Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 10 papers for 2026-09-05
Nothing cleared the bar today. Papers read in full today: 2.
The load-bearing claim is diagnostic rather than architectural: frozen MLLMs fail partly because they are given the wrong evidence modality, and a 3DSG can expose the appropriate evidence without retraining. The paper tests this claim module by module with same-backbone comparisons: deterministic tools substantially beat text-serialized graphs on metric VSI-Bench tasks, selective BEV modestly but consistently beats Struct2D, and geometry-guided top-2 frames raise Qwen2.5-VL-7B ScanQA CIDEr from 58.0 to 73.6. The full realistic-perception result (51.4 VSI-Bench average) is competitive with or exceeds several trained spatial models, although it is an evidence-routing system atop an upstream reconstruction pipeline rather than a new graph-learning method.
The strongest gains rely on an offline 3DSG with calibrated poses, object boxes, labels, and stored visible frames, while realistic perception sharply reduces several metric-task results and the paper does not establish robustness, latency, or amortized cost for genuinely noisy online scene graphs.
Many recurring text functions are easy to describe but difficult to implement with rules, while calling a large remote model for every input introduces repeated cost, latency, and dependency on a provider. We present compile by training, which turns a natural-language specification into a reusable n
As terminal-based code agents become prevalent, agent trajectories have accumulated at scale, while realistic, executable environments remain scarce. However, environments are what agent post-training actually requires: each can be re-queried into many verifiable tasks and provides execution feedbac
We introduce LLaDA-Image, a unified framework that pairs a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. Instead of relying heavily on paired image-text data from the beginning, we
Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, a
Large language models offer broad capabilities, but adapting them to evolving domains, tools, and requirements often entails repeated post-training. Autonomous systems automate parts of this process by proposing updates, training candidates, and using evaluation feedback to select subsequent proposa
Pattern Over-Generalization of Knowledge Graph Embedding analyzes how knowledge graph embedding models capture inference patterns such as symmetry, inversion, and composition, and discusses limitations arising from pattern over-generalization where embeddings learned for one pattern may erroneously apply to others.
AutoGraphForge describes a pipeline for automated graph-theoretic conjecturing, refuting, formalizing, and proving. Conjecture generation is counterexample-guided and proceeds in rounds; a Graffiti3 generator proposes conjectures on a growing snapshot table T of graphs and invariants, with growth driven by counterexamples to its own conjectures. A novelty filter of 559 classical and folklore relations, closed under transitive composition and linear inference, prunes proposals.
Privacy-Preserving Topology-Guided Safety for LLM-Based Multi-Agent Systems via Federated Graph Learning introduces FGLGuard, a privacy-preserving framework that trains a graph neural network over inter-agent graphs using federated learning. It localizes risky agents by examining edge features without pooling private traces across organizations, enabling topology-guided intervention while protecting prompts, tool outputs, and workflows.
DNative-Twin introduces a graph-native digital twin that records a committed agentic decision as a typed trajectory and allows re-execution of its decision mechanism under declared conditions. The graph links the observed state, the decision path followed, and the authority behind the resulting action, enabling reconstructability and auditability of agent decisions. This framework helps verify which evidence, tool states, constraints, and authorizations contributed to a given action.
PPO-STGNN proposes a reinforcement learning approach to DAG task scheduling in heterogeneous cloud-edge-end environments. The method combines Proximal Policy Optimization with a Spatio-Temporal Graph Neural Network to model system dynamics, node heterogeneity, and task dependencies. It aims to efficiently schedule complex, dependency-driven DAG tasks despite NP-hardness, improving performance and energy efficiency.
Semantic Bayesian World Models propose a world model where beliefs over knowledge graphs are described probabilistically rather than fixed facts. This enables probabilistic reasoning that is constrained by ontological axioms, bridging language models and knowledge graphs into a unified architecture. The vision is to treat the world as an evolving fabric of beliefs rather than a static database.
LevelSyn proposes physical-aware logic synthesis via Level-Asynchronous Graph Neural Networks to reduce PPA degradation in nanometer ICs. It couples logic synthesis with physical design by using GNNs that respect hierarchical level structure and signal flow, delivering more faithful spatial estimations than traditional wire-load models and shortening design closures.
SENTINEL-RL proposes an agentic-SOC architecture that offloads topological reasoning from LLM agents to improve scalability and action-topology consistency. It uses a heterogeneous graph attention encoder to summarize the live authentication graph and guide containment actions, effectively decoupling topology-aware decisions from semantic reasoning.
Evaluating Graph Neural Networks for Change-Criticality Classification in Maritime Navigation Charts investigates how different GNN architectures, message-passing schemes, and graph representations affect the classification of changes to objects in electronic navigational charts. The study compares configurations to identify which best capture change significance for chart maintenance and maritime safety.