Selected work

Representative Papers

GENADA: efficient generative time series adversarial attack framework

Aug 12, 2026

This work addresses the vulnerability of deep learning models for time series to adversarial attacks and the high computational cost of existing gradient-based attack methods. To this end, the authors propose GENADA, a novel framework that introduces generative modeling into time series adversarial attacks for the first time. Built upon generative adversarial networks, GENADA efficiently produces adversarial perturbations in a single forward pass and supports both single-step and iterative attack strategies. Experimental results demonstrate that GENADA achieves attack effectiveness comparable to strong baselines across multiple datasets and models, while substantially reducing the time required for perturbation generation and significantly improving inference efficiency.

0 citationsRead paper

Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training

Aug 11, 2026

This study addresses the unresolved trade-off between batch size and the number of negative samples under fixed memory budgets in training memory-constrained recommender systems. It theoretically and empirically demonstrates, for the first time, that when using sampled Softmax, prioritizing larger batch sizes over a greater number of negative samples yields faster convergence and superior recommendation quality within the same memory constraints. The proposed configuration principle is validated across four real-world sequential recommendation benchmarks—including MovieLens-20M—as well as synthetic data, offering clear guidance for efficient training in resource-limited scenarios.

0 citationsRead paper

Imagine to Ensure Safety in Hierarchical Reinforcement Learning

Jun 21, 2026

This work addresses the challenge of safety constraint violations in long-horizon reinforcement learning tasks, which often arise from accumulated errors and limited exploration. To mitigate these issues, the paper proposes a novel safety-aware hierarchical reinforcement learning framework that integrates a learnable world model with a two-level policy architecture. The high-level policy generates safety-oriented subgoals, while the low-level policy leverages imagined rollouts within the learned predictive environment to evaluate and correct unsafe actions before execution, thereby enforcing safety at both levels. This approach is the first to incorporate imagination-based mechanisms into hierarchical reinforcement learning, effectively reducing error accumulation. Empirical results demonstrate that the method significantly improves constraint satisfaction rates and consistently adheres to predefined safety budgets in high-dimensional navigation and manipulation tasks, outperforming state-of-the-art safe reinforcement learning baselines.

0 citationsRead paper

GENEB: Why Genomic Models Are Hard to Compare

Jun 03, 2026

This study addresses the lack of standardized evaluation protocols that hinder fair assessment of genomic foundation models’ performance and generalization. To this end, the authors introduce GENEB, a large-scale diagnostic benchmark that systematically evaluates frozen representations from 40 models across 100 tasks under a unified probing protocol, spanning 13 functional categories and supporting few-shot settings. This framework enables, for the first time, category-aware, fine-grained, and controllable multidimensional comparisons, revealing the instability of aggregate leaderboards and inherent trade-offs across tasks. Key findings indicate substantial variation in model rankings across functional categories, limited and inconsistent gains from increased model scale, and a more decisive influence of architectural design and alignment between pretraining data and downstream tasks than parameter count alone.

0 citationsRead paper

LDARNet: DNA Adaptive Representation Network with Learnable Tokenization for Genomic Modeling

Jun 03, 2026

This work proposes the first hierarchical foundation model for DNA sequence modeling—featuring 120 million parameters—that learns tokenization boundaries in an unsupervised manner, overcoming the limitations of fixed tokenization schemes (e.g., k-mers or BPE) which often disrupt biologically meaningful sequence structures. By employing a dynamic chunking mechanism, the model adaptively captures functional genomic elements, with learned token boundaries showing strong alignment to promoter motifs and splice sites, thereby significantly enhancing interpretability. The architecture integrates BiMamba-2 state space models, local attention, bidirectional routing, and ratio-based regularization. It achieves state-of-the-art performance on five histone modification tasks and leads in 11 out of 27 benchmark tasks overall, outperforming models with 20× more parameters. FLOPs-matched comparisons demonstrate that learnable tokenization yields up to a 14-percentage-point improvement over fixed-token baselines.

0 citationsRead paper

Recent publications

Latest Papers

GENADA: efficient generative time series adversarial attack framework

Aug 12, 2026

This work addresses the vulnerability of deep learning models for time series to adversarial attacks and the high computational cost of existing gradient-based attack methods. To this end, the authors propose GENADA, a novel framework that introduces generative modeling into time series adversarial attacks for the first time. Built upon generative adversarial networks, GENADA efficiently produces adversarial perturbations in a single forward pass and supports both single-step and iterative attack strategies. Experimental results demonstrate that GENADA achieves attack effectiveness comparable to strong baselines across multiple datasets and models, while substantially reducing the time required for perturbation generation and significantly improving inference efficiency.

0 citationsRead paper

Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training

Aug 11, 2026

This study addresses the unresolved trade-off between batch size and the number of negative samples under fixed memory budgets in training memory-constrained recommender systems. It theoretically and empirically demonstrates, for the first time, that when using sampled Softmax, prioritizing larger batch sizes over a greater number of negative samples yields faster convergence and superior recommendation quality within the same memory constraints. The proposed configuration principle is validated across four real-world sequential recommendation benchmarks—including MovieLens-20M—as well as synthetic data, offering clear guidance for efficient training in resource-limited scenarios.

0 citationsRead paper

Imagine to Ensure Safety in Hierarchical Reinforcement Learning

Jun 21, 2026

This work addresses the challenge of safety constraint violations in long-horizon reinforcement learning tasks, which often arise from accumulated errors and limited exploration. To mitigate these issues, the paper proposes a novel safety-aware hierarchical reinforcement learning framework that integrates a learnable world model with a two-level policy architecture. The high-level policy generates safety-oriented subgoals, while the low-level policy leverages imagined rollouts within the learned predictive environment to evaluate and correct unsafe actions before execution, thereby enforcing safety at both levels. This approach is the first to incorporate imagination-based mechanisms into hierarchical reinforcement learning, effectively reducing error accumulation. Empirical results demonstrate that the method significantly improves constraint satisfaction rates and consistently adheres to predefined safety budgets in high-dimensional navigation and manipulation tasks, outperforming state-of-the-art safe reinforcement learning baselines.

0 citationsRead paper

GENEB: Why Genomic Models Are Hard to Compare

Jun 03, 2026

This study addresses the lack of standardized evaluation protocols that hinder fair assessment of genomic foundation models’ performance and generalization. To this end, the authors introduce GENEB, a large-scale diagnostic benchmark that systematically evaluates frozen representations from 40 models across 100 tasks under a unified probing protocol, spanning 13 functional categories and supporting few-shot settings. This framework enables, for the first time, category-aware, fine-grained, and controllable multidimensional comparisons, revealing the instability of aggregate leaderboards and inherent trade-offs across tasks. Key findings indicate substantial variation in model rankings across functional categories, limited and inconsistent gains from increased model scale, and a more decisive influence of architectural design and alignment between pretraining data and downstream tasks than parameter count alone.

0 citationsRead paper

LDARNet: DNA Adaptive Representation Network with Learnable Tokenization for Genomic Modeling

Jun 03, 2026

This work proposes the first hierarchical foundation model for DNA sequence modeling—featuring 120 million parameters—that learns tokenization boundaries in an unsupervised manner, overcoming the limitations of fixed tokenization schemes (e.g., k-mers or BPE) which often disrupt biologically meaningful sequence structures. By employing a dynamic chunking mechanism, the model adaptively captures functional genomic elements, with learned token boundaries showing strong alignment to promoter motifs and splice sites, thereby significantly enhancing interpretability. The architecture integrates BiMamba-2 state space models, local attention, bidirectional routing, and ratio-based regularization. It achieves state-of-the-art performance on five histone modification tasks and leads in 11 out of 27 benchmark tasks overall, outperforming models with 20× more parameters. FLOPs-matched comparisons demonstrate that learnable tokenization yields up to a 14-percentage-point improvement over fixed-token baselines.

0 citationsRead paper