Showing 13 papers for 2026-06-06
We introduce SAGE-PTQ, an ultra-low-bit post-training quantization framework for LLMs that minimizes hidden scaling costs. It separates salient and unsalient weights via distributional statistics and models the subsampled unsalient weights as a sparse graph, reducing overhead while preserving accuracy.
Beyond Vector Similarity: A Structural Analysis of Graph-Augmented Retrieval for Industrial Knowledge Graphs. We compare eight retrieval architectures for aerospace supply chain knowledge graphs, progressing from text retrieval to graph traversal and graph computation. Using a 46-node KG with 64 typed edges and 23 queries across 10 intents, we show that five query classes are structurally unreachable for vector-only retrieval, supporting the operator vocabulary thesis that effective retrieval requires graph-based reasoning beyond vectors.
Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents. We introduce MRAgent, a memory framework that couples an associative Cue-Tag-Content graph with an active reconstruction mechanism, enabling dynamic, evidence-driven memory access during inference rather than static retrieval followed by reasoning. This graph memory enables reconstruction of relevant memory traces to support long-horizon reasoning.
TokenMizer: Graph-Structured Session Memory for Long-Horizon LLM Context Management. We present TokenMizer, an open-source proxy system that models LLM session history as a graph rather than flat text, preserving structured information such as architectural decisions, task transitions, and file histories across long sessions. This graph-structured memory helps maintain context beyond the Maximum Effective Context Window (MECW) and enables resumable long-horizon work.
CausalPOI: Spatio-Temporal Graph-Based Causal Modeling for Cold-Start POI Check-in Forecasting. We introduce CausalPOI, a spatio-temporal graph-based causal modeling framework for forecasting Point-of-Interest check-ins under cold-start conditions. Unlike proximity-based graphs and correlation-driven models, it captures functional dependencies and causal effects of urban interventions, improving prediction accuracy for evolving city environments.
HDST-GNN: Heterogeneous Dynamic Spatiotemporal Graph Neural Networks for Multi-Object Tracking in UAV Aerial Imagery. We propose HDST-GNN, a heterogeneous dynamic spatiotemporal graph neural network for multi-object tracking in UAV imagery. It addresses altitude variation and object heterogeneity by differentiating detections, active tracklets, and lost targets, and introduces altitude-adaptive edge construction to better reflect spatial relations, leading to improved MOT performance under challenging aerial conditions.
Beyond Soft Masks: Hard-Perturbation Mixup Explainer for Robust GNN Explainability. We present a hard-perturbation mixup explainer to improve the robustness of post-hoc GNN explanations beyond traditional soft-masking approaches. By generating perturbed subgraphs and mixing them, the method produces more faithful and OOD-resilient explanations for GNN predictions.
Reducing Hallucinations in Complex Question Answering using Simple Graph-based Retrieval-Augmented Generation (long version). We propose a simple graph-based RAG approach to reduce hallucinations in complex QA, leveraging graph structures to ground answers with reliable evidence from proprietary data. Experiments show improved factuality and reduced hallucination rates compared to baseline RAG methods.
AttackPathGNN: Cross-function vulnerability detection in smart contracts using state interference graphs and conjunction pooling. We introduce AttackPathGNN, a GNN-based detector that reasons over cross-function vulnerability paths. By modeling state interference graphs across functions and using conjunction pooling to capture combined conditions, it detects vulnerabilities that emerge from inter-function interactions beyond single-function analysis.
Harnessing Structural Context for Entity Alignment Foundation Models. We propose ContextEA, an entity alignment foundation model that leverages structural context to improve cross-KG alignment. It addresses two issues: weak cross-KG interaction during encoding and heavy reliance on coarse similarity in final ranking, by incorporating richer structural signals into both encoding and ranking.
Bridging the Semantic-Collaborative Gap: An Asymmetric Graph Architecture for Cold-Start Item Recommendation. We design an asymmetric graph architecture for cold-start item recommendation in Tubi's production system. The model produces standalone embeddings for new content for immediate retrieval, while learning device-specific embeddings to support approximate nearest-neighbor search, enabling seamless cold-start recommendations.
Generating Graph-Like Logical Rules for Knowledge Graph Reasoning via Diffusion Models. We propose a diffusion-model–based approach to generate graph-like logical rules for KG reasoning, enabling cycles and branches beyond simple chains. This addresses the limitations of traditional rule mining and helps manage the combinatorial search space.
A2RAG: Adaptive Agentic Graph Retrieval for Cost-Aware and Reliable Reasoning. We present A2RAG, an adaptive and agentic Graph Retrieval-Augmented Generation framework that tailors retrieval to mixed-difficulty workloads and mitigates extraction loss by preserving fine-grained source qualifiers. The system uses an agent to guide evidence routing for cost-efficient, reliable reasoning.