Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 32 papers for 2026-09-01
Nothing cleared the bar today. Papers read in full today: 7.
The load-bearing contribution is an explicit interaction hierarchy: additive aggregation loses the pairing between residue identity and relative geometry, whereas a geometry-conditioned matrix can retain it; the binding-swap proposition is correct for the narrowly defined additive layer. The operator-only ablation is the paper's strongest evidence: matrix gating improves pocket AP@0.5 from 0.432 to 0.467 over channel gating, while channel gating already captures nearly all of the EC gain (0.777 versus 0.789 Fmax), so the empirical result supports a useful diagnostic but not a new general expressivity paradigm. The advertised “hypergraph” is not a substantive new graph-learning mechanism, since the paper explicitly uses it only to partition sequence versus contact neighbors rather than performing hypergraph propagation.
Its “bilinear ceiling” is largely a characterization conditional on restricting a single message to second-order functions of a selected invariant geometry descriptor, and therefore does not establish a ceiling for nonlinear multi-layer, equivariant, or existing ECC-style protein encoders.
The load-bearing contribution is the local-privacy definition: a server need only be unable to distinguish among requested messages it actually stores. The paper backs it with explicit schemes and converses, most notably C(C_N)=1/2 independent of N, C(S_N)=1, and C(P_N)=(N-1)/(2N-4) for odd N, demonstrating that standard graph-PIR's vanishing-rate behavior can be an artifact of imposing privacy at irrelevant servers. This is a meaningful information-theoretic reframing with concrete consequences, but it is specialized PIR theory rather than a broadly reusable graph-learning mechanism.
The disconnected-union capacity theorem is supported by an invalid-looking converse step—an aggregate rate exceeding a weighted average does not imply it exceeds the rate of any individual component—so the headline multiplicative disconnected-graph gain is not rigorously established by the proof shown.
Unsupervised Multi-Scale Gromov-Wasserstein Hypergraph Alignment. We study unsupervised node correspondences between hypergraphs using multi-scale GW, without features or labels. Higher-order formulations are expensive, so we propose a graph-reduction approach that maintains alignment signals while improving efficiency.
TAAL tackles early beam pruning in generative recommender systems where items are encoded as hierarchical semantic IDs. Most failures occur within the first two decoding steps; TAAL (Temporal Autoregressive Alignment) aligns temporal transitions to reduce brittle pruning and improves retrieval across benchmarks.
We propose ESNN, Equivariant Sheaf Neural Network, which learns directed, matrix-valued transport between neighboring vector features on graphs while preserving exact Euclidean equivariance. By keeping scalar and vector features at first order, ESNN enriches geometric transport without increasing representation order.
Converse and Collision-Based Achievability for Node Localization with Hybrid Distance-Spectral Graph Positional Encodings. We propose a hybrid distance-spectral encoding as an observation map, derive a simplex-refined converse and a collision-based factorization of identifiability, and demonstrate results on random regular graphs.
Beyond Ranking Accuracy evaluates LLM-cited feature rationales for next-basket repurchase. While ranking is common, the work argues that users benefit from concise, interpretable evidence grounded in behavioral signals; the paper studies how LLM-generated rationales can accompany item rankings to improve user trust and action.
We introduce Conservative Hybrid Graph Network (CHGN) to handle dynamically changing topologies in industrial process networks. CHGN learns routing with physical conservativity constraints, linking measured trajectories to latent routes while maintaining interpretability and stability.
BiG-SURE introduces a cross-temperature uncertainty estimator for LLMs and VLMs, enabling reliable uncertainty estimates when parameters are inaccessible. It uses low-temperature responses as stable semantic anchors and high-temperature responses as probes under meaning-preserving transformations, and builds an anchor-probe Bipartite Graph (BiG) to quantify agreement and uncertainty.
CASTANET proposes a causality-aware spatio-temporal adversarial network to predict non-periodic traffic congestion caused by incidents. Incidents are sparse and context-dependent, making forecasting hard; CASTANET models incident effects with causal relationships to improve predictions beyond regular daily patterns.
Graph4BiLO: Graph Neural Network Approximation for Bilevel Mixed-Integer Linear Optimization. We introduce GNNs to approximate the follower's value function in bilevel MILO, enabling faster evaluation of leader decisions and scalable learning.
MolLedger: An Additive Graph Neural Network with Chemically Grounded ADME Attributions. MolLedger outputs predictions as the sum of per-atom scores, delivering exact interpretability of ADME properties without sacrificing predictive performance.
CoMPASS: Collaborative Molecular Property Prediction via Adaptive Small-Large Model Synergy. This framework uses retrieval-calibrated integration of small calibrated models with an LLM to decide when and how much the LLM should influence molecular property prediction, improving reliability.
TopoCompress offers training-free, model-agnostic long-context compression by selecting coherent semantic spans. It first scores spans by dense and lexical relevance together with semantic acceleration, then constructs a hybrid graph to guide compression without requiring target-model training or alignment.
CAMIE introduces a co-engagement-aware multimodal item embedding framework for Snap Dynamic Product Ads retrieval. Building on LLM/MLLM backbones, CAMIE mitigates fragmentation from separate encoders and content-only training by aligning multimodal embeddings with co-engagement signals that drive downstream conversions.
HF-SID proposes High-Fidelity Semantic IDs for generative retrieval in location-based services. Since POI are encoded as SIDs, preserving fine-grained geographic differences is crucial; existing SIDs blur numerical coordinates, hurting retrieval. HF-SID preserves fine-grained geography in embeddings, enabling more accurate generation-based retrieval.
We present IsleNet for spatiotemporal graph unlearning. It uses spatial-entropy based partitioning to achieve exact and efficient removal of unauthorized data without full retraining.
Unlearning on Spatio-Temporal Graphs through Subgraph Virtual Edge Reconstruction. We propose subgraph virtual edge reconstruction to erase a node's influence by reconstructing virtual edges within subgraphs, enabling efficient unlearning with preserved performance.
LLMODE: Aligning ODEs with LLMs via Gated Token Injection for Irregular Spatio-Temporal Forecasting. We propose a token-efficient framework that uses a graph-aware ODE encoder to reconstruct irregular observations as a continuous-time latent trajectory, then utilizes a frozen LLM with gated token injection to forecast under irregular sampling.
GraM-Diff: A Unified Graph-Mamba Diffusion Framework for EEG-Based Alzheimer's Disease Data Generation and Diagnosis. It embeds Graph Convolutional Networks within a diffusion framework with classifier guidance to synthesize realistic EEG data and improve AD diagnosis by capturing spatial-temporal brain dynamics.
TopGQ: Fast GNN Post-Training Quantization Leveraging Topology Information. TopGQ uses dual-axis scale absorption to quantize activations along outer and inner dimensions by merging one into the adjacency matrix, and introduces TopPIN to group nodes by local structure, reducing quantization overhead while preserving accuracy.
ToxLens: A Reproducible Graph-Learning Framework for Leakage-Aware, Uncertainty-Calibrated Molecular Toxicity Prediction. It provides a reproducible multi-task pipeline for 11 toxicity endpoints, combining conservative chemical curation and sphere-exclusion to prevent leakage and calibrate predictive uncertainty.
Agents as Knowledge Integrator and Utilizer investigates the semantic misalignment between multimodal content and recommendation objectives. The paper argues multimodal signals should be interpreted together with user behavior before constructing graphs or adjusting ranking, bridging knowledge integration with recommender performance.
Diachronic Hypergraphs for Orchestrated Multi-Agent Multimodal Memory Curation proposes diachronic hypergraphs to preserve and transfer evidence, role-specific context, decisions, procedures, and experiences in multi-agent memory. This approach overcomes limitations of vector/graph memories, enabling richer, traceable knowledge sharing across interactions.
This work develops Context-Aware Interpretable Representations for Retrieval and Graph Convolutional Network Classification. It addresses the Geometric Gap and Interpretability Gap by designing representations that align with the intrinsic geometry of data manifolds, improving retrieval and GCN classification and enabling interpretability.
From Extraction to Governed Memory: Multi-Agent Knowledge Graph Construction with Domain-Expert Review. MAGG presents a principled multi-agent framework for building Governed Knowledge Graphs with explicit governance decisions and domain classification to ensure trustworthy knowledge sharing.
Validating FKG.in: Soundness Assessment in LLM-Augmented Indian Food Knowledge. We develop a semi-automated workflow to validate structured recipe data extracted/augmented by LLMs, focusing on Indian cuisine, to detect hallucinations and ensure knowledge soundness.
ButterMamba: Butterworth-Enhanced Spatial-Temporal Mamba for Efficient Traffic Flow Prediction. ButterMamba improves real-time traffic prediction by integrating Butterworth filtering into a spatial-temporal Mamba framework, reducing attention complexity and mitigating high-frequency sensor noise.
EDGE: Engine for Deterministic Graph Evaluation through Conversation Simulation from Graph Structured DSL Configuration. The paper presents a formal evaluation engine that uses AgentGraph and a graph-structured DSL to simulate conversations and exhaustively enumerate paths, enabling deterministic evaluation of agent behavior.
Beamforming Design Via GNN in mmWave Cell-Free Massive MIMO Using Sub-6 GHz CSI. Represent CFmMIMO as a wireless graph and train a GNN to approximate beamformers that maximize downlink sum-rate using sub-6 GHz CSI, reducing training overhead while approaching optimal performance.
LLMs Interpret, Embeddings Organize, Graphs Emerge presents ASKS, the Agent-Driven Scientific Knowledge System. For each source, an LLM constructs a readable wiki view and a machine-facing semantics; deterministic checks derive a GraphDelta, and embeddings plus explicit graph rules integrate changes into persistent state, enabling scientific knowledge compilation across tasks.
RENSA introduces Rich Environment Metadata to navigate shared and distributed endpoints for automated federated SPARQL query generation. Federated querying is hard due to limited structural knowledge; RENSA provides rich environment metadata to guide query planning and optimization across datasets.