PulseAugur
EN
LIVE 15:43:54

LLM Features Can Harm GNN Performance on Homophilous Graphs

By PulseAugur Editorial · [2 sources] ·

A new research paper reveals that incorporating features generated by large language models (LLMs) into graph neural networks (GNNs) can sometimes decrease performance on specific benchmarks. This effect, termed 'concatenation interference,' was observed when LLM features were simply appended to existing data, leading to significant accuracy drops on datasets like PubMed and Cora. The study suggests that the effectiveness of LLM features depends on factors beyond simple concatenation, with performance improvements seen on datasets with medium homophily. AI

IMPACT This research suggests that simply appending LLM-generated features to GNNs may not always yield improvements and can even degrade performance on certain graph types.

RANK_REASON The cluster contains a research paper detailing novel findings about the interaction between LLM features and GNNs.

Read on arXiv cs.CL →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

LLM Features Can Harm GNN Performance on Homophilous Graphs

COVERAGE [2]

  1. arXiv cs.AI TIER_1 English(EN) · Zhongyuan Wang, Pratyusha Vemuri ·

    LLM Features Can Hurt GNNs: Concatenation Interference on Homophilous Graph Benchmarks

    arXiv:2606.17579v1 Announce Type: cross Abstract: 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 concatena…

  2. arXiv cs.CL TIER_1 English(EN) · Pratyusha Vemuri ·

    LLM Features Can Hurt GNNs: Concatenation Interference on Homophilous Graph Benchmarks

    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…

RELATED ENTITIES

RELATED TOPICS