← News

ResearchSources: arXiv · 20 Nov 2026

TETHER uses a document's table of contents to structure long files without training

A new arXiv paper proposes a training-free framework that treats the table of contents as a structural prior, claiming state-of-the-art zero-shot hierarchy recovery on long, design-driven documents.

In plain words

Turning messy documents into organized, machine-readable structures is a common bottleneck for large language model pipelines. TETHER is a new method that needs no training: it first builds a document outline from the table of contents, then attaches content blocks to that outline. The authors report it beats existing parsers and models on their multilingual benchmark.

What was announced

The paper argues that existing bottom-up approaches depend on localized visual cues, which become unreliable in high-entropy layouts where visual prominence does not match logical depth, while direct LLM extraction lacks spatial grounding and can hallucinate in long contexts. TETHER instead runs a two-stage neuro-symbolic mechanism: a Large Multimodal Model builds a document skeleton from the raw table of contents under spatio-semantic constraints, and an anchor-based alignment strategy then tethers content blocks to that skeleton to resolve boundary ambiguities.

The authors evaluated TETHER on a multilingual benchmark of ESG reports, product manuals, and annual reports, reporting state-of-the-art zero-shot hierarchy recovery that outperforms deep hierarchical recovery models, dedicated parsers, and general-purpose LMMs. Stress tests described in the paper indicate resilience to sparse omissions and textual noise, and ablations support the framework's design. The authors also report a 90% token reduction and drastic speedups over end-to-end baselines. These results are self-reported by the paper's authors and have not been independently verified.

What it does not show

  • Evidence is limited to the paper's abstract; the full methodology, benchmark details, and results were not available for review.
  • All performance claims, including the 90% token reduction and speedups, come from the authors and are not independently confirmed.

For security and IT teams

Scope
The framework targets design-driven long documents such as ESG reports, product manuals, and annual reports, per the authors' benchmark; applicability to other document types is not established in the abstract.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.