{
  "id": 4661,
  "url": "https://arxiv.org/abs/2605.09011v1",
  "title": "A Geometric Perspective on Next-Token Prediction in Large Language Models: Three Emerging Phases",
  "summary": "We investigate the geometry of predictive information across the layers of large language models (LLMs). We repurpose representation lenses-learned affine maps trained to predict the next token from intermediate residual streams-as geometric diagnostic tools. Rather than asking what the model predicts at each layer, we ask where predictive information resides and how it evolves across depth. We define at each layer a predictive readout subspace as the dominant k-dimensional singular subspace of ",
  "authors": "Gianfranco Lombardo, Giuseppe Trimigno, Stefano Cagnoni",
  "category": "research",
  "topics": "healthcare,finance-investment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-09T15:51:59.000Z",
  "fetched_at": "2026-07-14T16:31:08.358Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4661",
  "original_url": "https://arxiv.org/abs/2605.09011v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}