How an Everyday Word in a Tool Schema Changes AI Agents' Searches While Every Call Stays Valid

An AI agent fills the fields of a tool with values from a fixed list, the tool's schema. Some allowed values are also everyday words. A traveller who asks for "a hotel" can mean any place to stay, but a booking tool can separate hotel from guesthouse. This preprint asks whether such a value name changes what an agent searches for while every call stays valid. It uses public MultiWOZ booking dialogues and nine language models. In 5.6% to 18.8% of valid first answers, depending on the model, the search would have excluded every place that fits the traveller's task. In tests registered before we opened the answers, a rename of the field and the value removed hotel filters that the task did not need, and also filters that it needed. The preprint also gives a release test for schema updates. It shows that the verdict on a repair depends on the reference record. The evidence concerns one word in one corpus.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24270644
Primary Topic
Speech and dialogue systems
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

How an Everyday Word in a Tool Schema Changes AI Agents' Searches While Every Call Stays Valid

Dev D. Goyal
2 links
Zenodo (CERN European Organization for Nuclear Research)
Speech and dialogue systems
preprint

How an Everyday Word in a Tool Schema Changes AI Agents' Searches While Every Call Stays Valid

Dev D. Goyal
preprint en

Abstract

An AI agent fills the fields of a tool with values from a fixed list, the tool's schema. Some allowed values are also everyday words. A traveller who asks for "a hotel" can mean any place to stay, but a booking tool can separate hotel from guesthouse. This preprint asks whether such a value name changes what an agent searches for while every call stays valid. It uses public MultiWOZ booking dialogues and nine language models. In 5.6% to 18.8% of valid first answers, depending on the model, the search would have excluded every place that fits the traveller's task. In tests registered before we opened the answers, a rename of the field and the value removed hotel filters that the task did not need, and also filters that it needed. The preprint also gives a release test for schema updates. It shows that the verdict on a repair depends on the reference record. The evidence concerns one word in one corpus.

Zenodo (CERN European Organization for Nuclear Research)
Speech and dialogue systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon