Selected work

Representative Papers

Dual-Context Analog Retrieval for Time Series Forecasting

Oct 02, 2026

This study addresses the limitations of single-step mapping, which neglects historical dependencies, and the unreliability of traditional retrieval-based matching in long-term time series forecasting. To this end, we propose DuoTS, a model-agnostic framework that employs parallel patch encoding to generate base predictions. It innovatively introduces a dual-context mechanism that integrates local dynamics with global analogical evidence, alongside a segment-wise progressive refinement strategy to balance historical information across varying temporal distances, thereby achieving patch-level optimization. Extensive experiments demonstrate that DuoTS attains state-of-the-art performance on multiple real-world datasets. Furthermore, the proposed refinement mechanism can be seamlessly integrated into existing forecasting models, and ablation studies comprehensively validate the effectiveness of each individual component.

0 citationsRead paper

Most-Recent Anchoring with Recurrent Ordering for Time Series Forecasting

Oct 02, 2026

This study addresses the limitation of existing long-horizon forecasting models that employ fixed stacking, which applies uniform computational depth to both recent and distant evidence, thereby failing to effectively distinguish information recency. To overcome this, we propose MARO, a model that recursively scans from the most recent to the oldest data chunk, initializing and conditioning subsequent steps with the latest observation as an anchor to fold historical context into a recency-centric representation. By leveraging a shared recurrent module, this architecture extends the scanning range without increasing parameter count, utilizing intermediate states to achieve differentiated weighting for short- and long-term history. Experiments on multiple real-world datasets demonstrate state-of-the-art performance in both short- and long-term forecasting, while ablation studies validate the critical contributions of recursive ordering, recency anchoring, and parameter sharing mechanisms.

0 citationsRead paper

MixBench-TS: A Multivariate Time Series Forecasting Benchmark Where Channel Mixing Pays Off

Sep 26, 2026

This study addresses the deficiency of existing multivariate time series forecasting benchmarks in validating cross-channel coupling, which leads to distorted model evaluation. We propose the first Channel Dependency (CD) Gain metric, integrating lagged mutual information, Granger causality, and transfer entropy, to reveal the insufficient coupling inherent in standard datasets. Building upon this insight, we construct MixBench, a real-world benchmark characterized by high inter-channel coupling, and systematically evaluate mainstream multivariate forecasting models. Experimental results demonstrate that on MixBench, channel-mixing models significantly outperform channel-independent approaches, confirming the critical value of cross-channel coupling for accurate forecasting. This work establishes a more rigorous evaluation paradigm for multivariate time series forecasting.

0 citationsRead paper

Rethinking the Fully Hyperbolic Vision Transformer in Polar Coordinates

Sep 26, 2026

This study addresses the numerical instability and performance limitations of hyperbolic visual Transformers in ambient coordinates caused by large radii. To overcome these issues, this work proposes a fully hyperbolic Vision Transformer formulated in polar coordinates. Methodologically, it introduces a first-of-its-kind polar-coordinate fully connected layer that independently learns radius mappings, designs logarithmically growing hyperbolic displacements as relative positional encodings, and reformulates residual connections as Lorentz boosts based on the mean radius. These innovations effectively decouple directional and radial computations, mitigating numerical errors while fully exploiting the hierarchical properties of hyperbolic geometry. Experimental results demonstrate that the proposed method significantly outperforms both Euclidean and existing hyperbolic baselines on standard vision tasks, further exhibiting strong generalization capabilities on ImageNet.

0 citationsRead paper

NAE: Normalizing AutoEncoder

Aug 12, 2026

This work addresses a critical inconsistency between the existing loss functions used in flow-based autoencoders and their reconstruction objectives, which leads to suboptimal training dynamics. The paper provides the first theoretical analysis of this misalignment and introduces a novel conditional loss function designed to align the surrogate gradients of the encoder and decoder with the true reconstruction loss. By preserving the established architecture that combines normalizing flows with autoencoders, the proposed method significantly enhances generative performance. It achieves state-of-the-art results across diverse benchmarks—including molecular generation, tabular data modeling, and image synthesis—demonstrating both its effectiveness and broad applicability.

