惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

博客园 - 司徒正美
大猫的无限游戏
大猫的无限游戏
腾讯CDC
J
Java Code Geeks
博客园 - 【当耐特】
Microsoft Azure Blog
Microsoft Azure Blog
V
Visual Studio Blog
人人都是产品经理
人人都是产品经理
博客园 - Franky
博客园 - 聂微东
阮一峰的网络日志
阮一峰的网络日志
美团技术团队
云风的 BLOG
云风的 BLOG
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
U
Unit 42
雷峰网
雷峰网
B
Blog RSS Feed
博客园_首页
量子位
F
Fortinet All Blogs
罗磊的独立博客
H
Hackread – Cybersecurity News, Data Breaches, AI and More
酷 壳 – CoolShell
酷 壳 – CoolShell
C
Check Point Blog

cs.SI updates on arXiv.org

Hiding in Plain Sight: Finding MAHA on Reddit Prism: Structural Symmetry Scanning via Duality-Constrained Laplacian Projection MV-Gate: Insider Threat Detection via Multi-View Behavioral Statistics and Semantic Modeling Algorithmic Cultivation: How Social Media Feeds Shape User Language Universal Dynamics of Punctuated Progress AI-Mediated Communication Can Steer Collective Opinion CitePrism: Human-in-the-Loop AI for Citation Auditing and Editorial Integrity Explainable Detection of Depression Status Shifts from User Digital Traces Can Visual Mamba Improve AI-Generated Image Detection? An In-Depth Investigation ScioMind: Cognitively Grounded Multi-Agent Social Simulation with Anchoring-Based Belief Dynamics and Dynamic Profiles Humanwashing -- It Should Leave You Feeling Dirty When Do LLMs Generate Realistic Social Networks? A Multi-Dimensional Study of Culture, Language, Scale, and Method Moltbook Moderation: Uncovering Hidden Intent Through Multi-Turn Dialogue Linking Extreme Discourse to Structural Polarization in Signed Interaction Networks Predicting Channel Closures in the Lightning Network with Machine Learning Latent Causal Void: Explicit Missing-Context Reconstruction for Misinformation Detection Predictive Maps of Multi-Agent Reasoning: A Successor-Representation Spectrum for LLM Communication Topologies Large Language Models for Causal Relations Extraction in Social Media: A Validation Framework for Disaster Intelligence When Can Digital Personas Reliably Approximate Human Survey Findings? RAwR: Role-Aware Rewiring via Approximate Equitable Partition GravityGraphSAGE: Link Prediction in Directed Attributed Graphs Structure-Centric Graph Foundation Model via Geometric Bases Attention-based graph neural networks: a survey When AI Meets Science: Research Diversity, Interdisciplinarity, Visibility, and Retractions across Disciplines in a Global Surge Scalable inference of spatial regions and temporal signatures from time series Can LLMs Emulate Human Belief Dynamics? Predicting Post Virality with Temporal Cross-Attention over Trend Signals H3: A Healthcare Three-Hop Index for Physician Referral Network Prediction Dynamic Graph with Similarity-Aware Attention Graph Neural Network for Recommender Systems Spectral Graph Sparsification Preserves Representation Geometry in Graph Neural Networks
LLM Features Can Hurt GNNs: Concatenation Interference on...
Zhongyuan Wang, Pratyusha Vemuri · 2026-06-16 · via cs.SI updates on arXiv.org

Adding LLM-generated node features to graph neural networks (GNNs) is widely reported to improve accuracy on standard benchmarks. We document a contrasting observation: when LLM features are introduced through pure input concatenation (rather than joint training, distillation, or prompt-conditioning), they can systematically degrade accuracy on the same homophilous benchmarks where end-to-end LLM pipelines succeed. With an MLP backbone on the Planetoid public split and bag-of-words original features, concatenating SBERT-encoded GPT-4o-mini TAPE features reduces PubMed test accuracy by -17.0 +/- 0.3 pp and Cora by -4.3 +/- 0.6 pp (CiteSeer -0.6 +/- 0.8 pp, within seed noise). The drop attenuates as we relax each condition (GCN / GCNII / GAT backbones, random splits, smaller encoders) and reverses on medium-homophily WikiCS (+4.4 pp) and ogbn-arxiv (+11.7 pp). To predict when concatenation helps versus hurts, we report a simple measure of LLM-alone discriminability, Delta_sig. Across 9 datasets Delta_sig correlates with the concatenation cost more strongly than homophily at point estimate (r^2 = 0.38 vs. 0.06; N=9, bootstrap CIs overlap). The bootstrap-best change-point is tau = 13.8 pp, and the rule "Delta_sig <= tau predicts non-positive concat cost" classifies 7/9 datasets correctly; since 60% of bootstrap samples place tau in [5, 30] pp, we treat Delta_sig as an interpretive lens rather than a precision filter. A dimension-controlled ablation on PubMed places the LLM-feature drop between same-source PCA (-2.3 pp) and same-dim Gaussian noise (-37.3 pp), ruling out dimensionality and weight-decay artifacts. Nine PubMed configurations fit a power law |Delta_concat| proportional to (sqrt(d_l/n))^1.31 with r^2 = 0.97; the low-Delta_sig, small-n corner is exactly where the headline -17 pp PubMed deficit appears.