Showing 36 papers for 2026-04-29
We propose Time-varying Interaction Graph ODE to model dynamic graphs. Unlike existing GODEs with fixed message passing, our method allows time-varying interaction patterns by using different message functions over time, improving dynamic graph representation learning. Experiments on benchmark dynamic graph tasks show accuracy gains and better capture of evolving relations.
We introduce GraphPL for patchwork learning, where clients have varying modalities and missing data. GraphPL uses GNNs to exploit cross-modal relationships and imputes missing modalities in an unsupervised way, leveraging the modality graph across clients. It outperforms modality-subset baselines in experiments on patchwork benchmarks.
We study three RGNN stopping paradigms: converging RGNNs, output-converging RGNNs, and halting RGNNs. We establish expressiveness relations among them on undirected graphs, showing a hierarchy where halting is at least as expressive as the others (with strictness under conditions). This clarifies when different stopping criteria matter for representational power.
We conduct a controlled empirical study of PLM-GNN hybrids for code tasks. By pairing three code-specialized PLMs with three GNN architectures, we compare hybrids against PLM-only and GNN-only baselines on Java250 and Devign, including obfuscation. Results reveal which PLM-GNN combinations matter most, showing consistent gains from hybridization.
We adapt GNNExplainer, GNNShap, and GradCAM to LundNet's Lund-plane graphs for jet tagging. Through perturbation, Shapley, and gradient explanations, we compare interpretability and correlate explanations with physical jet features. The study provides guidance on which explainability method to use in high-energy physics GNNs.
FARM introduces functional group-aware representations for small molecules, enriching SMILES and graphs with FG annotations. Atoms carry FG membership info, enabling FG-enhanced SMILES and FG graphs that align with chemical semantics. The approach improves molecular property prediction and downstream generation tasks.
We propose Grothendieck Graph Neural Networks (GkGNN), an algebraic framework that generalizes neighborhoods to covers. This strictly extends the primitive of message passing and yields topology-aware GNNs beyond the WL test. The framework provides algebraic operations to compose new architectures with controlled expressivity.
We present a physics-inspired image classifier where CNN features are treated as spins on a sparse multi-edge quasi-cyclic graph, forming a Random-Bond Ising Model at the Nishimori temperature. This RBIM-based ensemble captures topological structure of image manifolds and achieves competitive accuracy with efficient computation in natural image classification.
We propose WG-SRC, a white-box signal-subspace probe to diagnose graph datasets. By replacing learned message passing with a fixed graph-signal dictionary, WG-SRC exposes which graph-learning mechanisms a dataset requires. The approach yields actionable dataset fingerprints and diagnostic insights.
PhaseGraph calibrates heterogeneous graph and vector retrieval scores for multi-hop QA. By applying percentile-rank normalization to map scores to a common scale, it enables stable fusion without discarding magnitude information. Experiments on MuSiQue show improved retrieval performance.
We investigate whether belief graphs help LLMs in cooperative games like Hanabi. Across thousands of trials, we find architecture matters: belief graphs often help weak models when used as action-gating structures but may be decorative for strong models in other settings. The results delineate when graphs are beneficial.
We evaluate granularity effects in AML scoring by mapping transaction-level scores to actor-level actions. Our projection framework formalizes four aggregation operators and budgeted investigation selects, showing how granularity changes queue composition and downstream actions. The results highlight tradeoffs between granularity and efficiency.
We construct domain-filtered knowledge graphs from sparse autoencoder features. A multi-stage filtering with contrastive activations yields a domain concept universe, enabling two aligned domain concept graphs and relations graphs, supporting more interpretable and domain-relevant graph representations.
We propose a three-phase framework for evaluating sustainable city-trip recommendations with LLM judges and human-in-the-loop. Dimensions include relevance, diversity, sustainability, and popularity balance; calibration rules and few-shot prompts correct systematic misalignments. Experiments show consistent improvements across two recommendation tasks.
XGRAG introduces a graph-native explainability framework for KG-based retrieval-augmented generation. It moves beyond text-only explanations by leveraging the KG structure to trace how specific facts influence the LLM output. The method enables end-to-end interpretability of graph-grounded RAG.
Learning in Blocks presents a multi-agent debate-assisted framework for personalized adaptive language learning. Debates between agents provide evidence-based feedback, guided by LLM-based judging, to tailor pacing and review according to each learner’s needs. The approach aims to improve learning outcomes and engagement.
Epicure analyzes FlavorGraph ingredient embeddings to reveal a multidimensional flavor structure. A curation pipeline consolidates thousands of ingredients into canonical entries, uncovering latent flavor dimensions and cultural associations that can inform culinary AI applications.
