Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 29 papers for 2026-09-07
Nothing cleared the bar today. Papers read in full today: 6.
The load-bearing empirical result is unusually valuable for this area: two of three expert-selected aging-related hypotheses received assay support, for TP53→BAMBI expression and RAF1→TNF expression. But THiGERLLM itself is largely a standard GraphSAGE temporal encoder plus hierarchical Transformer, LLM prefix-token fusion, label-wise ensembling and calibration; its reported macro-F1 gain over THiGER is only 51.16 versus 50.98, while THiGER remains better on weighted F1 and nDCG.
The central benchmark treats relations not observed after the cutoff as negatives despite explicitly acknowledging open-world incompleteness, uses mostly weak or dated baselines rather than modern temporal-KG/text-KG forecasters, and cannot rule out Mistral pretraining leakage; the wet-lab sample is too small and selectively curated to establish general discovery precision.
PACE introduces propagation-aware collaborative correction for one-shot personalized federated graph learning. Instead of replacing a local predictor, clients upload compact rank-r updates that correct the Local model using information from others, mitigating negative transfer and improving personalization under limited communications.
DCFA offers Dual-view Causal-inspired Attribution for failure reasoning in LLM-based multi-agent systems. It blends causal analysis with interactive traces to identify the decisive earliest action whose correction would reverse a failure, addressing shallow attribution and improving debugging of agent coordination.
TROVE introduces trace-grounded route orchestration via validation and editing, a method that revises only the parts of an agent's plan that runtime evidence invalidates rather than replanning from scratch. It uses offline distillation of evaluation data to guide runtime validation and editing, enabling adaptive skill orchestration with reduced waste and increased robustness.
Compact Bellman-Grounded Cognitive Maps (BCM) create a reusable, memory-efficient cognitive map for cost-aware navigation by grounding the map in additive heterogeneous route costs. The approach improves efficiency over state-indexed or high-rank spectral maps and supports flexible reuse for new goals without re-solving from scratch.
CABAL offers an end-to-end multi-agent simulacra framework to study the effects of collusive bidding in peer review, linking bidding, assignment, and review manipulation in a unified simulation. It enables counterfactual analyses and observation of collusive intent lifecycles, helping evaluate policy and mitigation strategies for reciprocal advantages.
We study self-supervised pretraining of molecular GNNs with LeJEPA, a predictor-free joint-embedding approach regularized by Sketched Isotropic Gaussian Regularisation (SIGReg). Evaluations on antibiotic activity and ogbg-molhiv show that pretraining improves property prediction, with robust multi-seed bootstrap results.
We develop a GNN-guided graph coarsening method with adaptive QUBO penalties for CVRPTW on a quantum annealer. By merging nearby customers into super-nodes based on time-window compatibility, the reduced QUBO is solved and expanded back, with a learned coarsening strategy that requires less tuning and improves reliability across random instances.
We address backmapping from coarse-grained beads to molecules using a free-energy-conditioned generative model across chemical space. The method learns to reconstruct chemically plausible molecules from CG beads, enabling reversible coarse-graining and more accurate high-resolution screening.
Continual Graph Memory enables adaptive recommendation under intent drift. By maintaining a memory of evolving relational evidence in a graph and updating representations over time, the system detects drift in user intent and sustains robust recommendations.
CAGE (Coherence-Aware Graph Encoding for Retrieval-Augmented Generation) reranks retrieved passages by modeling between-chunk coherence with four dimensions: intra-domain relevance, noise resistance, informational bonding, and factual consistency. It converts passages into directed heterogeneous graphs and leverages graph structure to enforce coherence and factual grounding in RAG outputs.
WEECFP provides a parameter-free 1024-d continuous fingerprint for molecules by distributing Morgan substructures across signed vector positions. WEECFP-SuRGE uses a transformer with Substructure Rotary Graph-distance Encoding, parameterized by shortest-path distances, to attend to substructure tokens. A 7-model blend achieves state-of-the-art regression ranking on TDC ADME datasets.
We propose physics-aware random walk fingerprints for scalable power-grid graph classification. By incorporating grid-physics into random-walk-based features, the approach yields interpretable fingerprints that explain propagation patterns and matches large-scale power-grid datasets with scalable performance.
