Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 15 papers for 2026-09-11
Nothing cleared the bar today. Papers read in full today: 6.
The load-bearing idea is a clean representational reframing: replace dataset-specific relation embeddings with anonymous relation-type nodes and six fixed meta-relations, then test whether transfer survives after removing bespoke relation-graph architecture. The full text does more than report a single benchmark win: it evaluates five backbones on 40 splits, exposes the ULTRA released-checkpoint versus published-score discrepancy, and diagnoses a concrete failure mode—plain candidate-only readout ranks isolated entities catastrophically—while finding GAT trails controlled ULTRA by 0.0167 MRR overall and is effectively tied only on a post hoc, NELL-dominated 13-split mutually-unseen-family subset.
The headline “matches ULTRA” is too strong: ULTRA leads on the primary all-40 comparison, newer KGFMs are stronger still, the symmetric subset was selected after results were known, and the database result is only a two-database featureless probe that does not test the proposed table/class vocabulary or compare against relational foundation models.
The load-bearing contribution is a set of exact structural and bifurcation results: stochastic normalization prevents a heterogeneous eigenmode from destabilizing before the homogeneous fold, while weakly coupled equitable blocks turn scalar bistability into combinatorial multistability. The full text does deliver these claims with direct contraction/spectral/implicit-function arguments and coefficient-level numerical checks for three threshold laws; the strongest new result is the two-block heterogeneous-cusp scaling, not the quotient reduction itself.
Its conclusions hinge on deterministic, row-stochastic coupling and exact equitability (with only local persistence after perturbation), provide no stochastic binary-agent derivation or broad network-class evidence, and therefore do not establish topology-driven tipping behavior for realistic heterogeneous networks.
This work identifies a fundamental scaling tension in AutoResearch agents: the agent generation and environment execution scale very differently. Generation relies on batched compute, while environment interactions are sequential and often bottlenecked by environment runtimes, creating inefficiencies as models scale. The paper analyzes implications for RL-based scaling in AutoResearch and discusses potential remedies and design considerations.
Show-Harness presents an Embodied Harness that enables VLMs to 'play' robots through a compact semantic interface linking intent to action. It exposes discrete semantic action units that the VLM can reason over, while embodiment-specific interpreters deterministically ground them into local robot actions, keeping the VLM responsible for high-level decisions and the concrete actions handled by the grounders. The approach provides a unified interface to bridge vision-language understanding and robotic control.
This paper critiques two stages of textual gradient MAS prompt optimization: gradient extraction and gradient aggregation. It argues that prior methods often select a target prompt without verifying whether its modification actually resolves failures, and that gradient aggregation can dilute useful signals. It proposes improvements to more reliably identify which prompt to modify and how to combine gradient information for better MAS prompting.
Programmable World Model decouples world-state evolution from visual observation generation, addressing persistent state and rule enforcement. An agent translates natural-language instructions into executable programs that specify states and state-transition rules for entities, enabling direct control over interactions. A lightweight engine runs these programs to simulate and render the resulting dynamics and interactions.
T1 introduces a 122B Mixture-of-Experts reinforcement learning agent designed for long-horizon terminal tasks, operating a real shell in a cloud sandbox and handling 300+ tool calls per task, rewarded by the task verifier. The paper provides a recipe including an aggressively warm-started training to stabilize actor-critic optimization and dense rewards based on the number of verifiers passed, along with stable optimization strategies.
We introduce quantum-inspired structural features for learning on signed graphs by encoding them as Ising models and extracting density-of-states moments. These moments count signed closed walks and are switching-invariant and size-free, enabling learning tasks such as predicting the frustration index on signed networks.
We develop a GNN-based estimator to assess relative algebraic-connectivity loss after multi-edge deletions, blending spectral theory with learning. The approach provides a bounded correction to a first-order Fiedler sensitivity and is tested under independent, spatially clustered, and betweenness-targeted failures, with zero-shot transfer to multiple real-world areas. It facilitates rapid evaluation where exact spectral recomputation is costly.
We present a variational, physics-informed graph neural network (PI-GNN) for heterogeneous solid mechanics, where bimaterial interfaces cause stress jumps while displacement remains continuous. The variational formulation integrates physics priors without relying on ad hoc width regularization, improving robustness to phase-contrast handling on heterogeneous materials.
Kernel-Complexity Edge Sanitization (KCES) is a training-free, model-agnostic defense against structural graph attacks. It leverages kernel-complexity principles to sanitize manipulated edges without retraining, offering a theoretically motivated, practical protection for GNNs.
HERALD introduces high-fidelity exemplar retrieval with adaptive landmark distillation for heterophily-aware graph condensation. It provides a gradient-free framework that adapts node scoring and feature learning to produce a compact surrogate graph that preserves downstream node classification performance under heterophily. The method targets retaining essential structural information while significantly reducing graph size.
LiteRAG introduces a cost-efficient retrieval-augmented generation framework that replaces expensive retrieval-time LLM control with query-conditioned algorithmic exploration and reasoning-chain construction. It achieves high quality on multi-hop retrieval benchmarks while reducing context and compute load.
We propose using directed state graphs to encode task connectivity within goal-conditioned hierarchical reinforcement learning (GCHRL). The graph is treated not just as a sampling tool but as an environmental model that captures connectivity and state-accessibility, enabling dense reward learning and more informative subgoal sampling. This approach aims to leverage topology for more efficient learning in complex tasks.
Multi-Agent Agentic Graph Learning (AGL) uses multiple role-based agents powered by LLMs to sequentially sample a graph as evidence. Unlike single-policy approaches, heterogeneous agents tailored to different graph regions enhance reasoning and prediction over complex graphs.
Motif-oriented graph captioning is studied as a bidirectional graph-text translation task, where captions must preserve topology for graph recoverability and express recognizable motifs (hubs, paths, cycles, cliques, bridges). This yields captions that both describe structure and enable reconstruction.
EXYGEN is a framework for scalable knowledge graph understanding that enables conversational access using LLMs. It leverages automatically derived structured metadata (VoID, ShEx) in a retrieval-augmented generation pipeline for text-to-SPARQL generation and evaluation on the SciQA benchmark.
SynCo presents synthetic, community-aware attributed graph generators to benchmark GNNs, addressing the lack of high-quality datasets with ground-truth communities. By controlling community structure and attribute distributions, it enables rigorous evaluation of community detection and related graph tasks across synthetic graphs.
SAFA-MZ proposes distributed physical-layer authentication and collaborative RSMA for multi-zone non-terrestrial networks, maximizing secrecy spectral efficiency while ensuring authentication. The framework emphasizes privacy-preserving, joint authentication and communication design aided by graph-based reinforcement learning.
A causal meta-learning framework for distributed physical-layer authentication and attack detection in 6G NTNs (DPLA) is proposed to combat distribution shifts due to Doppler, delays, and channel variations. It designs a multi-feature fingerprint and leverages causal meta-learning for robust PLA under unseen environments.