Selected work

Representative Papers

Decoupling Classifier for Boosting Few-shot Object Detection and Instance Segmentation

May 20, 2025Neural Information Processing Systems

This work addresses classification bias in few-shot object detection (FSOD) and few-shot instance segmentation (FSIS) caused by instance-level label omission—where unlabeled instances in images are erroneously treated as negative examples. To mitigate this, we propose a lightweight dual-head decoupled classifier that splits the standard classification head into two parallel branches: a positive branch for explicitly modeling ground-truth instances, and a noise-negative branch for modeling pseudo-negative instances arising from label omissions. Our approach is the first to formally define and address instance-level label-omission bias. It introduces no additional parameters, computational overhead, or reliance on meta-learning, fine-tuning, or external data; instead, it operates seamlessly within standard two-stage detectors (e.g., Mask R-CNN) under conventional supervised training. Extensive experiments on PASCAL VOC and MS-COCO benchmarks demonstrate consistent and significant improvements over state-of-the-art methods for both FSOD and FSIS. The code is publicly available.

26 citations3 influentialRead paper

Understanding and Overcoming Cross-modal Fusion Bias in Multimodal Anomaly Detection From A Fisher Information Perspective

Aug 02, 2026

This work addresses the performance bottleneck in multimodal anomaly detection caused by cross-modal fusion bias. It introduces, for the first time, the Fisher Information Matrix to quantify this bias and proposes an Unbiased Cross-Modal Fusion Framework (UCFB) that enables plug-and-play bias calibration. UCFB leverages Fisher information-guided dynamic regularization and Canonical Correlation Analysis to adaptively adjust fusion weights, thereby mitigating inter-modal inconsistencies. Evaluated on RGB-D multimodal data, the method consistently achieves significant performance gains across one-class, multi-class, and few-shot settings on the MVTec 3D-AD and Eyecandies benchmarks.

0 citationsRead paper

A Density-Matrix Framework for Electronic-Structure Analysis of Functional-Group and Salt Effects in Lithium-Metal Electrolytes

Jul 28, 2026

Traditional quantum chemical methods struggle to efficiently unravel the complex interplay between functional groups and salt effects on electronic structures in lithium metal electrolytes. To address this challenge, this work introduces EMolStudio—the first AI-driven platform integrating density matrix prediction with idempotency constraints—to enable unified modeling of diverse electrolyte systems. By explicitly constructing the first solvation shell of Li⁺, generating functionalized molecules, and performing multiscale electronic structure readouts—including frontier orbitals, electrostatic potentials, bond orders, and electron localization functions—EMolStudio captures critical physicochemical features across scales. Trained on a dataset comprising 163,655 functionalized molecules and 22,500 solvation shells, the model systematically elucidates how functional groups modulate frontier orbitals and Li⁺ coordination, while revealing the pivotal role of anions in governing the spatial localization of HOMO and LUMO orbitals.

0 citationsRead paper

Synthesis4AD: Synthetic Anomalies are All You Need for 3D Anomaly Detection

Apr 06, 2026

This work addresses the challenges of scarce anomaly samples and long-tailed distributions in industrial 3D anomaly detection. The authors propose Synthesis4AD, an end-to-end framework that pioneers the use of multimodal large language models to interpret design specifications and generate executable instructions for synthesizing anomalies. Coupled with MPAS—a primitive-guided, high-dimensional controllable synthesis engine—it enables large-scale generation of geometrically realistic anomalies with precise pixel-level masks. The framework further enhances training of Point Transformer detectors through spatial distribution normalization and geometry-preserving data augmentation strategies. Extensive experiments demonstrate state-of-the-art performance on Real3D-AD, MulSen-AD, and real-world industrial part datasets. Both the MPAS synthesis method and the 3D-DefectStudio platform will be publicly released.

0 citationsRead paper

Towards an Incremental Unified Multimodal Anomaly Detection: Augmenting Multimodal Denoising From an Information Bottleneck Perspective

Mar 03, 2026

This work addresses catastrophic forgetting in incremental unified multi-modal anomaly detection, which is primarily caused by redundant and spurious features. To this end, the authors propose the IB-IUMAD framework, which, for the first time, elucidates the underlying mechanism from an information bottleneck perspective. The method employs a Mamba decoder to decouple feature entanglements among objects and introduces an information bottleneck fusion module that explicitly preserves discriminative information while filtering out redundancy. This enables both cross-category generalization and continual learning of novel classes. Experimental results on the MVTec 3D-AD and Eyecandies datasets demonstrate that the proposed approach effectively mitigates catastrophic forgetting and achieves competitive anomaly detection performance.

