Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 25 papers for 2026-10-08
Nothing cleared the bar today. Papers read in full today: 11.
The paper does deliver a substantive mechanism: ON-Loss is not just a basis change, but is backed by worst-case observable bounds and strong empirical evidence that NON-basis supervision misaligns with energies/forces, while GROOT enforces valid idempotent densities by construction and ROCKET strips away gauge directions that do not affect the occupied subspace. The key concrete finding is that these choices massively improve size extrapolation (QM9→QM40 and QMugs→660 electrons) and make label-free self-consistency fine-tuning work on Transition1x, so this is more than a benchmark tweak.
The evaluation is still confined to closed-shell B3LYP/def2-SVP and one backbone family, so the scope of the claims is broader than the tested regime.
The authors revisit Cross-Tokenizer On-Policy Distillation (OPD) and examine whether widening alignment coverage between teacher and student improves learning. Across three heterogeneous teacher–student pairs in mathematical reasoning and code generation, they find that strict 1:1 token groups already cover most student-generated tokens despite large vocabulary mismatches. Analyses on responses sampled from pre-distillation students show a shared vocabulary that remains sufficient for supervision, arguing that supervision reliability matters more than expansive alignment.
TRACE (Train-Rollout Quantization Alignment via Compact Guidance) proposes an FP4 quantization framework for reinforcement learning of MoE language models. It directly aligns the quantized train and rollout execution paths rather than optimizing them separately, addressing a key limitation of prior FP4 RL methods. By reducing the discrepancy between training and rollout paths, TRACE enables more efficient RL training at ultra-low precision.
DuoMatching advances distribution matching for few-step video generation by introducing a unified joint-marginal formulation beyond joint distribution matching. While prior DMD reduces drift by matching the joint distribution to a teacher, DuoMatching adds marginal distribution matching to better approximate the real video distribution, improving visual quality and semantic alignment in short-horizon generation. This unified formulation helps curb error accumulation during autoregressive generation.
EVISKILL grounds skill evolution in replayable evidence for continual LLM agent improvement without updating model parameters. It emphasizes not only what to change and why, but when to make changes persistent, preserving behavioral evidence and task context. The approach addresses problems where existing experience-driven methods lose evidence or rely on global validation that may discard locally well-supported edits.
From Evidence to Action investigates failures in the evidence-to-action chain for tool-using agents. Across ten model-harness configurations, there can be strong static assessment of actions alongside weak interactive execution, with failures often occurring before execution due to incomplete investigation or acting before evidence is established. The work highlights the breakdown points where evidence gathering and action execution diverge.
Edge Accuracy Is Not Enough shows that transferring structure learned from dynamics to inverse problems can fail despite reasonable edge-error bounds. The paper proves that approximate structure improves estimation when the edge error Δ satisfies Δ < n^2 - k n, reducing sample complexity from O(n^2) to O(k n + Δ). The results reveal when learned structure helps and its limits with NRI-driven methods.
Oscillatory Neural Dynamics over Sheaves (ONDA) introduces operator-valued information waves for long-range graph propagation. Stalk-valued representations evolve with second-order dynamics to enable distant nodes to influence each other effectively. The framework enhances long-range communication while preserving locality.
Pathwise Information Certificates present a decentralized adaptive sensing framework with a pathwise certificate based on Renyi-Chernoff information along sensing trajectories. It yields nonasymptotic MAP-error bounds and an anytime network-wide stopping rule. The results provide principled stopping criteria for adaptive sensing under communication constraints.
Reliability of LLM Judges for Evaluating Entity Alignment studies how well LLM-based judgments align with human labels across multiple EA datasets and systems. The study analyzes agreement, variability across models, and guidance for reliable use of LLM judges in structured prediction. It provides best practices and findings about reliability.
GraphOPD augments on-policy distillation for LLM agents by injecting graph-structured guidance that propagates through decision sequences. It addresses the limitations of step-level guidance based on teacher-student divergence when decisions become chained. The method improves sample efficiency and performance in sparse-reward, multi-turn settings.
Neighborhood Smoothing introduces train-time calibration via graph smoothing in representation space, encouraging neighboring samples to have similar predictive distributions. The authors derive bounds connecting predictive divergence to neighborhood structure and analyze effects on calibration metrics. Experiments show improved reliability and calibrated probabilities.
Distilling Graph Geometry investigates how to transfer the graph-induced geometry from a GNN teacher to an MLP student. It reveals spectral failure modes when geometry is not preserved and shows that node representations can underfit high-energy directions. The work provides guidance for geometry-preserving distillation.
EDiS decomposes a graph once into edge-disjoint subgraphs and then recombines them across epochs under budget constraints. This separates one-time structural extraction from per-epoch graph construction, enabling diverse sparse topologies without expensive recomputation. The framework achieves fast sparse GNN training while maintaining performance.
