Xinjian Zhao (赵鑫鉴)

MolEmb asks whether multimodal large language models can become the embedding layer for molecular intelligence, rather than only serving as generative scientific assistants. By conditioning molecular representations on images, SMILES, and natural-language scientific intent, it moves molecular embedding from fixed-vector encoding toward context-aware retrieval and decision support.

FORGE treats manufacturing as a hard testbed for deploying multimodal intelligence in the physical world, where small visual differences can imply costly process failures. Across workpiece verification, surface inspection, and assembly verification, it shows that current MLLMs often recognize objects but fail to bind visual evidence to manufacturing rules, tolerances, and domain constraints.

This survey frames Vision Meets Graphs as a new interface between visual intelligence and structural learning, where graph depictions become direct computational inputs rather than auxiliary figures. It reorganizes scattered progress in graph reasoning, graph learning, and scientific graph understanding into a taxonomy that clarifies how visual perception, symbolic topology, and cross-modal reasoning can be fused.

This work challenges the assumption that graph structure must be understood through graph-native message passing by testing whether structural intelligence can emerge from pixels. We show that pure vision models can compete with GNNs on graph-level benchmarks, and GraphAbstract further reveals their advantage on global structural perception and scale-invariant reasoning.

GraphOmni turns graph theory into a stress test for general reasoning agents: tasks are structurally complex, scalable in difficulty, and exactly verifiable. By decomposing LLM graph reasoning across task type, graph structure, input representation, and prompting strategy, it exposes failure modes that disappear under a single aggregate score.

DEL starts from a stronger view of graph learning: a graph should not be represented only as a fixed adjacency object, but as a system whose structural identity is revealed through its responses. By placing graphs in a physical world and sampling steady-state edge layouts under energy constraints, DEL produces distributional signatures that separate hard graph pairs and strengthen GNN/Graph Transformer backbones.

This work pushes graph SSL away from augmentation folklore by asking whether spectral augmentation helps because of spectral information, or simply because it imposes useful invariance pressure. For widely used shallow GNN encoders, we show that InfoNCE behavior is often dominated by perturbation strength rather than spectral characteristics, explaining why simple edge perturbations can match or outperform heavier spectral operations.

Protein language models can over-rely on co-evolution knowledge, causing representations to preserve misleading evolutionary correlations when transferred to downstream protein tasks. We mine negative protein pairs from disparate categories and reduce their residue-level alignment in cross-attention space, steering PLMs away from co-evolution shortcuts and toward task-relevant discrimination.

This work treats graph pooling as a structural abstraction problem rather than a node-pruning heuristic: coarsening should decide which topology deserves to survive. By injecting persistent homology into the pooling process itself, it aligns PH filtration with cutoff-style pooling so global topological evidence influences what structures are preserved.

ACGCL treats graph contrastive learning as controlled optimization over sample difficulty, rather than a recipe of random augmentations and fixed positives/negatives. It uses pair-wise augmentation to generate similarity-controlled mirror subgraphs, then applies adversarial curriculum training so optimization gradually moves toward harder and more informative samples.