Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 28 papers for 2026-09-25
Nothing cleared the bar today. Papers read in full today: 8.
The core idea is that instead of assigning one coarse trajectory reward to all steps, merge rollouts into a trajectory graph, estimate node values with Bellman backups, and use TD residuals/Graph GAE for step credit. The paper does support this with strong ablations showing that the graph-backed estimator plus the step-level objective beats GRPO and also outperforms GiGPO/GraphGPO on ALFWorld, WebShop, and SearchQA, but the contribution is still an improvement to credit assignment within an existing agentic-RL recipe, not a new paradigm.
The strongest limitation is that the gains depend on state canonicalization and sparse, task-specific rollout graphs, so the method’s clean RL interpretation may be less stable when semantic matching is noisy or the graph is too sparse to approximate a meaningful value function.
SpeakerMem-R1 introduces a speaker-centered dual-track memory for multi-party dialogue. It aims to preserve who said what, who each statement concerns, and how relationships and states evolve over time, addressing bottlenecks in attribution and cross-member reasoning. By maintaining parallel memory streams for individual speakers and for the group, it facilitates attributing messages, tracking social relations, and integrating clues distributed across members, groups, and time.
Spatial-Interactor presents a learning framework for spatial reasoning in vision-language models by enabling interaction with the observable physical world. It focuses on dynamic state transitions caused by object motion and viewpoint changes, and on integrating these transitions along long trajectories to keep an updated spatial state. Unlike static question-only training, it uses interaction trajectories as supervision to teach models how space evolves over time.
HappyWorld-Bench introduces a comprehensive benchmark to evaluate world models by testing their reliability as agents explore, interact with, and modify generated worlds. It defines a hierarchical capability framework with six world capabilities (W1-W6) and evaluates across three independent tracks: video world models, spatial world models, and embodied world models. The benchmark emphasizes both generation quality and consistency under exploration and interaction.
The Past Frames the Future: Memory for Autoregressive Video Generation discusses memory challenges in autoregressive video generation under bounded context, storage, and compute limits. As sequences grow, critical historical information—such as entity identities, dynamic states, and causal changes—tends to fall out of the active context, hindering long-horizon generation. The work advocates memory-based solutions to retain and retrieve important past facts to support coherent, long-range video generation.
Just-in-Time Memory argues that existing write-time curation is inflexible, producing task-agnostic summaries before future queries are known. It proposes learning a task-adaptive memory curator that selects what experiences to store during task execution, guided by anticipated downstream tasks. This approach aims to improve retrieval relevance and efficiency by curating memories that matter for future use.
EvLink is an evidence-linking retriever for GraphRAG that preserves passages as retrievable evidence units and builds evidence-supported transitions between them. It creates two reliable link types: relation-grounded evidence links and cross-passage transitions that justify reasoning steps, strengthening the evidential basis of multi-hop reasoning.
We introduce BLADE, a distilled LLM regularization for calibrated knowledge graph completion. BLADE uses a variational model that separates latent truth from graph recording and distills offline LLM judgments into a frozen teacher regularizer; the LLM is unused at inference time, and posterior samples provide calibrated probabilities and epistemic uncertainty. Across five benchmarks, BLADE remains competitive while offering reliable uncertainty estimates.
We propose Direct Message Approximation (DMA), a consistency-based framework for tractable approximate inference on factor graphs. DMA directly approximates factor-to-variable messages rather than the marginal, for normalizable factors, avoiding problematic message iterations and Dirac-delta collapses.
We present TopU-LBVS, a realistic multi-target benchmark for ligand-based virtual screening under hard-negative screening conditions. Built from curated ChEMBL data, it covers 93 targets across seven protein classes and constructs target-specific screening libraries with property-matched decoys. The benchmark addresses biases in random negatives and easy decoys to enable robust evaluation.
We introduce the Graph Dynamics Model (GDM), a world model for graph-structured observations that can handle evolving topologies in stochastic, partially observable environments. It uses a sparse recurrent adjacency matrix to model topology updates and learn transitions over dynamic graphs.
We propose HermNet, a spectral graph neural network built from Hermite polynomials that combines a nodewise predictor with normalized Hermite propagation. Its sparse recurrence avoids eigendecomposition or learned bases, and the paper discusses optional coordinate calibration, response normalization, and Gaussian-derivative regularization.
We extend the neural network verification (NNV) framework to graph-structured inputs via GraphStar sets, a generalization of Star sets that captures uncertain graph states. This reachability-based formal verification allows verifying GNNs deployed on power-system tasks such as PF, OPF, and CFA.
We introduce GridSFM, a foundation model for solving AC OPF, built as a physics-inspired 15M-parameter graph neural network pretrained across 54 topologies (500 to 4,000 buses) and fine-tuned with physics-informed objectives. The model achieves a 2.45% zero-shot generation-cost error on a 10k-bus case with no size-dependent degradation, and demonstrates scalable physics-informed performance.
Augur rehearses reactions to product and policy changes offline by building a typed knowledge graph from change documents, populating a grounded persona market, and simulating interactions. It then returns an auditable decision memo that recommends one of five actions. The method is evaluated on Gold-50, a set of fifty real product/policy episodes with known outcomes, and the five-action verdicts are scored against public records.