0 citationsRead paper

Recent publications

Latest Papers

Dual-Context Analog Retrieval for Time Series Forecasting

Oct 02, 2026

This study addresses the limitations of single-step mapping, which neglects historical dependencies, and the unreliability of traditional retrieval-based matching in long-term time series forecasting. To this end, we propose DuoTS, a model-agnostic framework that employs parallel patch encoding to generate base predictions. It innovatively introduces a dual-context mechanism that integrates local dynamics with global analogical evidence, alongside a segment-wise progressive refinement strategy to balance historical information across varying temporal distances, thereby achieving patch-level optimization. Extensive experiments demonstrate that DuoTS attains state-of-the-art performance on multiple real-world datasets. Furthermore, the proposed refinement mechanism can be seamlessly integrated into existing forecasting models, and ablation studies comprehensively validate the effectiveness of each individual component.

0 citationsRead paper

Most-Recent Anchoring with Recurrent Ordering for Time Series Forecasting

Oct 02, 2026

This study addresses the limitation of existing long-horizon forecasting models that employ fixed stacking, which applies uniform computational depth to both recent and distant evidence, thereby failing to effectively distinguish information recency. To overcome this, we propose MARO, a model that recursively scans from the most recent to the oldest data chunk, initializing and conditioning subsequent steps with the latest observation as an anchor to fold historical context into a recency-centric representation. By leveraging a shared recurrent module, this architecture extends the scanning range without increasing parameter count, utilizing intermediate states to achieve differentiated weighting for short- and long-term history. Experiments on multiple real-world datasets demonstrate state-of-the-art performance in both short- and long-term forecasting, while ablation studies validate the critical contributions of recursive ordering, recency anchoring, and parameter sharing mechanisms.

0 citationsRead paper

MixBench-TS: A Multivariate Time Series Forecasting Benchmark Where Channel Mixing Pays Off

Sep 26, 2026

This study addresses the deficiency of existing multivariate time series forecasting benchmarks in validating cross-channel coupling, which leads to distorted model evaluation. We propose the first Channel Dependency (CD) Gain metric, integrating lagged mutual information, Granger causality, and transfer entropy, to reveal the insufficient coupling inherent in standard datasets. Building upon this insight, we construct MixBench, a real-world benchmark characterized by high inter-channel coupling, and systematically evaluate mainstream multivariate forecasting models. Experimental results demonstrate that on MixBench, channel-mixing models significantly outperform channel-independent approaches, confirming the critical value of cross-channel coupling for accurate forecasting. This work establishes a more rigorous evaluation paradigm for multivariate time series forecasting.

0 citationsRead paper

Rethinking the Fully Hyperbolic Vision Transformer in Polar Coordinates

Sep 26, 2026

This study addresses the numerical instability and performance limitations of hyperbolic visual Transformers in ambient coordinates caused by large radii. To overcome these issues, this work proposes a fully hyperbolic Vision Transformer formulated in polar coordinates. Methodologically, it introduces a first-of-its-kind polar-coordinate fully connected layer that independently learns radius mappings, designs logarithmically growing hyperbolic displacements as relative positional encodings, and reformulates residual connections as Lorentz boosts based on the mean radius. These innovations effectively decouple directional and radial computations, mitigating numerical errors while fully exploiting the hierarchical properties of hyperbolic geometry. Experimental results demonstrate that the proposed method significantly outperforms both Euclidean and existing hyperbolic baselines on standard vision tasks, further exhibiting strong generalization capabilities on ImageNet.

0 citationsRead paper

NAE: Normalizing AutoEncoder

Aug 12, 2026

This work addresses a critical inconsistency between the existing loss functions used in flow-based autoencoders and their reconstruction objectives, which leads to suboptimal training dynamics. The paper provides the first theoretical analysis of this misalignment and introduces a novel conditional loss function designed to align the surrogate gradients of the encoder and decoder with the true reconstruction loss. By preserving the established architecture that combines normalizing flows with autoencoders, the proposed method significantly enhances generative performance. It achieves state-of-the-art results across diverse benchmarks—including molecular generation, tabular data modeling, and image synthesis—demonstrating both its effectiveness and broad applicability.

0 citationsRead paper