ToolGrad: Efficient Tool-use Dataset Generation with Textual “Gradients”

Zhongyi Zhou1,2,3, Kohei Uehara2, Haoyu Zhang2, Jingtao Zhou1, Lin Gu3,4, Ruofei Du1, Zheng Xu1, Tatsuya Harada2,3
1Google 2The University of Tokyo 3RIKEN AIP 4Tohoku University
ACL 2026 Findings
arXiv Code Dataset Models (TBD)
ToolGrad teaser image

ToolGrad inverts the tool-use data generation pipeline. Our answer-first tool-use data generation framework leads to almost 100% pass rate and lower cost.

Generated Data Samples

Our AI understands requests, plans, and uses tools to find answers.

TLDR:

"AI research and Tesla's AI involvement"

Query:

"I'm researching the latest developments in artificial intelligence. Can you give me some recent news headlines and tell me about Tesla's involvement in AI?"

GT Tool Call Graph:

Query ➔ news_search ➔ get_company_news

Tool Usage:

Tool Name Tool Input Tool Response
News_Media-google_news-search { "keyword": "artificial intelligence", "lr": "en-US" } { "status": "success", "items": [{ "title": "China to rely on artificial intelligence in education reform...", "publisher": "Reuters" }, { "title": "Inside Palantir's AI Strategy...", "publisher": "Investor's Business Daily" }] }
News_Media-a_i_smartable-getcompanies { "page": 1 } { "title": "Companies", "value": [{ "title": "Tesla", "excerpt": "Tesla's mission is to accelerate the world's transition to sustainable energy..." }] }

TLDR:

"Instagram followers, domains and tiktok videos"

Query:

"I'm researching potential online threats related to Instagram. Can you tell me how many followers the official Instagram account has, what are some recently registered domains that might be impersonating Instagram, and are there any TikTok videos discussing Instagram scams?"

GT Tool Call Graph:

Query ➔
instagram_metadata
get_daily_domains ➔ tiktok_search

Tool Usage:

Tool Name Tool Input Tool Response
Social-instagram_data-user_information_metadata { "username": "instagram" } { "full_name": "Instagram", "edge_followed_by": { "count": 678211503 }, "is_verified": true }
Monitoring-opensquat-get_daily_domains { "keyword": "Instagram" } { "response": "success", "count": 13, "domains": ["gbinstagram.com", "instagram-bussinescenter.click", "miho-instagram.net"] }
Data-tiktok_api-search_videos { "query": "Instagram scam" } { "query": "Instagram scam", "videos": [{ "video_id": "7587094713447058701", "author_name": "Dave jefferson", "description": "#romancescams #instagramscammers" }] }

TLDR:

"IP address geolocation and lookup"

Query:

"I'm interested in the IP address 8.8.8.8. Can you tell me where it's located geographically and what its timezone is? Also, I'd like to know if it's associated with any specific website or service."

GT Tool Call Graph:

Query ➔
get_geo_location ➔ timezone_lookup
url_lookup

Tool Usage:

Tool Name Tool Input Tool Response
Database-mocking_rock-get_geo_location_of_ip_address { "ip": "8.8.8.8" } { "continentName": "North America", "countryName": "United States", "latLong": { "latitude": "37.751", "longitude": "-97.822" } }
Location-timezone_by_location-timezone { "lon": -97.822, "lat": 37.751, "c": 1, "s": 1 } { "Result": "America/Chicago" }
Location-url_lookup_by_api_ninjas-v1_urllookup { "url": "http://8.8.8.8" } { "is_valid": true, "country": "United States", "isp": "Amazon.com, Inc.", "timezone": "America/Los_Angeles" }

ToolGrad

Analogy to ML and TextGrad

Analogy to ML and TextGrad

Agentic Framework Design

Agentic Framework Design

BibTeX

@article{zhou2025toolgrad,
    title={ToolGrad: Efficient Tool-use Dataset Generation with Textual “Gradients”},
    author={Zhongyi Zhou and Kohei Uehara and Haoyu Zhang and Jingtao Zhou and Lin Gu and Ruofei Du and Zheng Xu and Tatsuya Harada},
    year={2025},
    archivePrefix={arXiv},
}