TopoGraphRAG-Bench evaluates multimodal GraphRAG on layout-grounded evidence reasoning. It emphasizes recovering evidence topology across text, tables, figures, and captions in complex documents. The benchmark targets evidence topology recovery and cross-modal reasoning for real-world documents.
Know the Shape, Find the Fault proposes topology-conditioned diagnosis for failures in multi-agent LLM systems. By analyzing communication topology, the work shows how exchange patterns differentiate coordination failure modes, and develops methods to infer the relevant structure from execution traces to diagnose issues when traces are incomplete.
HCPN-GCN introduces Cone geometry into Hierarchical Prototype Networks to scale continual graph learning. It uses a prototype memory that captures graph topology without storing raw historical data, addressing catastrophic forgetting. The cone-based prototypes enable efficient clustering of topological patterns and improve transfer across tasks.
CircuitGate provides a logic-consistent circuit-level functional model for And-Inverter Graphs. Moving beyond gate-level message passing, it captures higher-level functional context to support reliable synthesis and verification. The approach improves robustness to functionality-preserving transformations and yields better predictive accuracy on circuit tasks.
We redefine forecasting with sparse sensors as a conditional generation task on spatio-temporal graphs to predict unobserved node states (FUNS). The approach uses LLM-guided guidance to infer missing node states from contextual information and graph dynamics, reducing reliance on historical data. This formulation enables end-to-end generation of plausible node states for nodes without prior records, improving forecast accuracy.
Continual Graph Multi-Agent Reinforcement Learning studies how to reuse structural knowledge across sequences of tasks with different graph topologies. The method leverages structural information to enhance transfer while mitigating forgetting in CMARL settings. It shows improved performance when tasks share structural patterns.
We propose a Node-level Graph Neural Architecture Search framework that assigns different aggregation operations at different nodes. The search exploits node features and local topology to select suitable message-passing modules, mitigating over-smoothing and improving accuracy. The framework reduces the need for uniform convolution across all nodes.
MovieSTAGE is a multiscale encoding framework for movie-fMRI ADHD classification. It combines scene-level hypergraph FC within scenes, differences across adjacent scenes, and whole-movie FC into a unified representation. On a 260-subject dataset, it improves classification performance by aligning neural dynamics with narrative structure.
Shared-Roadmap Generation and Evaluator uses a heterogeneous GNN to generate and evaluate shared roadmaps for multi-agent path planning. By leveraging graph heterogeneity, it balances roadmap density and feasibility. The evaluator component assesses solution quality under dynamic, multi-agent constraints.
BEACON-SP is an ontology-grounded Graph Retrieval-Augmented Generation framework for clinical suicide risk assessment. It combines patient knowledge graphs with ontology-guided retrieval to support multi-hop reasoning about diagnoses, medications, risk and protective factors, life events, and temporal relationships. The framework supports clinician-facing decision support with explainable, evidence-backed outputs.
This paper introduces HGP, a hybrid graph memory framework for on-device personalized agents. It tackles challenges in routing and retrieving heterogeneous long-term traces, addressing limitations of single-vector memory representations. By using a lightweight self-enhancement classifier for personalized memory routing and constructing episodic and semantic memories (and relational structure), HGP enables on-device personalization and more accurate retrieval.
CircuitATLAS presents a provenance-grounded systems-neuroscience knowledge graph and an agentic reasoning framework for target discovery in circuitopathies. Rather than focusing solely on disease-altered molecules, it identifies unaltered molecular control points to restore pathological neural circuits toward functional states. The graph integrates literature-derived relationships across circuits to support principled target discovery and mechanistic reasoning.
Automatically Building and Updating a Knowledge Graph of MLIP Models describes an LLM-based pipeline to extract, validate, and maintain a knowledge graph of machine learning interatomic potentials. It includes SHACL-based validation loops to detect and correct errors and demonstrates automatic updating as the field evolves.
This dissertation outlines a scalable architecture for turning unstructured domain-specific text into structured knowledge for retrieval and reasoning. It integrates semi-automatic corpus curation, semantic structuring, retrieval, and inference into an interpretable pipeline. It also introduces Binary Bleed to accelerate low-rank search in NMF and Hierarchical NMF with automatic latent feature selection for depth-adaptive topic modeling.
RiftANN introduces an efficient graph traversal-based ANNS system designed for passive RDMA-based memory disaggregation. It addresses the inefficiencies of conventional best-first search by reducing remote reads and decoupling candidate evaluation from traversal expansion, enabling scalable, low-latency vector search across memory-disaggregated setups.
KGATE presents a Knowledge Graph Embedding Training Environment that supports end-to-end autoencoder-based KGEs. It bundles encoder/decoder components, tooling, and reproducible hyperparameters to streamline development and evaluation. The environment aims to improve usability and reproducibility of KGE research.