Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 11 papers for 2026-02-18
Experiential Reinforcement Learning (ERL) inserts an explicit experience-reflection-consolidation loop into reinforcement learning for language models. It addresses sparse and delayed feedback by having the model generate an initial attempt, then reflect on failures and consolidate improvements for future iterations. This aims to make learning from limited signals more robust and sample-efficient.
DeepImageSearch reframes image retrieval as an autonomous exploration task over visual histories. Traditional multimodal systems optimize semantic matching on isolated images, but this approach accounts for dependencies across temporal visual streams. Models must plan and perform multi step reasoning over raw visual histories to locate targets using implicit contextual cues.
BitDance introduces a scalable autoregressive image generator that predicts binary visual tokens with extremely high entropy, enabling a compact yet expressive discrete representation. To sample effectively from the huge token space it uses a binary diffusion head that performs continuous-space diffusion to generate binary tokens, and introduces next-patch diffusion for decoding.
Nanbeige4.1-3B is a unified generalist language model with 3B parameters that achieves agentic behavior, code generation, and general reasoning in a single model. It is claimed to be the first open-source small language model to reach such versatility. The authors combine point-wise and pair-wise reward modeling to improve reasoning and alignment, and design complexity-aware reinforcement learning rewards to optimize code correctness and efficiency.
REDSearcher offers a scalable and cost-efficient framework for long-horizon search agents. It addresses the bottlenecks of sparse high-quality search trajectories and costly tool-driven interactions by co designing task synthesis, mid training, and post training to optimize search agents across large tasks.
This paper studies Graph Transformers (GTs) with GNN-based positional encodings and builds a theoretical link to Manifold Neural Networks by analyzing manifold limit models for graph sequences. It also extends transferability results to understand how GTs generalize across graphs of different sizes and structures.
This work analyzes the geometric coherence of global aggregation in federated GNNs with heterogeneous client graphs. It shows that global aggregation can numerically converge yet distort relational behavior due to heterogeneity, revealing a geometric failure mode. The paper offers diagnostic insights and directions for more coherent aggregation design.
The authors reframe oversmoothing in deep GNNs as a bifurcation to a stable homogeneous fixed point, using a bifurcation theory lens. They show that stability can be broken by replacing standard monotone activations with non-monotone or alternative activations, mitigating representational collapse. They discuss how incorporating topological priors can further help preserve expressivity as networks deepen.
MRC-GAT introduces a meta-relational copula-based graph attention network for interpretable multimodal Alzheimer's disease diagnosis, addressing fixed structural designs that limit generalization across heterogeneous patient data. The model supports flexible fusion of multimodal information through meta-relational and copula-based dependencies, enabling improved interpretability and cross-patient accuracy.
NeuroLifting uses GNN-based reparameterization to convert inference in large-scale MRFs into a graph-based learning problem, balancing efficiency and solution quality. It enables applying standard neural and optimization tools to MRF inference, improving scalability compared with belief propagation or exact solvers.
Flock proposes a knowledge graph foundation model learned from random walks to support zero-shot link prediction. While conventional KG foundation models enforce equivariance over nodes and relations, deterministic equivariance can limit expressivity; Flock relaxes this constraint to distinguish nuanced structural patterns across KGs with similar structure.
Grappa introduces gradient-only communication for distributed GNN training. Partitions compute in isolation and only gradients are exchanged for the global update, reducing cross-partition communication. To avoid accuracy loss, Grappa periodically repartitions to expose new neighborhoods and applies a lightweight coverage-corrected gradient adjustment.
RUVA presents a personalized transparent on-device graph reasoning framework with a glass-box architecture for human-in-the-loop retrieval-augmented generation. It addresses black-box retrieval concerns and privacy by making retrieval causes inspectable and allowing users to correct errors, enabling on-device personalization.
NeuroSymActive presents a differentiable neural-symbolic reasoning framework with active exploration for knowledge graph question answering. It integrates neural reasoning with symbolic KG structure to enable efficient, multi-hop QA, and uses active exploration to locate relevant subgraphs and refine answers.
SIGMUS proposes semantic integration for knowledge graphs in multimodal urban spaces to fuse diverse sensor modalities and reason about incidents. It addresses data fragmentation and reduces reliance on human-driven reasoning when linking multimodal data to events.
Embedding Retrofitting examines how embedding retrofit relies on knowledge graph constraints and text preprocessing, and highlights that KG quality is critical. The paper offers a data engineering framework to mitigate annotation artifacts, improving retrofitting reliability. It shows that hashtag annotations can inflate KG density and create spurious edges that degrade retrofitting performance.