Showing 15 papers for 2026-05-16
GraphBit proposes an engine-controlled DAG framework for non-linear agent orchestration. It uses typed function agents and a Rust engine to manage routing, state transitions, and tool invocations, avoiding hallucinated routing and loops, ensuring reproducibility and auditability.
We study tokenization of resting-state functional connectivity for MAE-based representation learning. Traditional region-centric or graph-based tokenizations fail to respect large-scale modular brain organization. We propose Network-Aware Representations (NERVE) that encode network structure into tokens to align tokenization with brain modules, improving learned representations.
HEAR is an enterprise agentic reasoner built on a Stratified Hypergraph Ontology. The Graph Layer virtualizes provenance-aware data interfaces, while the Hyperedge Layer encodes n-ary business rules and procedural protocols, enabling evidence-driven, auditable multi-hop reasoning across heterogeneous systems.
This work argues that legal reasoning relies on constrained symbolic processes rather than semantic similarity search, addressing hallucinated precedents and outdated statutes in LLM-based Indian judicial AI. Falkor-IRAC presents a graph-constrained generation framework that enforces graph-based constraints for precedent propagation, procedural state transitions, and statute-bound inference, improving reliability.
COREKG introduces coreset-guided personalized summarization of knowledge graphs. Large KGs are unwieldy for QA and visualization, so the method uses coresets to select representative subgraphs that preserve query-relevant information. The result is compact, user-tailored KG summaries.
KGPFN unlocks KG foundation model potential via in-context learning. It studies how in-context learning can generalize KG foundation models to unseen entities and relations, leveraging structured and heterogeneous KG contexts. Effective prediction requires conditioning on both local context around query entities and global context that summarizes relations.
We frame citation faithfulness as a trajectory-level problem in agentic GraphRAG, where an agent traverses a knowledge graph before generating an answer and citations. Controlled ablation experiments show that incorporating neighborhood traversal and provenance improves citation faithfulness and accounts for visited-but-uncited entities.
We apply reinforcement learning to train tool-calling agents operating on FHIR graphs to correctly select and traverse healthcare resources. Agents must perform multi-step reasoning across multiple resource types while respecting traversal constraints; RL improves resource selection and reasoning accuracy.
We conduct an internal study of RAG, examining how retrieved evidence influences answer generation. Using circuit tracing, we construct attribution graphs that model information flow, shedding light on failure modes and guiding improvements.
Video2GUI proposes a fully automated framework that extracts grounded GUI interaction trajectories directly from unlabeled Internet videos. This enables scalable pretraining of GUI agents without costly manual annotation.
Graphs of Research (GoR) is a supervised fine-tuning method that extracts a 2-hop reference neighborhood for each seed paper. It leverages evolving citation structures to ground and scaffold idea generation, improving coherence and relevance.
WAR-R1 is an explainable Web API recommendation framework that integrates semantic reasoning with adaptive, variable-cardinality retrieval. It provides explanations for recommended APIs to improve transparency and user trust.
CUICurate uses GraphRAG to automate the construction of concept sets—synonyms, subtypes, and related concepts—for NLP tasks. It reduces manual curation and improves downstream NLP performance.
PRAETORIAN defends GNN backdoors by targeting intrinsic requirements of effective backdoors rather than surface cues. By analyzing internal correlations and external influences of triggers, it disrupts the conditions needed for backdoors to function, offering robustness against adaptive attackers.
This paper provides a structural diagnostic for choosing multi-agent LLM communication topologies using the successor representation M = (I - γP)^{-1}. It predicts how topology choice affects drift, convergence, and robustness before inference.