{
  "id": "2607.03502",
  "title": "Reading Between the Dots: Decoding Hidden Computation across Filler Tokens",
  "authors": "Kaley Brauer, Claudio Mayrink Verdun, Samuel Marks",
  "first_public": "2026-07-03",
  "added_at": "2026-10-23",
  "venue": "NeurIPS 2026",
  "arxiv": "2607.03502",
  "paper_url": "https://arxiv.org/abs/2607.03502",
  "html_url": "https://arxiv.org/html/2607.03502v2",
  "code_url": "https://github.com/kaleybrauer/filler-token-reasoning",
  "lens": [
    "learned-tokens",
    "analysis"
  ],
  "representation": [
    "discrete-token",
    "hidden-state",
    "cache-memory"
  ],
  "computation": [
    "token-wise",
    "depth-wise"
  ],
  "training": [
    "analysis-only",
    "inference-only"
  ],
  "questions": [
    "interpretability",
    "evaluation",
    "limitations"
  ],
  "evidence": [
    "empirical",
    "mechanistic",
    "probing"
  ],
  "modality": [
    "language",
    "symbolic"
  ],
  "summary": "Studies how models use filler tokens to compute without writing intermediate steps. Reads intermediate values from hidden states and swaps cached states between examples to test their effect on answers, then evaluates an unsupervised decoder on fact retrieval, arithmetic and string tasks.",
  "provenance": "Primary-source literature review, 2026-10-23",
  "author_names": [
    "Kaley Brauer",
    "Claudio Mayrink Verdun",
    "Samuel Marks"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2607.03502/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2607-03502",
  "citation_date": "2026-07-03",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2607.03502/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2607.03502/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2607.03502/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2607.03502/citation.csl.json",
  "pdf_url": "https://arxiv.org/pdf/2607.03502"
}