Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 19 papers for 2026-10-09
Nothing cleared the bar today. Papers read in full today: 1.
This paper questions whether hindsight can teach an agent what it could have anticipated before acting, addressing a limitation of verifiable-reward RL where group-relative objectives lose useful signal when all rollouts share the same reward. It introduces prospective learning, which uses post-hoc experiences to supervise foresight predictions made before actions, turning hindsight into proactive guidance. The goal is to recover informative learning signals from trajectories under uniform rewards for group-relative objectives.
This paper examines quantizing recurrent states in delta-rule linear attention, where fixed-size memory is used instead of growing KV caches. They show that naive low-precision quantization can cause large accuracy drops because quantization errors propagate through repeated state updates. The authors reveal that error impact depends on when long-lived memories are used (temporal dimension) and where in the memory the errors occur (spatial dimension) across key rows and state magnitudes, and they introduce STEPQuant to measure and mitigate these effects by adaptive precision across time and space.
This paper advocates nanoMuse, an open-source personal agent designed to operate across all devices owned by a person, managing accounts, memory, and conversations over weeks. It argues that such agents are software systems rather than single models, and it outlines five questions and three horizons to define the personal agent’s scope, capabilities, and deployment. The report sketches how Muse-inspired principles can be realized in an open, cross-device implementation.
Long-WAM introduces a model-system framework to scale the context of causal world-action models under real-time control constraints. It shows that longer histories are not enough; benefits are largest when the video foundation is pretrained autoregressively, enabling better history-to-future reasoning. The approach first learns causal predictions from unlabeled robot and egocentric videos and then preserves this history-to-future structure as it scales to action-conditioned world models.
UltraText Bench is a bilingual benchmark designed to evaluate how image generation models render dense visual text from prompts. It pushes beyond short strings by testing long, multi-region strings across a diverse set of 24 real-world scenes and three difficulty levels, with prompts provided in English and Chinese. Each prompt includes exact strings for 4–12 text regions and structured references to guide evaluation.
LEVER is an adaptive cost-aware proof search over AND/OR graphs. It makes the objective programmable and tunes search toward cheaper, higher-quality proofs by scoring partial proofs with realized values and subgoal predictions. This enables dynamic cost-aware optimization during automated theorem proving.
We propose using Weisfeiler–Lehman features for automated algorithm selection in constraint optimization. The approach automates feature extraction by graph-converting instances and applying WL graph kernels to generate robust structural representations for predictors, enabling more accurate algorithm selection across diverse problem instances.
RIT-RAG navigates document corpora with Retrieval-Induced Trees to provide structure-aware iterative search for retrieval-augmented generation. It addresses the limitation of agentic RAG that commits to a single document and cannot recover from a wrong choice by introducing tree-structured retrieval that preserves document structure and scales to large corpora.
We present C4-equivariant flow matching on anisotropic power-diagram graphs for microstructure generation. By representing polycrystalline microstructures as anisotropic power diagrams, the model employs flow matching with graph neural networks to learn a generative process that preserves symmetry and realism. This yields controllable, realistic microstructure samples for materials design.
MPGE is a Multi-Perspective Graph Explainer for molecular classification. It unifies factual support, counterfactual sensitivity, and exemplar tolerance into explanations for a frozen classifier, delivering a compact, prediction-preserving rationale. This enables users to understand which features support a decision, how changes could reverse it, and which similar examples the model tolerates.
Regularized Small Area Estimation with Graph Laplacian Benchmarking Priors develops a Bayesian SAE framework with benchmarking priors induced by regularization. The family includes Benchmarking Priors, Single-View and Multi-View Laplacian Benchmarking Priors that integrate benchmarking constraints and graph-based borrowing across areas.
MGRASRec is a multimodal graph retrieval-augmented framework for sequential recommendation. It fuses multimodal user–item representations with retrieval over collaborative-filtering paths to provide relevant context while avoiding excessive LLM inference. By retrieving and integrating information from graph paths, it leverages collaborative signals to improve recommendation quality.
Compactness and Consistency proposes a conjoint framework for deep graph clustering that jointly optimizes cluster compactness and alignment with global relations. It targets the shortcoming of purely local message passing in capturing global structure and reduces the impact of noise and redundancy in graph data. The resulting clusters exhibit stronger separation and global coherence.
Evi-VN offers Hard Region Guided Virtual Node Evidence Injection to strengthen GNN-based fraud detection. It identifies a shared hard region where fraud evidence concentrates beyond what graph topology and standard features reveal, and injects virtual-node evidence to emphasize these regions. Across diverse GNNs, this approach improves discrimination of fraudsters and disguised accounts.
Batch Before You Lift proposes scalable topological deep learning on large graphs by avoiding full-domain lifting materialization. It introduces batching strategies that construct lifted representations on the fly during training, enabling exploration of higher-order structures on large networks like Reddit without prohibitive memory costs.
Higher-Order Morphology Priors for Quadruped Reinforcement Learning Under Actuator Degradation introduces higher-order morphology priors by modeling the Unitree Go1 as a cell complex with limb- and body-level rank-2 cells and applying Hodge-based message passing. Under actuator degradation, this architecture with a Hodge-based actor leverages the body-limb hierarchy to maintain robust locomotion.
MSGAT: Multi-Head Spiking Graph Attention with Similarity-Space Fusion for Image-Text Retrieval proposes a spiking graph attention mechanism with multiple heads and a similarity-space fusion module to improve cross-modal alignment in energy-efficient spiking neural networks for image-text retrieval.
Automated Detection of Match Phases in Football from Spatio-Temporal Tracking Data Using Graph Neural Networks combines a graph neural network with a sequential model to classify match phases on a second-by-second basis across seven classes, validated on 203 matches.
We develop a graph neural network for global daily FRP prediction at medium-range lead times. The work addresses the latency and fixed-input limitation of operational FRP products used to initialize forecasts by training a GNN to predict FRP several days ahead, integrating updated fire observations within the forecast horizon to improve air-quality predictions.
Machine Learning Meets High-Energy Nuclear Physics reviews how ML methods are maturing toward physics-integrated workflows, including calibrated Bayesian inference for QCD matter properties and physics-informed discovery in dense matter, marking a shift from mere pattern recognition to physics-grounded analysis and inference.
CORAL introduces Cross-modal Vector Retrieval via Incremental Graph Construction at Scale to tackle out-of-distribution robustness and dynamic updates. It incrementally builds graphs to better exploit query modal characteristics, improve GPU parallelism, and support scalable, dynamic cross-modal retrieval.
Traveling with a Map studies reducing the search space of link traversal queries in decentralized networks through RDF shapes. It aims to address high data transfer costs and long executions in LTQP by pruning queries with RDF shape-based constraints, enabling more efficient query processing in decentralized web settings.
Introducing NEMORA, a neural equivariant multipole operator approach for long-range atomistic learning. It overcomes finite-cutoff limitations of traditional equivariant GNNs by using neural multipole operators to propagate long-range information without fixed kernels. The method preserves rotational and translational symmetry while enabling scalable long-range communication in atomic environments.
An Explainable Header-Centric Framework for Large-Scale Semantic Table Interpretation and Data Quality Assessment describes a header-centric metadata interpretation approach for metadata-only Semantic Table Interpretation. It maps headers to 39 interpretable FinalFormat types, enabling traceable knowledge graph preparation and explainable data quality assessment.