Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 6 papers for 2026-09-22
Nothing cleared the bar today. Papers read in full today: 2.
EvoOntology proposes a self-evolving ontology layer to bridge the agent-data gap in data agents that execute natural-language instructions over heterogeneous data sources such as tables, files, and databases. By providing a scalable, adaptive semantic layer that operates beyond static prompts, EvoOntology aims to scale to large data sources and adapt to different agent behaviors.
CodeMidas presents a pipeline that turns implemented functionality in existing codebases into executable reinforcement learning environments, using source code as the sole task-specific input. This approach aims to scale agentic coding RL by avoiding reliance on development artifacts and by allocating compute to every stage of environment construction.
Code2Skill introduces a fully automated pipeline that transforms selected code units into implementation-anchored, reusable skills for agentic intelligence. By leveraging executable code as evidence, it grounds abstractions and overcomes limitations of trajectory-based synthesis and document-derived skills.
RecreationWorld introduces a five-platform framework for hybrid computer-use agents that autonomously decide when to explore interfaces, implement software, and run and visually verify artifacts. Given a running reference, the agent must discover the behavior and build a faithful implementation with no prescribed workflow, enabling scalable and verifiable hybrid environments.
IntBMoE proposes Integrating Block-Level Conditioning into Expert Composition for Full-Participation Mixture-of-Experts, enabling independent control of participation, execution, and materialization costs. By introducing block-level conditioning, it aims to achieve full token participation without prohibitive compute or memory overhead, addressing the trade-offs of sparse routing and dense output mixing.
This paper studies the temporal generalization and explanation stability of Graph Neural Networks over Control Flow Graphs for malware detection. It uses strict temporal splits (train on one period, test on a later one) and analyzes both detection performance and the stability of explanations across two CFG corpora, where each node carries 37 features.
We introduce a scalable analytical framework to approximate Neighborhood Locally Fair PageRank and Uniform Locally Fair PageRank, addressing the scalability bottlenecks of exact convergence. The approach uses a group-aware heterogeneous mean-field representation to enable fast, accurate locally fair rankings on large graphs.
This paper studies cross-model migration of graph-based indexes for approximate nearest neighbor search by exploiting residual reachability. It investigates reusing parts of an existing graph index when the embedding model changes to avoid full rebuilds.
This work addresses Consistent Relexicalization of Clinical Documents using a Graph-Based Approach. It aims to mask sensitive information while preserving longitudinal structure, relational coherence, and temporal consistency across records; existing entity-by-entity replacements often introduce cross-time inconsistencies.
We propose Efficient Dense Vector Search within Knowledge Graph Content Embeddings, enabling SPARQL engines to natively operate on dense embeddings integrated with KG content for neurosymbolic reasoning. This enables tensor operations in SPARQL and tighter coupling of language models with knowledge graphs.
The study provides a comparative evaluation of graph databases using a large synthetic biomedical property graph (1.02 million nodes, 5.34 million total rows) and a twenty-query workload. It analyzes query, ingestion, and update performance across engines, highlighting differences in query planning, indexing, and data-readiness costs.