Structure-guided RAG for factual queries defines the Exact Retrieval Problem, enforcing that retrieved evidence satisfies query constraints. It pairs structured retrieval with LLMs to reduce hallucinations and improve factuality. Experiments show improved accuracy on factual queries.
An experience report on AI-assisted code review embedded in GitHub PRs shows improvements in code quality and learner self-regulation. Mixed-methods with logs, reflections, and surveys reveal how AI reviews shape engagement, feedback quality, and learning outcomes in capstone projects.
We present a training-free, model-agnostic context compression method using hybrid graph priors to move from similarity-based retrieval to structure-aware summarization. The approach preserves relevance, topic coverage, and cross-sentence coherence under fixed token budgets without training.
EAD-Net introduces emotion-aware talking head generation with spatial refinement and temporal coherence to produce expressive, lip-synced portrait videos. By enriching emotional semantics beyond simple labels, it achieves more natural expressions while maintaining efficient, globally coherent motion for longer videos.
SMSI automates threat modeling for cyber-physical systems using a hybrid neuro-symbolic pipeline that begins with a SysML architecture model and ends with a prioritized list of NIST 800-53 controls. It comprises a deterministic parser mapping components to vulnerabilities via the NVD, a retrieval/classification layer linking vulnerabilities to MITRE ATT&CK techniques, and a control recommender.
Crystal structure prediction using graph neural combinatorial optimization investigates CSP by formulating it as a graph-neural combinatorial optimization problem over unit-cell grids. It aims to assign atom positions to minimize interaction energy efficiently, leveraging end-to-end learning to improve accuracy over traditional discrete-search methods.
Agentic Witnessing proposes a privacy-preserving auditing framework that moves verification from attested execution to attested reasoning. By combining trusted execution environments with attested reasoning, it enables scalable auditing of qualitative properties in proprietary data without leaking sensitive content.
Latent-Hysteresis Graph ODEs introduce HGODE to model coupled topology and feature evolution in graphs via continuous phase transitions. It overcomes the monostability trap of positive-irreducible mixing operators by coupling state evolution with hysteresis, enabling richer long-term dynamics and regime changes.
Modeling Behavioral Intensity and Transitions for Generative Recommendation argues for modeling intensity and transitions among multiple user behaviors in generative sequence models. It addresses the assumption of uniform dependency among historical behaviors and improves conversion predictions by capturing varying behavior signals over time.
Large Language Models as Virtual Survey Respondents evaluates using LLMs to generate sociodemographic responses and proposes task abstractions and a unified evaluation framework. It demonstrates across diverse datasets how LLMs can serve as scalable virtual respondents for sociological surveys.
Reliable Microservice Tail Latency Prediction proposes decoupled dual-stream learning with gradient modulation to predict window-level P95 tail latency in microservice architectures. It explicitly separates traffic metrics from resource metrics to better capture their interactions and improve tail-latency forecasting.
Enabling Transparent Cyber Threat Intelligence combines ontology-driven structured outputs with Large Language Models to build an AI agent that improves extraction, interpretation, and transparency of cyber threat intelligence from logs. The method emphasizes semantically enriched, traceable results.
Verifying Quantized GNNs With Readout Is Decidable But Highly Intractable introduces a logical language for quantized ACR-GNNs and proves that verification tasks with a global readout are (co)NEXPTIME-complete. It highlights fundamental computational challenges while showing quantized GNNs can remain lightweight in practice.
BEAR proposes beam-search-aware optimization for LLM-based recommendation to align training objectives with beam-search retrieval. It addresses training-inference inconsistency where SFT optimizes positive-item probability but does not guarantee retrieval by beam search, improving end-to-end recommendation performance.
Anonymization-Enhanced Privacy Protection for Mobile GUI Agents studies privacy risks of mobile GUI agents that capture screen content and proposes anonymization-enhanced defenses to reduce exposure while preserving task functionality.
Harmonizing Generative Retrieval and Ranking in Chain-of-Recommendation introduces RecoChain, a unified framework that bridges generative retrieval of semantic IDs with downstream ranking to close the gap between generation and ranking in chain-of-recommendation.
Action-Aware Generative Sequence Modeling for Short Video Recommendation proposes an action-aware sequence model that accounts for diverse segments within short videos, reflecting user attitudes toward different parts. This temporal granularity improves recommendation accuracy for short-form content.
IoDResearch introduces IoDResearch, a private-data–centric framework to enable deep research over heterogeneous data sources. It emphasizes privacy protection and data interoperability, aligning with FAIR principles to improve retrieval and reuse of private data.
MemRec introduces Collaborative Memory-Augmented Agentic Recommender System, a memory-augmented framework that shares semantic memories across users and items to reveal hidden preferences and improve recommendations for data-sparse users.