0 citationsRead paper

Recent publications

Latest Papers

Understanding and Overcoming Cross-modal Fusion Bias in Multimodal Anomaly Detection From A Fisher Information Perspective

Aug 02, 2026

This work addresses the performance bottleneck in multimodal anomaly detection caused by cross-modal fusion bias. It introduces, for the first time, the Fisher Information Matrix to quantify this bias and proposes an Unbiased Cross-Modal Fusion Framework (UCFB) that enables plug-and-play bias calibration. UCFB leverages Fisher information-guided dynamic regularization and Canonical Correlation Analysis to adaptively adjust fusion weights, thereby mitigating inter-modal inconsistencies. Evaluated on RGB-D multimodal data, the method consistently achieves significant performance gains across one-class, multi-class, and few-shot settings on the MVTec 3D-AD and Eyecandies benchmarks.

0 citationsRead paper

A Density-Matrix Framework for Electronic-Structure Analysis of Functional-Group and Salt Effects in Lithium-Metal Electrolytes

Jul 28, 2026

Traditional quantum chemical methods struggle to efficiently unravel the complex interplay between functional groups and salt effects on electronic structures in lithium metal electrolytes. To address this challenge, this work introduces EMolStudio—the first AI-driven platform integrating density matrix prediction with idempotency constraints—to enable unified modeling of diverse electrolyte systems. By explicitly constructing the first solvation shell of Li⁺, generating functionalized molecules, and performing multiscale electronic structure readouts—including frontier orbitals, electrostatic potentials, bond orders, and electron localization functions—EMolStudio captures critical physicochemical features across scales. Trained on a dataset comprising 163,655 functionalized molecules and 22,500 solvation shells, the model systematically elucidates how functional groups modulate frontier orbitals and Li⁺ coordination, while revealing the pivotal role of anions in governing the spatial localization of HOMO and LUMO orbitals.

0 citationsRead paper

Synthesis4AD: Synthetic Anomalies are All You Need for 3D Anomaly Detection

Apr 06, 2026

This work addresses the challenges of scarce anomaly samples and long-tailed distributions in industrial 3D anomaly detection. The authors propose Synthesis4AD, an end-to-end framework that pioneers the use of multimodal large language models to interpret design specifications and generate executable instructions for synthesizing anomalies. Coupled with MPAS—a primitive-guided, high-dimensional controllable synthesis engine—it enables large-scale generation of geometrically realistic anomalies with precise pixel-level masks. The framework further enhances training of Point Transformer detectors through spatial distribution normalization and geometry-preserving data augmentation strategies. Extensive experiments demonstrate state-of-the-art performance on Real3D-AD, MulSen-AD, and real-world industrial part datasets. Both the MPAS synthesis method and the 3D-DefectStudio platform will be publicly released.

0 citationsRead paper

Towards an Incremental Unified Multimodal Anomaly Detection: Augmenting Multimodal Denoising From an Information Bottleneck Perspective

Mar 03, 2026

This work addresses catastrophic forgetting in incremental unified multi-modal anomaly detection, which is primarily caused by redundant and spurious features. To this end, the authors propose the IB-IUMAD framework, which, for the first time, elucidates the underlying mechanism from an information bottleneck perspective. The method employs a Mamba decoder to decouple feature entanglements among objects and introduces an information bottleneck fusion module that explicitly preserves discriminative information while filtering out redundancy. This enables both cross-category generalization and continual learning of novel classes. Experimental results on the MVTec 3D-AD and Eyecandies datasets demonstrate that the proposed approach effectively mitigates catastrophic forgetting and achieves competitive anomaly detection performance.

0 citationsRead paper

Evolving Interdependent Operators with Large Language Models for Multi-Objective Combinatorial Optimization

Jan 25, 2026

This work proposes E2OC, a novel framework for multi-objective evolutionary algorithms that addresses the challenge of modeling dynamic couplings among multiple neighborhood search operators—a task traditionally reliant on expert-designed heuristics and inadequately handled by existing large language model (LLM)-based approaches. E2OC formulates multi-operator optimization as a Markov decision process, explicitly capturing inter-operator dependencies for the first time. By integrating a co-evolutionary mechanism with an operator rotation strategy, it jointly optimizes both high-level design policies and executable code. Leveraging Monte Carlo tree search for progressive exploration and LLM-driven heuristic generation, E2OC consistently outperforms state-of-the-art methods across varying numbers of objectives and problem scales, demonstrating superior generalization and sustained optimization capability.

0 citationsRead paper