Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 19 papers for 2026-10-05
Nothing cleared the bar today. Papers read in full today: 7.
The central idea is load-bearing: four axioms (predictability, tightness, strictly k-range, topology-invariance) plus TRIP/GRIP as a construction that provably satisfies them, so the benchmark question becomes mathematically checkable rather than assumed. The full text does support the claim better than most benchmarking papers: it audits LRGB/GLoRa/ECHO, exhibits the averaging artifact of the common range metric, and shows on TRIP that balanced-Forman curvature is not predictive while ECHO rankings are partly driven by factors beyond pure long-rangedness.
The main limitation is that TRIP/GRIP are still synthetic tasks built from i.i.d. stable features and engineered range targets, so the framework certifies long-rangedness but does not by itself show what kinds of real-world graph phenomena it faithfully preserves.
The paper’s central mechanism is real: with incidence-matrix coordinates and tree blocks, the block subproblem can indeed be solved by recursive leaf-to-root messages, and the theorem gives an explicit linear rate that depends on the chosen partition. But the full text mostly formalizes and packages a known tree-elimination idea into decentralized optimization language; it does not demonstrate a qualitatively new class of problems or a broadly superior method beyond the graph topologies where tree decomposition is natural. The experiments are described only broadly here, so the evidence looks more like a principled reparameterization plus rate analysis than a field-changing advance.
The main limitation is that the gain is tied to graph structures and partitions that make tree tearing natural, so the claim is narrower than the abstract’s broad framing of a general framework.
The paper’s load-bearing idea is real: many plausibility benchmarks are invalid because the trained GNN may exploit shortcuts, so the authors compile a known logic formula into a GNN and then evaluate explainers against exact prime-implicant ground truth. The full text supports that claim with concrete shortcut audits on common datasets and with a benchmark where several explainers behave very differently on alternative implementations and multi-hop influence tasks. That said, the contribution is primarily a diagnostic benchmark and methodology paper, not a new explainer or a general solution to graph XAI; the gains are about making evaluation sound, not making explanations better.
Its scope is deliberately narrow: binary classification with categorical features and compiled, noiseless GNNs, so it is a minimum bar for evaluation rather than evidence about real trained models in the wild.
We analyze asymmetric margin supervision for instruction-guided LLM generative recommendation. History events do not always support the target item; we show how to design supervision that accounts for asymmetric influence of history on recommendations.
We formalize a vulnerability surface for GraphRAG pipelines with LLMs, revealing schema-level attacks. The Hop-Decayed Influence (HDI) attack and the 3S framework exploit Semantics, Structure, and Scoring to manipulate retrieval.
Proposes a hybrid generative framework for complex networks that combines graphon-based models with neural inverse operators. The goal is to retain interpretability from mechanistic models while enabling amortized, scalable generation beyond training-size, bridging two schools of graph modeling.
We show that graphifying multimodal data is not universally beneficial. Across diverse datasets and graph constructors, inserting relational graphs can hurt performance or be neutral, motivating a principle that graphification should be used judiciously (No-Free-Graph).
We examine whether high-level safe plans reflect actual execution traces in RoboGuard. By comparing safety verdicts on surface plans versus graph-refined traces under the same LTL specification across 28 cases, we reveal a refinement gap where execution differs from plan-only safety assumptions.
We address non-stationary distribution shifts in graph learning by crystallizing memory rather than generating new memory graphs. Our Efficient Memory Crystallization method reduces computational overhead while maintaining performance under drift.
We introduce MIRROR, a multipath quorum integrity mechanism for LLM multi-agent communication, defending against Agent-in-the-Middle attacks by cross-checking messages across multiple independent paths.
We introduce GraphBind, a topology-driven binding approach for Omni Multimodal Graph Foundation Models. It handles incomplete node attributes and uses topology to guide multimodal binding, unifying representations across heterogeneous node modalities.
We present HyperFuse, a fast self-supervised node embedding method for attributed hypergraphs. By avoiding heavy deep encoders and leveraging a spectral relaxation of hypergraph modularity to compute structural coordinates, HyperFuse enables scalable embeddings for evolving hypergraphs.
We compare electronic density versus geometry as inputs for machine-learned predictions of molecular absorption spectra, highlighting how representation choices affect predictive accuracy and computational efficiency.
We present A Residual Tree Gaussian Process Modeling Framework (ResTGP) for large spatial data with heterogeneous structures. By decomposing a Gaussian process via residual trees, it enables scalable, flexible modeling in high-dimensional domains.
We present a no-label feature-centric graph data augmentation method that does not rely on labels, leveraging latent structure to augment graphs and improve downstream performance. The approach is scalable and naturally inductive, applying across diverse graphs.
We propose Multimodal Hypergraph Flow Matching (MHFM) to generate joint structural and functional brain connectivity data by modeling higher-order, multimodal interactions with hypergraphs. This preserves cross-modal structure and relationships beyond pairwise graphs, improving neuroimaging data generation.
We propose Ansatz, a continual graph memory system for mathematical research agents to organize and reuse vast intermediate proofs and results across long-horizon explorations, enabling scalable collaborative reasoning.
We introduce DiffGCMS, a spectrum-conditioned discrete diffusion model for de novo molecular structure elucidation from GC–EI–MS, and fuse it with an LLM-based reranker to provide explainable structure predictions.
We propose RaBitQ-SSD, extending IVF-RaBitQ for SSD-resident vector search by splitting codes and pipelining I/O to reduce reads and better hide I/O latency, enabling efficient SSD-backed vector search for very large collections.