Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 15 papers for 2026-08-21
MileGPO introduces Milestone Inference with Local Evidence for graph-based policy optimization in long-horizon LLM agents. It derives process-level credit from grouped on-policy rollouts using three designs, including milestone discovery and local-evidence credit assignment. This enables finer credit allocation to intermediate progress rather than only final rewards.
Multi-Source Wasserstein Distributionally Robust Graph Learning proposes a distributionally robust framework for graph topology inference with multiple source domains. It uses the Wasserstein metric to preserve distributions while fusing heterogeneous sources, preventing biased consensus that collapses distinct geometries. The method improves robustness when target-domain samples are scarce.
G-MARK presents grounded multi-agent reasoning for cooperative driving via knowledge graphs. Under partial observability and occlusion, existing methods hide which agent observed what and how evidence affects decisions. G-MARK grounds multi-agent evidence in knowledge graphs to reveal object visibility, observer provenance, and the impact of conflicting evidence on downstream decisions.
Reward-Guided Autoregressive Graph Generation for Efficient Multi-Agent Communication Topology Design proposes RGA-Designer. It introduces reward signals to guide autoregressive graph generation, encouraging sparse and efficient communication topologies for MAS. This addresses ARG-Designer’s lack of explicit incentives for topology sparsity.
GraphPFN: A Prior-Data Fitted Graph Foundation Model extends the PFN paradigm to graphs. It pretrains on carefully designed synthetic data to enable in-context learning and improved transferability across diverse domains. GraphPFN aims to mitigate data scarcity challenges in graph foundation models.
PROBE-Web is an interactive system for probing evaluation landscapes of knowledge graph completion models. It lets users adjust perspectives, such as predictive sharpness and popularity-bias robustness, to flexibly compare multiple KGC models via a GUI. The system helps diagnose model behavior under different evaluation criteria.
Better Call Graphs introduces a new Android malware dataset based on function call graphs for classification. FCGs capture behavioral structure beyond surface signatures, improving malware detection in the mobile domain. The dataset enables robust evaluation of graph-based malware classifiers.
Self-Evolving Agents as Dynamic Graph Transformation provides a survey on LLM-based agents as self-evolving systems that persist, maintain memories, use tools, acquire skills, refine workflows, and coordinate. The paper emphasizes agent states as dynamic graphs that evolve with evidence and environment, advocating viewing graphs as evolving substrates rather than mere supports.
RDFdL integrates RDF with Differential Dynamic Logic to reason about both static knowledge and continuous dynamics of physical systems. It syntactically represents differential equations within RDF-based knowledge graphs and provides a framework for dynamic reasoning.
GenEx proposes a graph-based representational paradigm for SARS-CoV-2 variant detection via codon co-occurrence networks. It converts raw gene sequences into codon co-occurrence graphs and extracts 25+ graph features to capture contextual interdependencies that traditional linear methods miss.
LEDGER: Claim-to-Evidence Trace Graphs for Auditing LLM Agents introduces a trace-graph based approach to audit LLM agent outputs. It links claims, actions, artifacts, and validation steps to provide end-to-end evidence for conclusions.
GrabVG proposes Graph-Attentive Binding for Visual Grounding in UAV imagery. It addresses high visual redundancy and topological ambiguity due to many small, similar objects by leveraging graph-attentive mechanisms to improve grounding.
G-ReAct: Graph-Guided Deep Search via Structure-State Co-Evolution presents a reasoning framework that uses graph guidance to co-evolve search structure and state. It preserves intermediate states and constraints across long-horizon multi-hop tasks, addressing context forgetting, drift, and exploration inefficiency.
Wildfire Suppression discusses complexity, models, and instances of allocating suppression resources on graphs. The paper proves strong NP-completeness for planar graphs and certain directed grids for two related variants, and analyzes modeling choices and example instances.
Fairness-Aware Network Embeddings reviews methods to mitigate bias in network embeddings while preserving predictive utility, surveys applications and challenges, and discusses fairness constraints and evaluation.