EVODiff: Entropy-aware Variance Optimized Diffusion Inference, Shigui Li, Wei Chen, Delu Zeng* [Bib]

NeurIPS 2025 | Paper | Code | News🎉

• Introduces an information-theoretic view: successful denoising reduces conditional entropy in reverse transitions.
• Proposes EVODiff, a reference-free diffusion inference framework that optimizes conditional variance to reduce transition and reconstruction errors, improving sample quality and reducing inference cost.

Entropy-informed weighting channel normalizing flow for deep generative models, Wei Chen#, Shian Du#, Shigui Li#, Delu Zeng*, John Paisley [Bib]

Pattern Recognition (PR) 2025 | Paper | Code

• Proposes EIW-Flow, enhancing normalizing flows with channel-wise weights and latent variable shuffling.
• Achieves state-of-the-art density estimation with minimal computational overhead.

ReciprocalLA-LLIE: Low-light image enhancement with luminance-aware reciprocal diffusion process, Zhiqi Lin, Wei Chen, Jian Xu, Delu Zeng*, Min Chen [Bib]

Neurocomputing 2025 | Paper

• Proposes a reciprocal diffusion process within DDPM for low-light image enhancement.
• Introduces Luminance Adjustment Block for robust global luminance control.

To-Flow: Efficient Continuous Normalizing Flows with Temporal Optimization Adjoint with Moving Speed, Shian Du#, Yihong Luo#, Wei Chen#, Jian Xu, Delu Zeng* [Bib]

CVPR 2022 | Paper | Code

• CNFs via neural ODEs are costly; To-Flow proposes temporal optimization via coordinate descent.
• Accelerates flow training by about 20% while maintaining performance.

Towards Disentangled Preference Optimization Dynamics Beyond Likelihood Displacement, Wei Chen, Yubing Wu, Junmei Yang, Delu Zeng*, Qibin Zhao, John Paisley, Min Chen, Zhou Wang [Bib]

arXiv 2026 | Paper | Code

• Presents a unified incentive-score decomposition showing that diverse preference optimization objectives share identical local update directions and differ only in scalar weighting coefficients.
• Identifies the disentanglement band (DB), a simple, testable condition characterizing when training avoids likelihood displacement by suppressing the rejected response while maintaining the chosen one.
• Proposes a plug-and-play reward calibration (RC) that adaptively rebalances chosen/rejected updates to satisfy the DB, mitigating likelihood displacement without redesigning the base objective and often improving downstream performance.

One-Step Score-Based Density Ratio Estimation, Wei Chen, Qibin Zhao, John Paisley, Junmei Yang, Delu Zeng* [Bib]

arXiv 2026 | Paper | Code

• Proposes OS-DRE, a partly analytic and solver-free score-based DRE framework that decomposes the time score into spatial and temporal components.
• Represents the temporal part with an analytic RBF frame, turning the otherwise intractable temporal integral into a closed-form weighted sum and enabling DRE with only one function evaluation.
• Establishes approximation error bounds for both finitely and infinitely smooth temporal kernels, grounding the framework in approximation theory.

A Minimum Variance Path Principle for Accurate and Stable Score-Based Density Ratio Estimation, Wei Chen, Jiacheng Li, Shigui Li, Zhiqi Lin, Junmei Yang, John Paisley, Delu Zeng* [Bib]

ICLR 2026 | Paper | Code

• Resolves the path schedule paradox in score-based DRE by identifying the overlooked path variance term.
• Proposes MVP Principle with closed-form variance expression for tractable optimization.
• Achieves state-of-the-art results on challenging DRE benchmarks.

Diffusion Secant Alignment for Score-Based Density Ratio Estimation, Wei Chen, Shigui Li, Jiacheng Li, Jian Xu, Zhiqi Lin, Junmei Yang, Delu Zeng*, John Paisley, Qibin Zhao [Bib]

arXiv 2025 | Paper | Code

• Proposes ISA-DRE, replacing the high-variance instantaneous tangent with its interval integral (the secant) as a provably lower-variance and smoother learning target along diffusion interpolants.
• Introduces the Secant Alignment Identity (SAI) to enforce self-consistency between secant and tangent, and Contraction Interval Annealing (CIA) for stable convergence.
• Achieves comparable or superior accuracy with fewer function evaluations and effectively mitigates the density-chasm problem under large distribution discrepancies.

Dequantified Diffusion-Schrödinger Bridge for Density Ratio Estimation, Wei Chen, Shigui Li, Jiacheng Li, Junmei Yang, John Paisley, Delu Zeng* [Bib]

ICML 2025 | Paper | Code

• Proposes D3RE, a unified framework for robust DRE under distribution mismatch.
• Introduces dequantified diffusion/SCHRödinger bridge interpolants for support expansion and stabilized scores.

DeepAR-Attention probabilistic prediction for stock price series, Jiacheng Li, Wei Chen, Zhiheng Zhou, Junmei Yang, Delu Zeng* [Bib]

Neural Computing and Applications 2024 | Paper

• Proposes DeepAR-Attention for probabilistic stock price forecasting.

Neural ordinary differential equation networks for fintech applications using IoT, Jiacheng Li, Wei Chen, Yican Liu, Junmei Yang, Delu Zeng*, Zhiheng Zhou [Bib]

IEEE Internet of Things Journal (IoTJ) 2024 | Paper

• Develops neural ODE network approaches for fintech applications in IoT settings.

Integrating Ordinary Differential Equations with Sparse Attention for Power Load Forecasting, Jiacheng Li, Wei Chen, Yican Liu, Junmei Yang, Zhiheng Zhou, Delu Zeng* [Bib]

IEEE Trans on Instrumentation and Measurement (T-IM) 2025 | Paper

• Proposes EvolvInformer: integrates ODE solver with ProbSparse attention for long-sequence load forecasting.
• Achieves 29.7% MSE reduction while preserving logarithmic memory complexity.

Generative Self-Supervised Time-Series Forecasting Leveraging Wavelet Diffusion, Jiacheng Li, Wei Chen, Yican Liu, Junmei Yang, Zhiheng Zhou, Delu Zeng* [Bib]

IEEE Trans on Instrumentation and Measurement (T-IM) 2025 | Paper | News🎉

• TimeWaveDiff: a lightweight self-supervised framework that integrates wavelet decomposition and diffusion modeling to capture multiscale periodicities and robustly learn complex noise patterns in measurement signals.
• Achieves superior long-term forecasting accuracy with significantly reduced computational cost.

Diffinformer: Diffusion Informer model for long sequence time-series forecasting, Jiacheng Li, Wei Chen, Yican Liu, Junmei Yang, Zhiheng Zhou, Delu Zeng* [Bib]

Expert Systems with Applications (ESWA) 2025 | Paper

• Proposes Diffinformer: combines conditional diffusion models with Informer’s ProbSparse attention.
• Demonstrates consistent improvements across five large-scale datasets.