We present BridgeMem, which estimates dyadic transition evidence as a residual added to the log scores of a frozen full-vocabulary forecaster for temporal knowledge graph forecasting. The residual captures how prior relations between the query actor and a candidate change the odds of the target relation.
We introduce ICE: Task-Aligned Clifford Latent Fields for Multimodal Graph Foundation Models. The approach uses Clifford algebra-inspired latent fields to preserve entity semantics, construct interaction states from graph neighborhoods, and expose geometry-aware representations to different prediction units. An empirical study shows that preserving semantics, interaction state, and geometry are inseparable for high-quality multimodal graph foundations.
We propose graph-based inference and topology-aware multi-agent reinforcement learning for large-scale railway network management. The approach uses graph representations to model spatial deterioration and coordination, enabling scalable planning and maintenance decision-making. It demonstrates improved coordination and efficiency over traditional centralized or decentralized methods.
Budget-constrained graph augmentation studies how to strengthen network robustness by adding new links and distributing weights under limited resources, using Kirchhoff index (total effective resistance) as the objective. The problem couples discrete edge selection with continuous weight allocation under heterogeneous per-unit costs and a total budget, typically with an exact-cardinality constraint, making it a challenging combinatorial optimization task.
We study learning and interpreting policies for simultaneous entanglement requests in quantum networks. The work focuses on scheduling link-level entanglement resources and producing multipartite entanglement across a network while minimizing resources and latency. We present methods to learn and interpret these policies.
WST-Graph reworks the front-end of wavelet scattering for speech deepfake detection to preserve the topology of scattering paths. It reconstructs paths as a sparse modulation-carrier grid for an AASIST graph backend, enabling path-aware representations. Modulation-level normalization and length-aware local attention pooling produce fixed relative-time representations while preserving acoustic axes for improved detection.
Asymmetric Dynamic Routing (ADR) proposes dynamic traversal in hypergraph RAG to balance reasoning depth and computation efficiency. It addresses the static retrieval fallacy where simple queries incur unnecessary traversal and complex queries lack sufficient cognitive context. ADR adapts routing decisions to query intent, enabling deeper reasoning when needed while reducing overhead for easier queries.
We present SMILESGNN, a multimodal architecture that fuses a SMILES Transformer encoder and a GATv2 graph encoder to predict clinical toxicity. The model addresses class imbalance and scaffold-based generalization, and provides interpretable predictions via cross-attention fusion of the two modalities.
We present AT-SKM-Net, an accelerated trainable Sampling Kaczmarz-Motzkin framework for linear hard-constraint feasibility on dynamic graphs. It improves upon T-SKM-Net by using an accelerated sampling-based approach that avoids full constraint processing and matrix factorizations, enabling scalable feasibility in changing graphs.
We introduce HyPER, Hypergraph for Particle Event Reconstruction, a graph neural network architecture that uses hypergraph representations for reconstructing short-lived particles in collider events with many final-state jets. The hypergraph approach yields more powerful representations and improved efficiency for complex final states such as all-hadronic tt̄ decays.
We present BiGraph-Diffuse, a bidirectional diffusion language model with graph-structured retrieval for mental health counseling. The bidirectional diffusion process helps capture layered emotional expressions and progressive disclosure, while graph retrieval supports context-aware responses.
SciWalker introduces a framework to synthesize scientific coding problems by sampling operator-chains and using execution feedback. It blends scientific library interfaces with operational graphs to generate diverse, realistic problems that align with algorithmic tasks in scientific computing. This framework aims to improve the quality of training data for LLMs in scientific coding domains.
We introduce a growth-inspired graph-generation framework for mechanical lattices. A discrete dot matrix supplies potential nodes, and the final architecture is created by sequential cross-layer and intra-layer growth, visualized in 2D as a leaf-vein-like developmental sequence. The approach enables GCNN-guided inverse design of lattice structures.
We propose TinyCardioUNet, a lightweight UNet that translates six-axis IMU data into ECG signals. It does not require channel selection, improves bottleneck representation with a graph neural network encoding inter-axis dependencies, and uses tensor-decomposition with Bayesian rank selection for parameter reduction. On a public dataset it achieves an RMSE of 0.098 and strong correlation.
We introduce Transformer-GAT for cross-modal dialogue emotion recognition, combining Transformer-based global context with Graph Attention Networks to fuse visual, language, and other modalities. The hybrid architecture captures synergistic information across global and local contexts, improving recognition performance.
CodeGraph proposes an open-taxonomy knowledge graph for source code grounded to wide-open ontologies. It leverages a code-specialized large language model to extract entities representing algorithms, paradigms, patterns, and domains, and grounds these entities to an open taxonomy to enable semantic annotation of code.
We propose an uncertainty-aware clinical knowledge graph for chest X-ray device reasoning, representing device instances, tip estimates, placements, provenance, report events, and temporal links as connected evidence. The system integrates image-derived geometry with textual provenance to support reasoning under uncertainty.
Physical Internet (PI) applies the digital internet paradigm to freight transportation to enable sustainable, interoperable, and collaborative logistics. This study explores Graph Retrieval Augmented Generation (GraphRAG) using GPT-4o mini and Neo4j to review and integrate PI literature, helping researchers organize and synthesize emerging work. The approach aims to streamline knowledge discovery in PI research.