Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 17 papers for 2026-09-30
Nothing cleared the bar today. Papers read in full today: 5.
The core idea is clean: replace dense all-pairs attention in graph ICL with sparse message passing, then pretrain on a richer synthetic graph prior. The full text does support the efficiency claim well—runtime scaling is near-linear and substantially faster than GraphPFN—and the main results show Ephris beating tuned GNNs and prior graph GFM on 51 datasets, so this is more than a cosmetic rewrite of TabPFN-style ICL.
The gain is still mostly benchmark-level: the method is a fairly elaborate combination of existing pieces, and the paper does not yet show a new theoretical understanding of when graph ICL should beat tuned GNNs or how robust the synthetic prior is under harder distribution shift.
The central idea is load-bearing: fixed-prefix training is a bad match for non-stationary temporal KG streams, and the paper tests that claim by comparing static training, naive fine-tuning, replay, regularization, and HiTS-CL across five backbones. The full text does support the claim that continual updating materially reduces long-horizon decay and usually beats adapted CL baselines, but the gains are mostly from the continual protocol plus a simple history cache, not from a fundamentally new reasoning mechanism.
The history-enhancement module is a lightweight recency/frequency cache and the distillation step is a fairly standard teacher-student recipe, so the method is more a strong protocol correction than a new graph-learning primitive.
The paper shows that post-training can leave a behavioral shadow on unrelated decisions by transferring capabilities via task-unrelated text. It introduces Active Taskless Distillation (ATD), which enables capability transfer using only a single word from the teacher per prompt. ATD is used to probe the post-training shadow by selecting prompts where teacher and student signals align, revealing how updates propagate through generations.
YuE2 unifies symbolic and audio music generation at frontier quality by using symbolic planning. A single AR-NAR Mixture-of-Transformers writes a readable score specifying melody and harmony, expands it into semantic tokens, and realizes it as full-song audio. In head-to-head comparisons at the same checkpoint, experts prefer symbolic planning for overall quality and musicality.
Domain-normalized Multi-Teacher On-Policy Distillation (DN-MOPD) studies how to merge multiple domain specialists into a single student via on-policy feedback. The routing chooses which specialist teaches for a given prompt, but existing methods do not properly normalize how strongly that feedback moves the student. Experiments on Qwen3.5 across three model sizes show that domain-normalized distillation yields better multi-domain performance and stability.
Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL (GAGAR) addresses the issue that executable tests provide binary rewards and that prior methods like GRPO treat all passing trajectories in a rollout group as equally valuable. This ignores differences in implementation quality and task adherence. GAGAR introduces quality-aware credit redistribution with dynamic sampling to reward better code quality and targeted behavior.
Self-Evolving Coding Agents propose a pattern where agents maintain explicit state and use executable-code-style loops to revise actions based on feedback, enabling robustness in physical-world tasks. Building on digital coding agents, LLMs call tools, verify results, and revise actions as executable code. This framework aims for more interpretable, adaptable agents for vision-language-action and world-action tasks.
Measuring trainable degrees of freedom in materials GNNs: a random-subspace intrinsic dimension analysis. The authors introduce trainable-degree dependence and random-subspace intrinsic-dimension analysis to study how much of the parameter space is needed. They compare CGCNN, ALIGNN, and DimeNet++ across six prediction tasks, showing how much of the parameter space is truly needed for good performance.
Better Nearest Neighbor Graph Indices via (Efficient) LLM-Guided Pruning. Graph-based approximate nearest neighbor search (ANNS) indices are usually built from geometric relationships rather than semantic relevance, causing a geometry-semantic mismatch during downstream retrieval. The authors propose an efficient LLM-guided pruning approach to align indices with semantic objectives, improving downstream semantic retrieval with limited geometry loss.
ImbalancE tackles degree imbalance in link prediction at inference time. It finds that the target entity degree in test triples heavily influences prediction quality, with low-degree targets harder to predict. The paper proposes an inference-time latent search to mitigate this imbalance and improve link prediction accuracy.
Luxury resale volume and composition shifted during the COVID-19 tourism shock, but co-purchase network tests were inconclusive and construction-sensitive. The study analyzes Japan's 2020 second-hand luxury resale data, comparing 2019 and 2020 brand co-purchase projections using Holm-corrected metrics and Portrait Divergence; results do not clearly separate years and are sensitive to design choices.
Spatiotemporal Hyperedges for EEG Seizure Detection and Prediction. Seizures are rare and propagate across channels; existing methods use sequences of pairwise channel edges but miss spatiotemporal coupling and are costly. The proposed HyBrain summarizes evidence with a small set of soft spatiotemporal hyperedges, reducing training cost while capturing seizure phenomena.
Graph-Split Bayesian Causal Forest for Spatial Heterogeneous Treatment Effect Estimation. It addresses spatial dependence and spatially varying treatment effects using a Graph-Split Bayesian Causal Forest, enabling robust estimation of HTE in spatially complex settings.
TSG Suggester: Tree-Structured Knowledge-Graph Retrieval for Troubleshooting Guide Recommendation in Cloud Incident Management. The system retrieves relevant troubleshooting guides directly from incident descriptions, evaluating five retrieval strategies—Text-Only RAG, Image Augmented RAG, RAPTOR, Tree Structured Retrieval, and the proposed method—and demonstrating effectiveness in guiding engineers.
This paper proposes Mycelium, a generalizable cross-grid multi-task model for electrical distribution systems. It defines a unified grid ontology and represents variable-sized networks as heterogeneous graphs that preserve topology, asset types, and electrical relationships, enabling diverse physics-grounded tasks and better generalization to unseen networks.
Dagger: Decoupling-based Model Stealing Attack against Graph Neural Networks. It presents a decoupling-based attack that can construct a functionally equivalent surrogate model from a victim GNN API under weaker assumptions than prior work, highlighting security risks for GNN-based services.
GeoOutageBench benchmarks LLM-based geospatiotemporal KGQA for outage and resilience analysis. It constructs a spatiotemporal knowledge graph that integrates outage records, remote sensing, weather observations, storms and power events, geographic entities, and domain ontologies, including multimodal data. It provides a competency query taxonomy at increasing difficulty levels, addressing spatiotemporal containment, proximity, co-occurrence, and related queries.
Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis. The work addresses high-order correlations and long-range dependencies in brain networks, noting redundancy across information types. It introduces an Information Bottleneck-guided Adaptive HyperGraph Transformer (IBAHGT) to integrate high-order and long-range information efficiently for brain disease diagnosis.
Graph neural networks for sampling-invariant embeddings of organized signal sets. The paper studies encoders for organized signal sets where signals have heterogeneous sampling parameters, aiming to project these into a fixed-size vector space invariant to sampling. This representation supports robust downstream tasks on signals organized on grids or graphs.
PowerZooJax is a JAX-based reinforcement learning benchmark suite for power-system operation. It provides five constrained Markov decision process tasks spanning generation, transmission, distribution, and distributed energy resources, designed for scalable evaluation on modern accelerators and to overcome CPU-based simulation bottlenecks.
Concealing LLM-Based Multi-Agent Topology via Phantom Structure Injection. As LLM-based MAS reveal topology through observable traces, we propose phantom structure injection to conceal the true topology; this mitigates topology leakage risks.
BrainNet Studio is a unified toolkit for brain network construction, analysis, and visualization. It supports static and dynamic brain networks and integrates dynamic modeling with modern graph and sequence learning methods to streamline brain network research.