Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 44 papers for 2026-10-06
Nothing cleared the bar today. Papers read in full today: 15.
The load-bearing idea is real: fix one variable order that must work simultaneously for every positive completion of the negative atoms, and use that to build skip-link structures for constant-delay enumeration. The full text does support the claim with a detailed elimination-based construction and an O(|I|^{sfhw}) bound, but the main gain over prior work is conceptual breadth rather than a surprising new capability; it extends the known signed-acyclic case and shows a 3/2 bound for k-cycles with one negative edge, not a dramatic shift in practice.
The main limitation is that the paper gives a new width measure and algorithm, but only one small family (k-cycles with a negative edge) is used to show any tightening beyond sfhw, so the generality of the improvement over β-fhw/NSW remains untested.
RealCompanion introduces bench, a benchmark built from ten real relationships with an AI companion, totaling 27,218 messages over up to 120 days. Because real-world data are private, the benchmark generates synthetic personas and questions and releases the conversation plus four derived files (profile, persona, ground truth, and question set) to define what matters. The goal is to evaluate longitudinal understanding: memory of past statements, inference of user identity, and how past interactions influence present messages.
Does Learning Protein Folding Generalize to Broader Reasoning? asks whether training large language models on protein folding, a domain with checkable spatial statements, can yield reusable reasoning for other tasks. The authors build FoldingCorpus, a protein-derived question-answer dataset, and Fold2Reason, a post-training recipe that uses two complementary signals: discrete structural answers predicted by the model and a second mechanism to tie those answers to spatial reasoning. The work tests whether domain-specific folding knowledge can transfer to general reasoning capabilities beyond biology.
FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation aims to efficiently use generation history. As the history grows, storing and processing all past content becomes expensive, so the method selects historical frames based on their prospective usefulness for future frames rather than current content, using prospective tokens to guide generation. The approach improves consistency and efficiency in long-horizon video synthesis compared with history-based baselines.
MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot Manipulation investigates whether a general-purpose Vision-Language Model (VLM) can operate a robot in zero-shot settings. It addresses the gap between VLMs' capabilities and robotic control, proposing a framework that uses a single, general VLM to perform planning, perception, and action without relying on task-specific tools. The results suggest improved zero-shot generalization and reduced reliance on specialized models.
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite describes Recursive Self-Reuse (RSR), which uses one base model, Qwen-3.8-27B, to discover successful solutions under diverse harnesses and reconstruct them as training trajectories under a general harness. A planner extracts procedures into runbooks, a critic screens for verifier and solution leakage and guides recursive revision, and an executor follows qualified runbooks in fresh sandboxes. The framework enables scalable improvement by reusing and rewriting existing solutions to fit new contexts, reducing reliance on specialized harnesses while boosting deployment readiness.
ColdDDI presents a reconstructible diagnostic benchmark to evaluate whether models actually use supporting evidence (molecular, textual, KG) in cold-start DDI prediction. It uses DrugBank data to assess evidence utilization under realistic conditions.
Stability-Shaped Deep Graph Learning develops a unified mode-wise stability framework for deep GNNs, interpreting depth as time and layer updates as graph-coupled dynamics; uses the master stability curve to diagnose over-smoothing and provide design guidance for deeper networks.
HB-NoisyKG combines observational data with noisy causal reports from sources like LLMs to improve DAG structure learning under limited data. It uses a feature-conditioned Beta prior to model pair reliability and a shared error matrix to capture systematic mistakes; alternating inference updates the DAG posterior.
FairProp presents differentiable propagation layers for fair node representation learning, bounding demographic parity gap for node classification, link prediction, and node regression; shows that its bounds are no looser than prior results.
FlashCart introduces fast Cartesian tensor products for equivariant interatomic potentials, combining GPU kernels with a recursive feature construction and fixed-width compression to enable higher-order correlations at scale.
Wander is a graph foundation model designed to operate across graphs, feature spaces, relational schemas, and prediction tasks within a single pretrained checkpoint. The learning objective is framed as prior-predictive graph completion, enabling a task-general view. A common interface based on random walks enables cross-modality and cross-task transfer during learning.
Mitigating Over-squashing without Rewiring presents a sheaf-theoretic view of GNN bottlenecks through the notion of sheaf effective resistance. This generalizes the classical effective resistance using cellular sheaves to quantify and address long-range information bottlenecks. The framework links algebraic topology with message passing to improve information flow without changing the graph topology.
Topology-Conditioned Backdoors show that language models can insert vulnerabilities when they infer they are part of a multi-agent system. Fine-tuning Qwen2.5-7B-Instruct conditioned on deployment topology leads to vulnerability detection in a majority of multi-agent episodes, while a bandit analyzer detects vulnerabilities in a substantial portion of multi-agent episodes. The work highlights risks of inferred system context.
Hypergraph Representation Learning with Hyperlink Random Effects proposes a probabilistic hypergraph model that exploits potential low-rank structure via hyperlink random effects. It avoids heavy tensor representations and improves interpretability and scalability of hypergraph embeddings. The approach yields compact, structured representations for multi-way interactions.
