Xinjian Zhao (赵鑫鉴)

We introduce FORGE, a fine-grained multimodal benchmark for manufacturing that combines real-world 2D images and 3D point clouds with domain-specific semantic annotations. Evaluating 18 state-of-the-art MLLMs across workpiece verification, surface inspection, and assembly verification reveals that domain knowledge, rather than visual grounding alone, is the key bottleneck.

This survey provides a first systematic consolidation of the emerging 'When Vision Meets Graphs' area, where graph depictions are treated as first-class inputs. Using the RPI lens and a three-thread taxonomy, it organizes fragmented literature and outlines concrete future directions toward graph foundation models.

For the first time, we conduct a controlled comparison between pure vision backbones and GNNs for graph structural understanding, and introduce GraphAbstract to isolate topology reasoning. Pure vision models achieve comparable overall performance and outperform GNNs on holistic structure tasks with stronger cross-scale generalization.

We introduce GraphOmni, a modular benchmark that factorizes graph-theoretic reasoning by graph type, serialization format, and prompting strategy. Large-scale evaluation reveals strong interactions among these factors and substantial capability gaps that are hidden by single-score comparisons.

We propose Distributional Edge Layouts (DEL), modeling graph topology as a distribution by globally sampling edge layouts with Langevin dynamics under Boltzmann energy. As a model-agnostic preprocessing module, DEL consistently improves diverse GNN backbones and reaches strong state-of-the-art performance on multiple datasets.

We re-examine spectral augmentation in contrastive graph self-supervised learning via broad empirical studies and theoretical analysis of InfoNCE bounds. Simple edge perturbations (edge dropping for node-level, edge adding for graph-level) often match or outperform spectral methods with much lower computational cost.

We introduce negative sample mining for protein language models by constructing hard negatives from protein pairs across distinct categories. The approach improves downstream performance and shows better alignment with biological mechanisms such as protein-protein interaction.

We propose a graph pooling mechanism that injects persistent homology directly into coarsening, motivated by the alignment between PH filtration and cutoff-style pooling. The method is plug-and-play across pooling families and yields consistent, substantial gains on standard benchmarks.

We propose ACGCL, combining pair-wise augmentation with controllable similarity and subgraph contrastive learning for graph-level representation learning. An adversarial curriculum progressively increases sample difficulty and focuses optimization on harder examples, outperforming strong baselines on six benchmarks.