{
  "id": "2604.12426",
  "title": "Do Transformers Use their Depth Adaptively? Evidence from a Relational Reasoning Task",
  "first_public": "2026-04-14",
  "paper_url": "https://arxiv.org/abs/2604.12426",
  "authors": "Alicia Curth, Rachel Lawrence, Sushrut Karmalkar, Niranjani Prasad",
  "arxiv": "2604.12426",
  "added_at": "2026-11-09",
  "venue": "arXiv",
  "provenance": "Primary-source literature review, 2026-11-09",
  "lens": [
    "analysis"
  ],
  "representation": [
    "hidden-state"
  ],
  "computation": [
    "depth-wise"
  ],
  "training": [
    "analysis-only",
    "supervised"
  ],
  "questions": [
    "interpretability",
    "evaluation",
    "capability"
  ],
  "evidence": [
    "empirical",
    "probing",
    "mechanistic"
  ],
  "modality": [
    "language"
  ],
  "summary": "Studies when transformer layers combine family relationships to answer multi-hop questions. Uses logit-lens probes and causal activation patching to compare depth use before and after task fine-tuning.",
  "author_names": [
    "Alicia Curth",
    "Rachel Lawrence",
    "Sushrut Karmalkar",
    "Niranjani Prasad"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2604.12426/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2604-12426",
  "citation_date": "2026-04-14",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2604.12426/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2604.12426/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2604.12426/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2604.12426/citation.csl.json",
  "pdf_url": "https://arxiv.org/pdf/2604.12426"
}