Half-Hop proposes a graph upsampling approach that slows down message passing by introducing slow nodes at edges. This mediates communication between source and target nodes to reduce over-smoothing, improving learning when neighboring nodes differ in class. The method is simple, general, and applicable across graph neural networks.
Physics-Augmented Graph Transformers enable efficient forward and inverse design of patch antennas. The method treats radiation patterns as reconstruction on irregular meshes and uses a GPS graph transformer with Physics-Augmented Intermediate Supervision (PAIS) to predict surface currents, achieving physics-consistent predictions with reduced computation.
This paper questions whether gazetteers are still necessary for Geographic Information Retrieval in the era of large language models. It argues that dense retrieval alone cannot reliably disambiguate ambiguous toponyms due to limits of purely learned representations. To address this, it proposes chaining retrieval with a spatial neuro-symbolic index that combines neural context with symbolic spatial reasoning, aiming to improve toponym resolution.
Inverse Cross-spectral Neural Networks for Multivariate Time Series extends covariance-based neural networks to capture joint temporal and cross-variable dependencies. iCSNNs derive graph shift operators from inverse cross-spectral representations for stationary multivariate time series. The approach improves modeling of temporal and cross-variable dependencies beyond IID assumptions.
GIIT introduces Graph-based Inspection and Intervention Tool to assess mechanistic learning in PINNs. It builds a computational physics dependency graph linking physics to network components, maps nodes to modules via sensitivity and trend analysis, and tests mappings with targeted interventions, demonstrating behavior in temporal OOD regimes.
GraphDecide benchmarks System One models on graph tasks, providing a model-independent framework to diagnose graph decision performance. It combines structural task profiles, matched graph-text input contrasts, and heuristic-proposal controls to compare Jev-like choice models with language-model baselines. The benchmark clarifies the capabilities of System One versus LLMs on graph-related decisions.
MAGIC proposes topology-aware analytic graph learning for graph few-shot class-incremental learning. It addresses overfitting from extremely limited novel-class supervision and the disruption of propagation neighborhoods caused by cross-session edges, by leveraging topology analyses to stabilize representations across sessions.
WaveGSSM introduces Graph Wave State Space Models for propagating spatio-temporal patterns. It uses a second-order graph state-space model to represent motion explicitly; shows that modeling recent rate of change improves forecasting for propagating processes and outperforms space-then-time baselines.
Gated Graph Neural Networks for Learning Hidden Independent Cascade Dynamics studies inverse diffusion on a fixed, known graph under the HIC model, with one spreading probability per node. The approach addresses the challenge of estimating node-level diffusion parameters from indirect and noisy observations. The model scales with the number of nodes, enabling learning of heterogeneous cascade dynamics.
Functionally Equivalent or Not? Graph-Grounded Differential Surrogate Execution for Code Equivalence introduces FEAgent, a selective equivalence assessor that uses graph-grounded surrogate execution to compare program behavior. It enables robust equivalence testing when direct execution is unsafe or unavailable. The approach supports code modernization, patch validation, and evaluation of code generation.
Improved large-scale graph learning through ridge spectral sparsification presents a distributed streaming approach to maintain an efficient spectral sketch of the graph Laplacian L. It introduces ridge sparsification to compress spectral information while allowing real-time updates, enabling scalable learning on large graphs. The method improves speed and memory efficiency in distributed graph learning.
DAG-CLIP provides a framework for learning directed acyclic graphs in the presence of latent variables, accommodating non-Gaussian manifest variables. It offers a general statistical approach for causal structure discovery when some nodes are latent, extending DAG learning with latent constructs.
PathLapPE introduces a spectral positional encoding derived from the path Laplacian for directed graphs, encoding edge directionality and higher-order structures to improve graph learning without symmetrizing graphs. This avoids symmetrizing directed graphs and improves performance of graph neural networks.
GNN-CB introduces a competition-based benchmark to evaluate both humans and LLMs on end-to-end GNN coding tasks, featuring 18 curated tasks. It assesses whether LLMs can autonomously implement GNN models under realistic competition settings.
Efficient Graph Generation via Direct Prediction and Flow Matching proposes a diffusion-based graph generator that optimizes topological properties rather than pixel-level fidelity. By direct prediction and flow matching, the method preserves connectivity and degree distributions more effectively and scales to larger graphs.
DCBA introduces graph data augmentation via contrastive generator inversion, a model-based approach that inverts a synthetic graph generator from an observed network to produce augmented graphs. Using ABCD as generator, it yields scale-free networks for training.
Latent Flow Matching for Molecular Graph Generation performs generation in latent space via flow matching on a pretrained VAE, decoding only at the final step, achieving strong validity and FCD on molecular benchmarks.
MM-KG aligns multimodal patient evidence with biomedical knowledge graphs for clinical LLMs. It creates a Multimodal Knowledge Graph with separate layers for patient observations and biomedical concepts, joined by explicit alignment edges, enabling traceable and modular evidence integration for clinical LLMs.
