Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 14 papers for 2026-09-21
Nothing cleared the bar today. Papers read in full today: 3.
The load-bearing idea is a minimax conversion from adaptive width to cheap fractional balanced separators, followed by Korchemna et al.’s FOCS 2024 separator-rounding theorem and a recursive decomposition argument. The paper actually proves the quantitative bridge and derives concrete consequences: under ETH, FPT iff PTIME iff bounded fhw for bounded-lambda pattern classes, plus a decision-to-exact-counting collapse and an FPT-to-quasipolynomial-time implication for arbitrary recursively enumerable pattern classes. This is a substantial structural result in CSP/query complexity, not a benchmark-driven incremental method.
Its strongest consequences are conditional on ETH and concern worst-case homomorphism/CQ complexity; the paper does not establish practical algorithms or direct consequences for modern graph-learning workloads.
OpenMAS-GCom provides a diagnostic benchmark for graph-enhanced multi-agent systems to attribute performance to communication graphs, roles, and information flows. It standardizes evaluation to separate effects of models, communication patterns, and computation costs.
This work analyzes attention-graph topology to detect LLM hallucinations, using Forman-Ricci curvature to identify bottlenecks in attention flow. It develops semi-local and global flow features of heads associated with hallucinations and evaluates across models and benchmarks.
VISPATH introduces Visual-Intent-Guided Path Reasoning to maintain multimodal cues throughout multi-hop KG reasoning for KGQA. It uses visual intent to guide evidence gathering and reasoning paths, improving multimodal KGQA performance.
EnSol proposes an environment-aware GNN for molecular solubility prediction. It goes beyond fixed-solvent assumptions by modeling solute–solvent interactions and continuous temperature effects through environment-aware representations, enabling more accurate solubility predictions across solutes, solvents, and temperatures.
MACE introduces Memory-Agent Co-Evolution with adaptive memory graphs to improve how multi-agent systems reuse procedural traces. Grouping dependencies into functional memory units enhances retention, while connecting units improves retrieval; the best configuration changes depending on whether the task uses instructions or checklists.
GraphSkillEvo proposes evolutionary optimization of graph-structured agent skills to provide explicit workflow guidance. By reducing redundancy and constraining skills into executable graphs, it improves the practicality and performance of LLM-guided agents.
M2G-LLM introduces a multimodal framework that augments LLMs with graph neural networks to integrate clinical text, labs, and imaging. By aligning modalities and reasoning on a graph, it enhances clinical prediction beyond text-only LLM capabilities.
KG-Chronos-2 couples a time-series foundation model (Chronos-2) with hydraulic knowledge to improve surrogate forecasting of water-surface elevation. It uses exact-state residual decoding, graph-conditioned historical retrieval, and input-aligned correction to outperform baselines.
PCC+GCN combines Particle Competition and Cooperation as a pre-training label-refinement stage to detect and reassign suspicious labels before GCN training, improving robustness to label noise in graphs.
HERMES is a graph-based framework that performs knowledge-graph reasoning directly on clinical text to predict patient outcomes, preserving relational and temporal structure in notes. It leverages contrast-aware reasoning to improve predictive performance.
The paper presents a co-evolving framework for zero-day jamming: online detection using graph attention and adaptive attack synthesis that evolves with detection, enabling robust evaluation against unseen jamming strategies.
kgsteward is a tool for building, reproducing, and maintaining distributed knowledge graphs in life sciences. It supports private consortium data integration with public references via RDF, with provenance and privacy controls to keep data private during projects.
The paper proposes an end-to-end reconciliation pipeline to harvest validated links between scholarly papers and their source code from software journals and repositories, modeling publication-to-repository pairs in a knowledge graph for enhanced discovery and reproducibility.