Embedded Graph Flows (EGF) learn continuous embeddings for node and unordered-edge categories, removing one-hot biases. A permutation-equivariant graph transformer transports Gaussian noise toward learned category embeddings, and a final readout maps embeddings back to graphs, enabling principled categorical graph generation.
We present APT-RAG, a Tree-based Retrieval-Augmented Generation framework for evidence-intensive QA. It uses adaptive planning and topology-aware evidence gathering to overcome rigidity and topology-ignorance in multi-document reasoning, enabling effective synthesis across many sources.
Graph-complexity-based UncerTainty (GUT) quantifies and analyzes the reasoning uncertainty of LLMs by modeling the branching graph of thought processes. It characterizes divergent reasoning paths and identifies unreasonable or nonsensical branches, enabling measurement and potential calibration of LLM reasoning.
Personalized Task Dependency Graphs address signal erosion in multi-task recommender systems by allowing task dependencies to vary with item characteristics. The method uses hierarchical message passing along flexible, personalized graphs to adaptively propagate signals, reducing attenuation and improving performance on sparse, deep-funnel objectives.
We propose a method to personalize medical concept representations by refining LLMs over budgeted text-attributed graphs (TKGs). By selectively incorporating patient-specific KG context along with trajectory information, the approach accounts for context-dependent meaning and predictive value, rather than treating all concepts uniformly. Experiments on EHR prediction show improved personalization.
This paper compares counterfactual explainers for GNNs that permit adding or removing edges. It benchmarks methods on quality, efficiency, and realism of generated explanations, providing guidance on when and how to use different counterfactual graph edits.
We propose a constraint-aware conditional generative framework for synthetic origin-destination demand in hierarchical logistics networks. The model generates OD demand conditioned on topology and operational constraints, enabling scenario planning under topology changes while respecting capacity and routing constraints.
MURAL introduces Multimodal Uncertainty-aware Recommendation via Adaptive edge Learning. It addresses rigidity from static similarity graphs and semantic fragility from noisy modalities by learning adaptive edges with uncertainty modeling, improving robustness and recommendation quality across modalities.
We propose Hierarchical Possession-Aware Graph Pointer Network for pass receiver selection under partial observation. The model reasons over anonymous candidates, opponent pressure, and recent context with a hierarchical graph and pointer mechanism to predict the intended receiver.
GRACE—Graph-grounded Reflective Agent Copilot Engine—decomposes LLM outputs into atomic claims and grounds them against trusted priors in a weighted bipartite graph. Edge weights encode evidence, uncertainty, and cross-document relationships to support grounding, verification, and expert-in-the-loop refinement.
Towards AI-Driven Nanomedicine Discovery formalizes a benchmark and multimodal learning framework for predicting nano self-assembly, addressing the lack of standardized tasks and public data. It models pairwise compatibility of components and provides a unified evaluation setup to accelerate nanomedicine discovery with fewer wet-lab screenings.
This work presents a robust watermark-based fingerprint framework to verify GNN ownership, addressing common limitations such as model degradation under out-of-distribution watermark graphs and potential attacks. It offers improved resilience for ownership verification in the face of distributional shifts and adversarial modifications.
This survey reviews dynamic heterogeneous graph representation learning. It discusses challenges of modeling evolving, multi-typed nodes and edges, summarizes current methods, benchmarks, and applications, and outlines promising directions for future research in DHGs.
We propose a low-rank adapted spatially attentive graph neural network for predicting spatiotemporal PM2.5 concentrations using mobile sensing data in Surat, India. The method uses two node definitions (uniform segmentation and DBSCAN clusters) and combines spatial attention with low-rank structure to deliver accurate, interpretable predictions.
We propose a unified physics-aware quantum machine learning framework that discovers compact parametrized quantum circuits for data-scarce device modeling. A graph neural network policy optimized with PPO searches circuit architectures using leave-one-group-out cross-validation on held-out process and geometry groups, achieving lower MAE and tighter variability than classical baselines.
Corporate Language Model (CLM) presents an architecture to turn tacit enterprise knowledge into a sovereign, auditable, executable corporate intelligence layer. It integrates knowledge capture, ontological grounding, secure deployment, and auditable actuation to empower enterprise decision making with LLMs.