/p/2026-10-15 · explainer
Paper explainer · 2610.14334 · Li, Qiao and Gao

Two tools, one name, and the schema decides.

A model choosing a tool has never run it. All it sees is metadata — a server name, a tool name, a description, and the shape of the parameters. Across 367,174 published tool definitions, 169 records named send_email expose 78 different parameter sets; swap one piece of metadata at a time across five models and five tasks, with candidate order rotated, and the parameter schema is the strongest effect in 19 of 30 conditions, about twice the size of the server name. Models want separate top-level fields, and they read your field names closely enough that renaming one halves the odds.

01 · The problem

The tools you are being compared against

which part of the registry to look at

02 · The mechanism

One schema change, in odds you can read

which model is choosing

03 · The evidence

Four things you can change, ranked

pick a factor to read its tally

04 · Why I care

It reads your field names, and your request

rename one field and nothing else

05 · Apply it

Audit one of your own tools illustrative

Results

What the paper actually measured

What it does not show

In practice