Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure
Sapienza · hf · 2026-10-27
Researchers from Sapienza, Cambridge, Pisa and ISTI-CNR present an in-depth analysis of graph soft prompting (G-Retriever) and identify a pathological behavior they call token collapse.
- Graph soft prompting trains a GNN to produce graph-specific tokens meant to inject relational structure into LLM prompts
- Across 3 benchmarks, 3 graph encoders, 2 LLM families (frozen and LoRA-adapted), the learned graph tokens become nearly indistinguishable across inputs, acting as generic prompts rather than conveying relational information
- The authors argue evaluation should rely on alignment probes rather than accuracy alone
More from Research
- Qwen3.8-Flash-Next's open 180B weights hailed as a model organism for backdoor research — jd_pressman · 2026-10-27
- Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27
- arXiv advice from T. Dietterich: wrong primary category will almost surely get your paper on hold — tdietterich · 2026-10-27
- arXiv has 4,000+ papers on hold amid submission surge; moderator shares how to avoid it — tdietterich · 2026-10-27