Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 11 papers for 2026-09-16
Nothing cleared the bar today. Papers read in full today: 4.
The load-bearing contribution is a nontrivial backward-transfer theorem, L(G) ≡ 3-WL L(H) ⇒ G ≡ 3-WL H, combined with Shrikhande/4×4-rook witnesses proving strictness. The paper tests the resulting prediction rather than merely reporting accuracy: ILG-3-WL separates all SR25 pairs and 359/400 BREC pairs versus 0/105 and 270/400 for root 3-WL, and it gives a concrete universal consequence—one global pair-refinement round determines the exact K4 count.
The result is a sharp but specialized expressivity statement—restricted to Whitney-general/componentwise Whitney-general graphs, with no downstream learning evidence and potentially unfavorable m² pair-state cost on dense line graphs—while the key long appendicial proof is not independently assessable from the supplied main text.
Vidu S2 introduces two components: Vidu S2-Avatar, a real-time interactive digital-character model, and Vidu S2-Editing, a real-time video editing model. The system enables real-time spatial video generation and supports updates to dynamic references during generation. Compared to Vidu S1, S2-Avatar delivers real-time 720p video, stronger instruction-following (e.g., dancing), and dynamic reference updates; S2-Editing can edit a streaming video in real time, including style rendering and clothing replacement, with other character editing capabilities.
Atria Dawn Preview is a foundation agentic language model designed for scientific research and engineering workflows, aiming to boost agent productivity in real-world settings. It uses a Verifiable Experience Pipeline that links tool-mediated interactions to executable environments and externally verified outcomes, enabling traceable progress across tasks. It is evaluated across 16 benchmarks spanning real-world research, engineering, and digital environments.
ZGCM-1 is a fully open 7B dense foundation model trained from scratch with high data, system, and algorithmic efficiency. It argues compact models can't rely on memorization alone; by coupling internal reasoning with active external tool use, it can operate with a 256K context. The training recipe emphasizes architecture and system co-design with interleaved gated sliding-window and full attention, plus a stable FP8 Muon optimizer; progressive training methods are proposed.
Dream-RSI presents a framework for recursive self-improvement through evolving worlds, addressing the bottleneck of exploration. It targets scalable and recursive self-improvement for autonomous AI agents by crafting adaptive exploration strategies that can improve as search spaces grow, avoiding fixed heuristics and the inefficiencies of online policy optimization with delayed feedback. The approach uses evolving environments to continually surface high-value exploration opportunities.
PhysBrain 1.5 introduces a unified model that understands physical environments, generates actions and predicts future states, unifying vision-language reasoning with embodied control. It encodes language responses, end-effector motions, and dense visual targets as discrete sequences and trains them jointly with autoregressive next-token prediction. Pre-training uses embodied supervision solely from human interaction videos, in a task-centered setup.
This paper investigates structural negative transfer in federated graph neural networks caused by heterogeneity in client graph structures, beyond label or feature distribution differences. It provides diagnosis methods and causal investigations, and reveals the limits of divergence-aware mitigation strategies in mitigating such transfer.
The paper explores whether interpretable NLI can be achieved with graph-based representations built from atomic propositions. It proposes a fully graph-based pipeline where sentences are decomposed into atomic propositions and converted into ConceptNet triples, forming three graphs per pair for NLI reasoning.
The authors address the limited discriminative power of GNNs due to the 1-WL bottleneck and propose repurposing Unified Topological Signatures to capture multi-scale global topology for graph representations. This approach aims to surpass local neighborhood information and improve graph discrimination.
RECAL is an unsupervised framework addressing relation heterogeneity in provenance-based intrusion detection by balancing relation frequencies and calibrating error levels across relations. This approach reduces false alarms and missed detections in PIDS settings with highly uneven relation distributions.
The authors propose a Unified Heterogeneous Graph Neural Network solver that uses a single Heterogeneous Residual Gated Graph Convolutional Network to solve Power Flow, Optimal Power Flow, and State Estimation with one shared backbone, emphasizing generality across tasks.
End-to-end cell detection is proposed via instance-aware graph modeling, moving beyond two-stage pipelines. The approach models spatial and relational interactions among cell nuclei within the tumor microenvironment to jointly detect and classify cells.
This study introduces a hypergraph-based extractive summarization method that leverages domination properties to select key sentences from a single document. It compares the proposed domination-based hypergraph approach with state-of-the-art graph-based methods to benchmark performance.
GPEvac presents a GNN-based PPO framework for adaptive evacuation routing during shooting events, aiming to minimize threat exposure while handling adversarial uncertainty and crowding dynamics. The approach seeks to deliver real-time, layout-flexible evacuation policies that scale to large environments.
This work introduces a virtual imaging trial framework with reinforcement learning to optimize CT protocols. By evaluating 468 parameter combinations across 63 computational models, the approach aims to improve diagnostic image quality while reducing radiation dose.
netseg is a Python package implementing a wide range of polarization and segregation indices for social networks. It supports more than two groups, directed and undirected graphs, and ports functionality from an R package with thorough documentation and testing.