBoost (72.76%). Results demonstrate superior predictive accuracy and enhanced biological interpretability.

0 citationsRead paper

The Geometry of Nonlinear Reinforcement Learning

Sep 01, 2025

This paper addresses the challenge of jointly optimizing multiple objectives—reward maximization, safe exploration, and intrinsic motivation—in reinforcement learning. Methodologically, it introduces the first unified geometric optimization framework that generalizes classical algorithms such as policy mirror descent and natural policy gradient to settings involving nonlinear utility functions and convex constraints, integrating differential geometry and convex optimization to construct a trust-region-style nonlinear policy optimization framework for deep RL. Theoretically, it uncovers a shared geometric structure of multi-objective trade-offs in the space of long-horizon behavioral trajectories. Algorithmically, it unifies the modeling of robustness, safety, and exploratory diversity within a single principled formulation. This framework establishes a novel theoretical foundation for safe reinforcement learning and efficient exploration, while providing a scalable and modular paradigm for algorithm design.

0 citationsRead paper

Multimodal Recurrent Ensembles for Predicting Brain Responses to Naturalistic Movies (Algonauts 2025)

Jul 23, 2025

This study addresses the challenge of improving prediction accuracy for distributed cortical fMRI responses evoked by naturalistic film stimuli. To model multimodal temporal dynamics, we propose a hierarchical multimodal recurrent ensemble model: modality-specific bidirectional RNNs encode temporal evolution of video, audio, and pretrained language embeddings; a hierarchical feature fusion mechanism is coupled with a curriculum learning strategy progressing from sensory to association cortices; and a lightweight subject-specific output head is trained with an MSE–correlation composite loss. Robustness is enhanced via ensemble averaging over 100 model variants. Our approach ranked third in the Algonauts 2025 Challenge (overall r = 0.2094), achieved a peak average correlation of 0.63 across individual brain regions, and notably improved prediction performance for the most challenging subject (Subject 5).

0 citationsRead paper

Recent publications

Latest Papers

Graph Neural Networks with Triangle-Based Messages for the Multicut Problem

May 13, 2026

This work addresses the NP-hard Multicut problem, which finds broad applications in bioinformatics, data mining, and computer vision, by proposing a specialized graph neural network architecture. The method models features exclusively on edges and introduces, for the first time, a message-passing mechanism that leverages triangular structures inherent in the input graph to align with the objective function and constraints of the Multicut problem. Experimental results demonstrate that the proposed approach outperforms state-of-the-art heuristic solvers in solution quality on both synthetic and real-world instances with up to 200 nodes. Notably, it obtains optimal solutions within seconds on certain instances where exact solvers require several hours.

0 citationsRead paper

Predicting Microbial Interactions Using Graph Neural Networks

Nov 03, 2025

Predicting interspecies microbial interactions is a fundamental challenge in deciphering microbial community structure and function. This paper introduces the first graph neural network (GNN) framework tailored for large-scale microbial interaction prediction, where species pairs are modeled as edges and co-culture experiments as nodes. The framework integrates multi-source features—including monoculture growth phenotypes, phylogenetic distances, and prior knowledge of known interactions—to enable fine-grained, directional prediction of interaction types (e.g., mutualism, competition, parasitism). Innovatively, it leverages edge-centric graph structures to capture cross-experiment shared information, overcoming limitations of conventional classifiers in modeling interaction directionality and type specificity. Evaluated on a benchmark dataset comprising over 7,500 experimentally validated interactions, our method achieves an F1-score of 80.44%, significantly outperforming XGBoost (72.76%). Results demonstrate superior predictive accuracy and enhanced biological interpretability.

0 citationsRead paper

The Geometry of Nonlinear Reinforcement Learning

Sep 01, 2025

This paper addresses the challenge of jointly optimizing multiple objectives—reward maximization, safe exploration, and intrinsic motivation—in reinforcement learning. Methodologically, it introduces the first unified geometric optimization framework that generalizes classical algorithms such as policy mirror descent and natural policy gradient to settings involving nonlinear utility functions and convex constraints, integrating differential geometry and convex optimization to construct a trust-region-style nonlinear policy optimization framework for deep RL. Theoretically, it uncovers a shared geometric structure of multi-objective trade-offs in the space of long-horizon behavioral trajectories. Algorithmically, it unifies the modeling of robustness, safety, and exploratory diversity within a single principled formulation. This framework establishes a novel theoretical foundation for safe reinforcement learning and efficient exploration, while providing a scalable and modular paradigm for algorithm design.

0 citationsRead paper

Multimodal Recurrent Ensembles for Predicting Brain Responses to Naturalistic Movies (Algonauts 2025)

Jul 23, 2025

This study addresses the challenge of improving prediction accuracy for distributed cortical fMRI responses evoked by naturalistic film stimuli. To model multimodal temporal dynamics, we propose a hierarchical multimodal recurrent ensemble model: modality-specific bidirectional RNNs encode temporal evolution of video, audio, and pretrained language embeddings; a hierarchical feature fusion mechanism is coupled with a curriculum learning strategy progressing from sensory to association cortices; and a lightweight subject-specific output head is trained with an MSE–correlation composite loss. Robustness is enhanced via ensemble averaging over 100 model variants. Our approach ranked third in the Algonauts 2025 Challenge (overall r = 0.2094), achieved a peak average correlation of 0.63 across individual brain regions, and notably improved prediction performance for the most challenging subject (Subject 5).

0 citationsRead paper