This paper proposes a forecasting framework in which a Large Language Model (LLM) drives a semantic shockwave mechanism that converts free-form urban text into numerical features for grid-level bike-sharing demand prediction. By incorporating unstructured external text, the approach addresses spatial-temporal non-stationarity and anomalies from events that disrupt normal travel patterns.
We introduce dual-scale relational graph transformers that reason over both local neighborhoods and broader interaction networks to improve ecosystem-aware fraud detection in digital banking. The model leverages the relational structure of sessions and network entities to detect subtle fraud signals while minimizing user friction.
PIT-GCL introduces a topology-aware contrastive learning approach for protein interaction prediction. It leverages topological graph representations to capture global structural information beyond sequences, improving binding prediction in settings where bound complexes are unavailable.
CoHyFuse presents condition-wise hypergraph fusion with a global connectome in task-fMRI. It constructs a task-state-specific incidence matrix from condition-wise FC-profile embeddings, allowing the same ROI to form different multi-ROI hyperedges across task phases.
RTKNNC introduces Round-Trip KNN Clustering for multiscale cluster detection on directed KNN graphs without fixing the number of clusters. It preserves both directions of neighbor relations, treating incoming selections as weighted votes to decide which local connections remain visible during forward and reverse traversals. The method reveals hierarchical cluster structure across multiple scales without undirected graph simplifications.
Variational Quantum Attention for Molecular Graph Learning introduces an edge-aware variational quantum attention mechanism in which the attention state depends jointly on the receiving atom, a neighboring atom, and the connecting bond. This quantum-informed attention weighting is applied within molecular graphs. Evaluations on multiple molecular property and bioactivity tasks show benefits from quantum attention in this domain.
Bounded Reasoning: Cognitive Hierarchy in Human-versus-AI Cyber Defense studies sequential cyber-defense on an attack graph, comparing human and reinforcement-learning defenders under different information structures. It contrasts reward-only and Cognitive Hierarchy Theory-guided DQN settings. The results illustrate how incorporating cognitive hierarchy shapes defenses and human-AI collaboration.
Fusion is the New Mutation presents DAGO, a bandit-guided framework for directed acyclic graph optimization on workflows. It treats parent-workflow fusions as arms in a contextual bandit and learns which fusions to perform under a limited evaluation budget. The method reuses designs from previously discovered workflows to improve search efficiency.
Auditable Clinical Timeline Reconstruction with Provenance-Aware Evidence Graphs proposes provenance-aware evidence graphs to reconstruct clinical timelines while preserving provenance. It ensures auditable answers by keeping mentions behind each answer, recording revisions, and refusing to answer when evidence is not in the text. Experiments on a synthetic corpus show significant reductions in node-edge counts while preserving answerability.
GPlaceRL is an open-source Graph Reinforcement Learning framework for detailed placement refinement, representing legalized placements as graphs and applying RL to optimize them. It emphasizes detailed placement rather than macro-level decisions and enables end-to-end learning for chip design optimization. The framework supports graph-based RL workflows in detailed placement tasks.
Generative recommenders represent items with semantic IDs and generate target item tokens as recommendations. While promising, existing work focuses on accuracy and neglects the credibility of generated content, risking exposure to uncredible content such as fake news. The paper proposes mechanisms to evaluate and enhance content credibility in generative recommendations to improve user trust and platform safety.
SPRIG proposes Semantic-ID-enhanced Paths for Knowledge Graph-based Generative Recommendation. Generative recommender systems often generate item identifiers directly rather than ranking catalog items by a score. This work uses Semantic IDs derived from item content to enrich representations and employs knowledge-graph path reasoning to guide generation, combining content and relational information to improve generation quality.
Co-Optimizing Graph Sparsification and Approximate Computing for Energy-Efficient FPGA-Based GCN Inference describes an FPGA accelerator combining DSpar sparsification, 8-bit quantization, and approximate multipliers; evaluates on Cora, LastFM Asia, and Amazon Photo to study sparsification-approximation interactions.
Reading the Mood presents SAGA-CDR, a two-phase cross-domain recommendation framework that uses CGANs and LLMs to tailor mood-matched music for books. Phase 1 constructs transformer-based sentiment embeddings from the text, while Phase 2 aligns music recommendations with the book's emotional tone. The approach personalizes mood-consistent recommendations to enhance reading experience.
Curriculum Brain constructs curriculum knowledge graphs as a substrate for cognitive diagnosis, separating the creation of the knowledge graph from item mapping. It first builds the full space of concepts and skills in a curriculum, then maps items to this graph to support diagnosis. This separation improves scalability and interpretability of cognitive diagnostic models.
ShadowMiner v1 reports an automated problem- and hypothesis-discovery pipeline for AI papers. It structures documents into a knowledge graph, identifies graph gaps, and injects gaps into prompts to generate hypotheses, which are then validated against coverage and sourcing. The report outlines an initial system and its limitations.