Understanding hypergraph evolution is essential for revealing high-order interaction patterns and generating credible synthetic data when real interaction records are scarce. Existing models suffer from two key limitations: (1) they rely on global topological heuristics that treat nodes as passive entities, yielding poor semantic consistency and generalization; (2) they ignore the influence of node attributes on structural evolution. We propose HyMAGE, a semantic-aware dynamic hypergraph generation framework based on semantic preferential attachment, without any graph-structure pretraining or centralized optimization objective. HyMAGE models each node as an autonomous agent and leverages LLMs for local-level semantic selection, so that hyperedge formation and dissolution emerge from local semantic affinity and structural context in a self-organizing manner. It serves both as a generative model explaining real-world high-order relationship evolution and as a scalable synthetic data factory that distills LLM domain knowledge into explicit hyperedge structures, producing topology-and-semantics-aligned attributed hypergraphs for downstream tasks. Extensive experiments show that HyMAGE significantly outperforms existing methods at both structural and semantic levels. It simultaneously reproduces nine high-order structural patterns of real hypergraphs and generalizes well to downstream tasks: hypergraph neural networks trained solely on HyMAGE-generated data achieve high accuracy, and its diffusion behaviors closely match those of real hypergraphs. These results demonstrate that HyMAGE offers a plausible explanation for high-order evolution mechanisms while providing rich semantic hypergraph training sets for hypergraph learning and mining.
B. Gu, Ji Zeng, Nuoran Zhou et al.· Proceedings of the 32nd ACM...· 0 citations
Auto Research uses language-model agents to propose, implement, and evaluate machine-learning changes in a closed loop, but is usually judged by its terminal pipeline. A terminal score cannot reveal which technical decision produced a gain or distinguish a reusable discovery from a change adapted to development feedback. We introduce intervention-centered Auto Research, which validates research decisions rather than only final artifacts and makes their reliability measurable. Feature, Model, Representation, and Data axes are searched independently with inner five-fold feedback. Each axis winner is frozen before an outer-holdout matrix compares all alternatives on evidence the loop never sees. Across 701 agent-executed attempts spanning ten Matbench endpoints, outer evidence confirms the selected intervention on nine of ten endpoints and preserves 89.3\% of non-tied intervention orderings. It also rejects an aggregate Representation gain that inner feedback endorsed. The resulting matrix reveals an information-dependent hierarchy. Composition-only tasks support several routes to improvement, whereas structure-informed tasks favor local geometry features and complementary tree ensembles. A subsequent compatibility test combines already frozen Feature and Model code without further search or tuning and raises mean outer-holdout improvement from 19.0\% to 26.3\%. By validating decisions rather than only artifacts, this design turns adaptive search into reusable evidence wherever agents propose executable alternatives against a fixed evaluator.
Jingjie Ning, Xiaochuan Li, Shan Zhong et al.· 0 citations