Tool-Use Grounding
Benchmarks and methods for evaluating and improving whether LLM-generated tool calls invoke the correct API endpoints.
14 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · Shandong University of Technology, Shandong Agriculture and Engineering University
Executability-oriented benchmark exposes LLM service-call grounding failures
Introduces an executability-oriented benchmark of 3,135 real RapidAPI functions to expose whether generated code invokes the correct endpoint.
1 further paper
Oct 2026 · Hong Kong University of Science and Technology, Tencent
Silent tool errors roughly halve success of e-commerce LLM agents
Introduces the RITE benchmark for evaluating LLM agent recovery from silent tool errors in e-commerce workflows.
3 further papers
Oct 2026
AI agents struggle to adapt learned workflows as tools change
introduces a benchmark that tests agents' ability to adapt to evolving tool APIs, directly evaluating grounding robustness.
1 further paper
Oct 2026 · Zhejiang University, Zhejiang Geely Holding Group Co., Ltd
AI agents’ clicks marked wrong can work a third of the time
Redefines GUI click correctness by outcome rather than bounding box, providing a more accurate grounding metric for tool actions.
Oct 2026 · Oracle AI
AI agents that succeed at tasks may still fail their mid-way steps
evaluates tool-call sequence prefixes for executability.
1 further paper
Nov 2026
Terminal prompts make AI agents report success without doing the work
Shows terminal tool calls requiring confirmation are treated as completed without the operation finishing, a tool grounding failure.
1 further paper
Nov 2026 · Yale University, Ant Group
Robot action checks trade off error detection against pinpointing causes
Studies detection of semantically invalid robot commands (e.g., swapped axes) using a model that links command structure to observed outcome, extending grounding to action-level correctness.
7 of 14 papers shown