Newest papers

On the Token Collapse Phenomenon in Textual Graph Soft Prompting

Andrea Giuseppe Di Francesco, Rishabh Jain, Franco Maria Nardini, Nicola Tonellotto, Pietro Liò, Fabrizio Silvestri · 2026 · arXiv

DOI: 10.48550/arXiv.2610.31572 · open-access PDF

German

Die jüngsten Methoden zur Injektion von Graph-Strukturen in LLM-Prompts umfassen Graph-Soft-Prompting, bei dem ein Graph-Neural-Netzwerk (GNN) grafspezifische Token lernt, die relationelle Informationen zwischen Entitäten in Text-Prompts erfassen. G-Retriever kombiniert Graph-Soft-Prompts mit einer verbalisierten Beschreibung desselben Graphs, um das Text-Prompt-Signal zu stärken. Wir führen eine eingehende Analyse von G-Retriever durch und enthüllen ein pathologisches Verhalten, das wir Token-Kollaps nennen. Über drei Benchmarks, drei Graph-Codierer, zwei LLM-Familien und beide gefrorenen und LoRA-

Machine translation (Meta M2M100); the English original follows.

Abstract

Recent methods for injecting graph structure into LLM prompts include graph soft prompting, in which a graph neural network (GNN) learns graph-specific tokens that capture relational information among entities in textual prompts. G-Retriever combines graph soft prompts with a verbalized description of the same graph to strengthen the textual prompt signal. We conduct an in-depth analysis of G-Retriever and uncover a pathological behavior we term token collapse. Across three benchmarks, three graph encoders, two LLM families, and both frozen and LoRA-adapted models, G-Retriever's graph tokens b

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Is this paper real? This page was built from OpenAlex and arXiv records, not generated. Software can verify any reference in one call: GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.31572, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txt

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange