Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 19 papers for 2026-09-18
Nothing cleared the bar today. Papers read in full today: 9.
The load-bearing idea is a useful graph-structural reframing: message-passing depth fixes an exact k-hop Hessian sparsity envelope (K=2L for MACE and 2L+1 for PET), so derivative cost is a direct consequence of graph receptive field rather than merely an implementation issue. The paper delivers a particularly valuable negative result—exact ASD is only 1.2–1.7× faster at the median, at most 6×, even on porous-material best cases—then shows that truncated ASD can preserve 300 K heat capacity within 0.1% with median 11× speedup for MACE and 13× for PET-S; this is a convincing, actionable result for MLIP derivative computation rather than a routine new GNN.
The empirical claim is narrow—three short-range foundation MLIPs, mainly porous solids, and essentially one harmonic observable (heat capacity)—with no theory or robust observable-dependent rule for choosing truncation, while zeolites and PET-XS already show that the favorable tradeoff can disappear.
We present LimiX-2, a scaled-up model in the LimiX family built under our scaling laws. It adopts Contextual Mechanism Networks (CMNs) and is pretrained with Context-Conditional Masked Modeling (CCMM). CMNs reframe in-context learning from predicting y given x and context to jointly modeling the context-dependent distribution over x and y, enabling mechanism-oriented representations.
We introduce ScienceIDE, infrastructure for turning the world’s scientific codebases into programmable environments for scientific agents. The platform is guided by expert-defined scientific cases and acceptance criteria to ensure useful learning experiences. Agents transform repositories into executable environments that support experimentation and reasoning.
We analyze PPO critics in RL for large language models and identify Value Flattening: state values estimated from multiple Monte Carlo continuations vary sharply across intermediate states, while critic predictions remain flat. The effect appears in controlled settings like FrozenLake and grows with larger state spaces. The paper provides theoretical and empirical insights into this flaw and discusses possible remedies.
We argue that existing confidence estimators only read the current inference, which is insufficient for reliable trust decisions. We propose XConf (eXperiential Confidence), a method to estimate confidence jointly with the model’s reasoning process and accumulated experiential evidence, improving calibration and robustness of deployed systems.
ProgramDistill introduces a benchmark for evaluating coding agents on features discovered by interacting with fully functional reference applications. It factorizes applications into features at various granularities, each paired with replayable behaviors executable via gold patches. The pipeline mine-craft-patch guides discovery, implementation, and patch-based verification.
GAVEL presents Graph World Models for verified and efficient long-horizon LLM task planning. It builds an explicit graph world modeling object relations, action preconditions/effects, and probabilistic beliefs to predict consequences and verify/repair long-horizon plans.
The paper investigates how spatial dependencies are encoded in graph attention based short-term traffic forecasting models. By enhancing interpretability of gated graph attention networks, it attributes influence to neighboring locations and analyzes which components are essential for accurate predictions. The study proposes evaluative methods to explain attention dynamics and verify spatial influence against measured data.
Not All Nodes Are Created Equal argues that evaluation splits in GNNs can be unstable due to graph homophily. It proposes homophily-aware stratification for cross-validation to ensure test folds reflect the dataset's label distribution and connectivity properties, improving reliability of comparisons.
VisKG-LM shows that knowledge graphs retrieved for QA can be compiled offline into a visual memory, decoupling graph encoding from online LM inference. The approach reuses the same subgraph across training runs and seeds, avoiding re-encoding while the knowledge graph remains static.
RAFT introduces a Stateful Retrieval-Augmented Framework for Troubleshooting Agents that models each historical case as a timeline of entries and retrieves at the entry level, matching intermediate states to surfaced guidance.
Labeled Incidence Structures LIS provide a uniform representation for text, knowledge graphs and hypergraphs by encoding each endpoint as (x_d, s, e). This preserves content, role and relation without flattening, enabling a single Transformer to process all data types natively and preserve structural information.
Beyond Similarity through Zero-Token Geometric Graphs for Multi-Hop RAG presents G3RAG a document only offline graph framework that constructs a geometric edge graph without LLM calls or token generation Edges carry geometric gain to enable multi hop reasoning while reducing reliance on dense, similar document retrieval.
JointMatch unifies heterogeneous graph neural solver for large-scale ride-sharing matching. It tackles both bundle pairing and vehicle assignment in a single framework rather than sequential stages, improving revenue and scalability.
Learning Submanifolds for Inference on Random Dot Product Graphs Part 1 Theory develops a framework for restricted inference when latent positions lie on unknown low-dimensional manifolds. It uses Isomap to learn a low-dimensional representation and designs semisupervised decision rules that map configurations to actions, analyzing their behavior.
GS-Power-UCT introduces a graph-based Monte-Carlo Tree Search variant that shares states reached at the same planning depth while keeping depth-specific values. This enables efficient planning in stochastic MDPs, including problems with cycles. The authors also show convergence of the root estimate to the finite-horizon value as the number of simulations grows.
SCGFM-ART introduces a structure centric graph foundation model framework that aligns arbitrary graphs onto a shared relational atlas using Amortized Relational Transport ART. The relational atlas provides a universal coordinate system to reduce domain shifts across graphs. This enables transferability of representations across heterogeneous graph domains.
HyperAMS-Net proposes Adaptive Multi-Scale Spatial Hypergraph Network for Brain Disorder Classification. It integrates adaptive multi-scale convolution, hypergraph attention, and spatial-channel interactions to model multi-scale brain connectivity from resting-state functional and structural MRI, addressing inter-subject heterogeneity.
Quantum Graph Convolutional Networks implements two representative quantum-inspired GCNs SGC and LGC, and analyzes their trainability on quantum primitives. The work provides implementation details and assesses scalability for large graphs.
AI-Driven Real-Time Relay Optimisation in Smart Urban NR-V2X Networks via Learning-to-Optimise Graph Neural Networks proposes a L2O framework using GNNs for real-time multi-hop relay selection in NR-V2X networks. The vehicular graph model optimizes relay decisions under latency and reliability constraints.
TRACE Accountable Agentic Retrieval for Source Discovery in Digital Archives presents a training-free agentic retrieval framework for accountable source discovery over historical corpora. It is designed to handle OCR degraded and heterogeneous archives and emphasizes source traceability for scholarly use.
WiCleanData automates taxonomy refinement and type constraint enforcement to guarantee Wikidata type consistency. It designs an automated pipeline that cleans taxonomy and enforces constraints, reducing inconsistencies and complexity.
From Rollout to Reset A Graph-Based Harness for Autonomous Long-Horizon Manipulation Evaluation introduces HALTER a graph-based harness that automates reset and evaluation for long-horizon manipulation, improving reproducibility and reducing operator time.
AURORA is a natural language driven agentic framework for understanding reasoning and orchestrating reliable air ground co-simulation. It treats scenario generation as compilation with verification, using an Air-Ground Scenario Graph AGSG to encode relationships and validate user requests.