Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 6 papers for 2026-09-12
Nothing cleared the bar today. Papers read in full today: 2.
We present NCP-ArchPreview, a latent-space language model that augments standard next-token prediction with Next Concept Prediction. This concept-level objective operates on discrete concepts spanning multiple tokens while preserving autoregressive token generation. The model builds a latent space by product-quantizing hidden states into a concept vocabulary and uses a dedicated predictor to forecast future concepts, enabling richer representations and planning beyond token-level objectives.
SenseNova-U1.5 is an 8B-MoT native unified multimodal model that understands, reasons about, and generates visual content without an encoder or VAE. It strengthens its visual interface with spatially coherent patch reconstruction and scales training with curated generation and editing data, refined task formulations, structural prompts, and 4K-native resolutions. In a post-training phase, specialized experts for visual aesthetics, bilingual rendering, infographic generation, and image editing are integrated to broaden its capabilities.
SpatialBlock introduces a synthetic block-stacking problem to teach LVLMs spatial intelligence, enabling them to infer 3D structure from 2D images without costly real-world geometric annotations. By leveraging synthetic data and a targeted training curriculum, it improves spatial reasoning in LVLMs while reducing reliance on external perception modules and expensive labeling.
EvoSafeHarness develops an adaptive safety framework for LLM agents by evolving model- and domain-specific harnesses. It acknowledges that different models and domains require different enforcement levels, and tunes constraints to balance safety with utility while guarding against both indirect prompt injections and direct harmful requests at the system level. The framework supports deployment-time adaptation through an automated or guided evolutionary process.
Mi-Ripple presents a diagnosis-guided restoration workflow that suppresses digital ripple artifacts produced during iterative AI editing while preserving image structure. It disentangles periodic lattice artifacts from content-entangled texture and applies selective spectral notching, structure-aware smoothing, and cleaned-reference regeneration to achieve low-distortion restoration; when removal would harm content, it regenerates content from a cleaned reference.
We break ground on Specified-Foil Counterfactuals for Temporal Graphs. Given an original prediction A and a foil B fixed in advance, we search for a low-cost past-event intervention that makes the predictor choose B as the top option. The method uses trace-guided search to efficiently identify feasible counterfactuals, enabling destination-specific counterfactual explanations.
We present Mosaic, a training-free framework that treats GraphRAG retrieval as a per-query control problem. An LLM analyzer translates each query into query-specific retrieval actions to steer exploration. This approach mitigates structural mismatches by enabling local neighborhoods, balanced coverage across targets, or deeper paths through connectors.
We propose DRG-MAPPO, a hierarchical dynamic Role-Graph MARL framework for cooperative air combat. It addresses two core limitations: lack of structured relational modeling and flat architectures, by introducing a dynamic role-graph and hierarchical policies to capture time-varying interactions among battlefield entities. The approach aims to improve tactical coordination and decision-making under complex, adversarial scenarios.
From Queries to Narratives proposes narrative-based data stories to support knowledge graph exploration and quality assessment for cultural heritage KGs. Cultural-heritage KGs like NFDI4Culture-KG hold millions of triples; although SPARQL is learnable, many users need starting points and guidance due to data-model complexity. We provide narrative-guided exploration and quality assessment workflows that translate graph structures into approachable stories, easing discovery and evaluation.
From Document Silos to Process Intelligence introduces a multi-layer knowledge graph and modular agentic-AI platform to support CMC process development. It converts heterogeneous process-development documents into a unified, queryable resource, enabling end-to-end traceability across discovery, development, and regulatory filing. The modular AI agents orchestrate data integration across domains to reduce knowledge-management costs and accelerate technology transfer.
Geospatial AI, Dataverse Metadata, and the Study of Place-Based Government builds a geospatial-aware knowledge graph for Harvard Dataverse. The graph connects 102,650 datasets with 528,003 edges to keywords, publications, subjects, journals, and locations, with 43,991 datasets carrying geospatial fields. This resource enables researchers to study place-based governance and the role of geospatial metadata in data discovery and policy analysis.