Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 18 papers for 2026-09-28
Nothing cleared the bar today. Papers read in full today: 9.
The load-bearing idea is real: replace DGN’s explicit coarse-step update with a semi-implicit Newmark-like nodal solve, and add an operator-weighted virtual hub to approximate the dense global coupling that an implicit mechanical solve would induce. The experiments do test this claim reasonably well: on the beam, the method is dramatically more stable than DGN, MGN, EGNO, EGHN, and even IGNS at coarse steps; the staged ablation shows the hub is the biggest gain, and the inferred forces/operators are compared against independent biomechanical moments and finite-element stiffness structure. However, the paper also admits a limit: on the longest beams, plain DGN can slightly beat the hub approximation, so the “rank-one global coupling” story is not universally dominant.
The main limitation is that the architecture’s best argument—the hub as a low-rank proxy for implicit global coupling—breaks down in the longest-beam regime, and several headline evaluations are single-seed rather than robust multi-seed studies.
This work studies implicit graph neural networks (IGNNs) where node representations are fixed points of message-passing operators, enabling effectively infinite-depth propagation and test-time flexibility. The authors address the requirement of a unique, attainable equilibrium and propose implicit neural sheaves to achieve convergence without restricting the diffusion updates, aiming to enhance expressiveness at equilibrium.
This study evaluates advection-aware graph neural networks for short-term forecasting of cloud-induced power ramps across a distributed sensor network. Using a controlled synthetic testbed with a known wind field and a self-supervised cloud-motion estimator to set edge time-lags, results show that realistic CMV estimates enable advection-aware propagation to improve ramp forecasts, while poorer estimates provide limited benefits.
Acacia is a graph foundation model trained on the web graph that supports arbitrary feature dimensionalities and semantics without extra training, handles diverse tasks with in-context learning, and does not rely on pretrained LLMs or extra heads.
Moment-guided edge sampling proposes a framework to quantify the effect of local edge edits on global graph structure through spectral moment changes of the random-walk transition matrix. It provides exact moment-change computations via a combinatorial, closed-form approach for edge additions and removals, enabling principled sparsification decisions.
Scaffold presents a topology-based, unsupervised sparsification framework for GNNs derived from support graph theory preconditioners. It explicitly controls dilation and related structural quantities to reduce edges while preserving essential communication patterns.
HySTAR introduces anchored hypergraphs for stable credit assignment in cooperative multi-agent reinforcement learning. It builds on a MAPPO-like framework and separates adaptive representations from topology, mitigating structural target drift as agents and coalitions evolve.
MARCEDES advances score-based causal discovery under non-Gaussian errors by introducing mean absolute residual risk and showing that, asymptotically, the true weighted DAG minimizes this risk among feasible matrices, with continuous optimization for high-dimensional settings.
Learning What to Skip introduces counterfactual credit assignment for efficient multi-agent LLM workflows by learning which steps to omit. Using full-workflow logs and controlled skip interventions, it learns to skip future steps that do not contribute to the final reward.
Externalized CPDAG Summaries Improve LLM Causal Deduction introduces Structured Thinking, a two-turn pipeline that externalizes a typed CPDAG summary before reasoning, improving causal deduction performance on Corr2Cause benchmarks.
The paper argues that sensor expansion changes evidence about a process rather than the underlying dynamics and introduces STFO (Spatio-Temporal Field Operator) to decouple dynamic learning from sensor-layout changes, promoting robust continual spatio-temporal forecasting as sensors grow.
FTB Graph analyzes first-token broadcasters and language-identity head circuits in multilingual LMs using Edge Attribution Patching and activation-patching verification across six model families, revealing causal circuits that determine the target language early in generation.
Adaptive Interaction Graphs for Particle Simulation introduces a dynamic interaction graph where each particle carries a variance head that predicts local uncertainty. High-uncertainty particles receive expanded neighborhoods, and the system is trained jointly with an acceleration head under a heteroscedastic Gaussian NLL loss to improve long-horizon trajectory accuracy.
Multivariate conformal uncertainty propagation in multitask atomistic simulation surveys the use of conformal methods to recalibrate surrogate predictions across scales and multiple targets, highlighting successes and pitfalls in multiscale uncertainty quantification.
BeatGraph presents heartbeat-centric self-supervised graphs for infant ECG representations in home environments. By modeling heartbeats directly rather than fixed patches, BeatGraph better aligns representations with neonatal cardiac structure and rate variability.
Do LLMs Understand Context? A Knowledge Graph-Based Evaluation Framework argues that true contextual understanding should be tested against a knowledge-graph-informed framework rather than pattern-matching prompts alone.
DynBranch tackles the branch-resolution barrier in dynamic agentic LLM serving by making unresolved branches addressable early. It enables speculative subgraph execution during resolution and caches results to reduce serialization latency.
Predicting Transmembrane Protein Topology from 3D Structure demonstrates that a SchNet-based GNN can infer membrane topology using all-atom embeddings rather than only C-alpha features, achieving promising results without pretrained weights on the same dataset as DeepTMHMM.