2026
MulRobBench: A Decision-Level Benchmark for Safe and Security-Policy-Compliant Multimodal UAV Agents
Belal S. Alsinglawi†, Weizheng Wang†, Junyi Wu†, Lianhai Lin, Mérouane Debbah, Izzat Alsmadi.
2026 · Submitted to Applied Soft Computing. arXiv:2607.23870
An offline, protocol-conditioned UAV decision benchmark with 3,024 samples, 17 task-taxonomy nodes, 12 scoring dimensions, and evaluation of 17 multimodal models. Separates semantic decision quality from structural compliance, unsafe actions, and parsing failures.
H2T-DEIM: Hilbert–Haar Representation with Tri-Modal Multi-Dimensional Pruning for Maritime Small-Object Detection from Uncrewed Aerial Vehicle Imagery
Junyi Wu†, Weizheng Wang†, Liqin Tian, Qing Tian, Yong Niu, Yi Jiang.
2026 · Submitted to Remote Sensing.
Combines Hilbert-Mamba context modeling, Haar wavelet downsampling, and structured pruning across CNN, Transformer, and state-space widths for maritime small-object detection and Jetson deployment.
TRIM: Dual-Budget Structured Pruning with Information-Modulated Saliency for CNN, Transformer, and Mamba Vision Models
Weizheng Wang†, Junyi Wu†, Qing Tian, Liqin Tian.
2026 · Submitted to IEEE ICASSP 2027. Repository
Combines dependency-closed pruning units, information-modulated Taylor/Fisher saliency, coupled MILP allocation, and measured-budget repair to control parameter and computation budgets.
When Retrieval Is Not Binding: A Diagnostic Benchmark for Visual Episodic Memory
Weizheng Wang†, Junyi Wu†, Liqin Tian, Qing Tian.
2026 · Submitted to IEEE ICASSP 2027. Repository
Uses matched memory-off, wrong-entity, and shuffled-binding interventions to distinguish content benefit from episodic binding sensitivity while holding queries, candidates, features, and memory envelopes fixed.