← News

ResearchSources: arXiv · 20 Nov 2026

TableRoute routes table question answering across strategies by model capability

A new arXiv paper proposes a framework that checks what evidence a question needs, then picks the inference strategy best suited to the table size and the model's reliability.

In plain words

Asking questions about tables is hard for large language models because tables vary from short to extremely long, and a method that works for one size often fails for another. The paper's framework, TableRoute, first verifies which columns and values a question depends on, then chooses among three inference strategies based on table length and how reliably the model handles the needed operations. In the reported benchmarks this improved accuracy while lowering token cost.

What was announced

The authors argue that existing table question answering systems tend to commit to a single inference strategy, which becomes brittle as tables range from compact to million-token sizes. Full-table reading works for small inputs but suffers from irrelevant context; table reduction helps with long inputs but relies on accurate evidence localization; and code-style execution is needed when a table exceeds the model's context window, yet remains sensitive to errors in grounding schema and values, according to the paper.

TableRoute addresses this by first generating validated question-table evidence and identifying required operations, then routing each instance among the three strategies using serialized table length and calibrated estimates of how reliably the base model handles those operations. The validated evidence grounds columns, values, and conditions for the chosen route. On the WikiTQ, WTQ-L, BirdQA, and TableBench benchmarks, the authors report consistent accuracy gains across table scales alongside reduced online token cost. Code is stated to be released upon publication.

What it does not show

  • The evidence is the paper's abstract only; benchmark results are author-reported and not independently verified.
  • The abstract states code will be made available upon publication, so availability is not yet confirmed.

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.