Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 17 papers for 2026-02-13
Step 3.5 Flash is a sparse Mixture-of-Experts model that pairs a 196B-parameter foundation with only 11B active parameters to enable efficient inference for frontier-level agentic tasks. It uses interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce latency in multi-round agent interactions. The authors also outline a scalable reinforcement learning framework to push toward frontier-level intelligence.
VidVec conducts a layer-wise analysis showing that intermediate (pre-trained) MLLM layers already encode substantial task-relevant information for video-text tasks. Building on this, it demonstrates that fusing embeddings from intermediate layers yields strong video-text representations and improves retrieval performance without relying on extensive end-to-end fine-tuning.
GENIUS introduces GEN Fluid Intelligence EvalUation Suite to assess Generative Fluid Intelligence in Unified Multimodal Models. It argues that current benchmarks overemphasize crystallized knowledge and neglect the ability to induce patterns, reason under constraints, and adapt to new scenarios. GFI is formalized as a synthesis of three primitives, including Inducing Implicit Patterns.
PhyCritic introduces a multimodal critic model tailored for physical AI tasks, addressing the gap where existing critics focus on generic vision tasks. It employs a two-stage training process to learn judgments for perception, causal reasoning, and planning in physical contexts, providing pairwise preferences, scalar scores, and explanatory rationales.
ASA proposes training-free representation engineering to improve tool-calling for domain-specific LLM agents. It identifies a Lazy Agent failure mode where tool necessity is decodable from mid-layer activations, yet the model remains tool-shy. The approach introduces Activation Steering to align internal representations with tool-usage behavior without additional training.
We propose GP2F, a cross-domain graph prompting framework that adaptively fuses multiple pre-trained GNNs to better adapt to downstream tasks when domain shift occurs. It analyzes why graph prompt learning remains effective under domain shifts and demonstrates empirical gains across benchmarks.
This work revisits LLM-guided OOD detection on text-attributed graphs (TAGs) and shows that both topology and text drive detection quality. It argues that relying on topology alone is insufficient and advocates joint consideration of textual and structural patterns to improve robustness of OOD detection with LLMs.
SpaTeoGL introduces a spatiotemporal graph learning framework for interpretable SOZ (seizure onset zone) analysis from intracranial EEG. It learns window-level spatial graphs among electrodes and a temporal graph across windows based on spatial similarity, all within a smooth graph-signal processing formulation, enabling interpretable seizure network insights.
TopoFair studies how topological biases in link prediction benchmarks affect fairness. It argues that biases beyond simple homophily limit current fairness interventions and proposes a framework linking topology with fairness considerations to improve generalization across diverse networks.
This work addresses privacy risks of unsupervised graph learning-based clustering revealing sensitive groups. It analyzes how communities could be inferred and proposes concealment or privacy-preserving strategies to protect group-level information.
RokomariBG is a large-scale heterogeneous book graph dataset for personalized Bangla recommendations, containing books, users, authors, categories, publishers and reviews with eight relation types. It enables research on multi-entity recommender systems in low-resource language settings.
GraphPFN introduces a Prior-Data Fitted Graph Foundation Model, pretraining on synthetic data to fit priors and enable better transferability across graphs and data-scarce regimes.
Self-Adaptive Graph Mixture of Models proposes a self-adaptive MoE of GNNs to select and combine models for a given graph task, addressing the difficulty of choosing a single best model and improving robustness across datasets.
CardinalGraphFormer introduces Cardinality-Preserving Attention Channels for Graph Transformers, adding a query-conditioned gated aggregation channel to preserve dynamic cardinality cues and incorporating sparse masking to scale. Applied to molecular property prediction under low data regimes.
EEG2GAIT presents a hierarchical graph convolutional network for EEG-based gait decoding, building a pyramid of spatial embeddings across EEG channels and integrating temporal-spectral features to improve gait decoding accuracy.
TF-DWGNet designs a directed weighted GNN with tensor fusion to classify cancer subtypes from multi-omics data, capturing directionality and interaction strengths between modalities and fusing them via tensor operations for better predictive performance.
Beyond Pixels proposes a Vector-to-Graph pipeline that converts CAD diagrams into property graphs with nodes as components and edges connectivity, enabling reliable structural reasoning for schematic auditing by LLMs.
Text2GQL-Bench provides a benchmark for Text-to-Graph Query Language, evaluating natural language to graph queries across datasets and languages, and enabling LLM-based graph querying with diverse query languages.
Statistical Parsing for Logical Information Retrieval extends a probabilistic logical model with negation factors to enable contrapositive reasoning and provides parser support for natural language, bridging inference, semantics, and syntax.
ModelWisdom is an integrated toolkit for visualizing, digesting, and repairing TLA+ models, offering improved explainability of counterexamples, interactive state graphs, and automated repair support.
Evolutionary Router Feature Generation for Zero-Shot Graph Anomaly Detection uses mixture-of-experts to generate diverse, transferable router features, addressing distribution shifts and improving zero-shot GAD performance.
Towards a theory of Façade-X data access: satisfiability of SPARQL basic graph patterns investigates Façade-X data access framework, formalizing RDF specialization and analyzing satisfiability of SPARQL BGPs when querying such heterogeneous sources.