Selected work

Representative Papers

Differentiable Rule Induction from Raw Sequence Inputs

Feb 14, 2026International Conference on Learning Representations

Existing differentiable inductive logic programming (ILP) approaches struggle to learn symbolic rules directly from raw continuous data—such as time-series or images—primarily due to the explicit label leakage problem: without supervision from feature-level labels, they cannot reliably map continuous inputs to symbolic variables. This work proposes an end-to-end neuro-symbolic framework that integrates self-supervised differentiable clustering with a novel differentiable ILP formulation, enabling direct learning of interpretable symbolic rules from raw data without requiring explicit labels. By circumventing the label leakage bottleneck for the first time, the method preserves rule interpretability while substantially improving generalization and applicability. Experiments on both temporal and visual tasks demonstrate its ability to discover accurate and intuitively meaningful symbolic rules.

2 citationsRead paper

Evolutionary Optimization of Physics-Informed Neural Networks: Advancing Generalizability by the Baldwin Effect

Dec 06, 2023

Existing Physics-Informed Neural Networks (PINNs) are trained task-specifically, exhibiting poor generalization across distinct physical problems and incurring high retraining costs. This work proposes an evolutionary meta-learning framework inspired by the Baldwin effect—the first to integrate evolutionary optimization into PINNs—via a two-stage stochastic programming scheme that jointly optimizes population-level structural priors (encoding physics-informed preferences) and individual-level task adaptation, enabling rapid, constraint-compliant transfer without gradient backpropagation. On strongly nonlinear PDEs (e.g., diffusion-reaction systems), our method achieves a 70× improvement in prediction accuracy and a 700× reduction in computational time. Compared to state-of-the-art gradient-based meta-learning approaches, it attains over one order-of-magnitude higher accuracy while significantly lowering computational overhead. The core contribution is the establishment of the first unified PINN generalization paradigm integrating evolutionary optimization, physics-informed priors, and lifelong learning.

2 citationsRead paper

The Blending Ratio Is Not Where the Performance Is: Diagnosing Prototype Blending for Few-Shot Adaptation of Vision-Language Models

Aug 23, 2026

本文探讨了视觉-语言模型少样本适应中混合比的问题，提出了一种无需验证数据估计最优混合比的方法，并证明性能提升关键不在于混合比的优化。

0 citationsRead paper

Corrupting Attention: Evasion-Based Adversarial Attacks on Encoder Attention in Detection Transformers

Aug 06, 2026

Existing adversarial attacks struggle to effectively disrupt the encoder attention mechanisms of detection Transformers. This work proposes a novel ℓ∞-constrained adversarial attack that, for the first time, directly optimizes encoder attention objectives by guiding gradients to steer model attention toward perturbed regions, thereby impairing its spatial reasoning capability. The method is agnostic to specific target formulations and supports diverse attention disruption strategies—including dispersion, reordering, permutation, and peak suppression—while remaining compatible with general attention architectures. Evaluated on COCO, it reduces the mAP of DETR-R50 and DINO-Swin-L to 0.97 and 1.44, respectively, substantially outperforming the strongest existing attacks and revealing that the attention structure itself constitutes a critical vulnerability in model robustness.

0 citationsRead paper

Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition

Aug 05, 2026

This work addresses the limitations of existing methods in privacy preservation, computational efficiency, and cross-domain generalization by proposing a lightweight motion reasoning framework for anomaly detection in real classroom settings. The authors construct a hybrid benchmark by fusing synthetic CCTV videos with real-world classroom pose data and, for the first time, distill multi-order dynamic features—encompassing direction, velocity, acceleration, and intensity—into a single-order lightweight model. Operating under strict privacy constraints, the proposed approach achieves superior performance at less than one-tenth the computational cost of baseline methods and demonstrates remarkable zero-shot transfer capability from synthetic to real-world domains.

0 citationsRead paper

Recent publications

Latest Papers

The Blending Ratio Is Not Where the Performance Is: Diagnosing Prototype Blending for Few-Shot Adaptation of Vision-Language Models

Aug 23, 2026

本文探讨了视觉-语言模型少样本适应中混合比的问题，提出了一种无需验证数据估计最优混合比的方法，并证明性能提升关键不在于混合比的优化。

0 citationsRead paper

Corrupting Attention: Evasion-Based Adversarial Attacks on Encoder Attention in Detection Transformers

Aug 06, 2026

Existing adversarial attacks struggle to effectively disrupt the encoder attention mechanisms of detection Transformers. This work proposes a novel ℓ∞-constrained adversarial attack that, for the first time, directly optimizes encoder attention objectives by guiding gradients to steer model attention toward perturbed regions, thereby impairing its spatial reasoning capability. The method is agnostic to specific target formulations and supports diverse attention disruption strategies—including dispersion, reordering, permutation, and peak suppression—while remaining compatible with general attention architectures. Evaluated on COCO, it reduces the mAP of DETR-R50 and DINO-Swin-L to 0.97 and 1.44, respectively, substantially outperforming the strongest existing attacks and revealing that the attention structure itself constitutes a critical vulnerability in model robustness.

0 citationsRead paper

Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition

Aug 05, 2026

This work addresses the limitations of existing methods in privacy preservation, computational efficiency, and cross-domain generalization by proposing a lightweight motion reasoning framework for anomaly detection in real classroom settings. The authors construct a hybrid benchmark by fusing synthetic CCTV videos with real-world classroom pose data and, for the first time, distill multi-order dynamic features—encompassing direction, velocity, acceleration, and intensity—into a single-order lightweight model. Operating under strict privacy constraints, the proposed approach achieves superior performance at less than one-tenth the computational cost of baseline methods and demonstrates remarkable zero-shot transfer capability from synthetic to real-world domains.

0 citationsRead paper

AirKey: Multimodal Acoustic-Assisted WiFi Sensing for Zero-Training Robust PIN Inference

Aug 04, 2026

This work addresses the challenge of achieving high-accuracy, training-free inference of contact-based keypad inputs in real-world settings, where pure WiFi-based sensing suffers from temporal ambiguity due to overlapping Channel State Information (CSI) waveforms caused by rapid keystrokes and network permission constraints. To overcome this limitation, the paper proposes AirKey, a cross-modal sensing framework that uniquely integrates acoustic signals with CSI extracted from WiFi ACK frames. By leveraging acoustic cues as precise temporal anchors to segment overlapped CSI traces, AirKey enables passive, zero-training PIN code inference without requiring network infrastructure modifications. Experimental results demonstrate that AirKey achieves over four times higher accuracy than existing zero-training, single-modality approaches in realistic environments, successfully recovering target PIN codes within six attempts.

0 citationsRead paper

GGC: Selective Query Correction for Reliable Text-to-SPARQL Generation

Jul 30, 2026

This work addresses the issue that SPARQL queries generated by large language models, while syntactically executable, often suffer from semantic inaccuracies leading to erroneous retrieval. To mitigate this, the authors propose the Generator-Gate-Corrector (GGC) framework, which first produces an initial query via a generator, then employs a gating module to assess its risk of error. Only queries flagged as high-risk are selectively refined by a corrector module, thereby avoiding unnecessary modifications to already accurate queries. This selective correction strategy enhances overall accuracy while substantially reducing inference overhead. Evaluated on the MCQA dataset, GGC improves query accuracy from 90.23% to 98.33% and cuts inference cost by 45% compared to applying correction universally.

0 citationsRead paper