Intel / article

The Menu Is an Execution Prior: State-Path Tool Menus for Online Agents

Source: arxiv.org
Author: Bo Yan, Weikai Lin, Song Wang
Date: Published Sep 10, 2026

Context

An can fail because its available tools omit a necessary earlier step. The State-Path Tool Menu paper addresses this by building the tool menu, the short ordered list of tools an agent is allowed to call, as a connected route toward the requested outcome. It includes tools that produce missing inputs and places them before tools that consume those inputs.

The authors report raising online task success on ToolBench from 73.7% to 89.8% without changing the executing agent, and say the menu outperformed retrieval, , generation and routing baselines. That points to which tools are shown, and in what order, as a factor separate from the agent itself. The reported result concerns this and menu-construction method.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
reranking — A second pass that re-scores retrieved candidates by reading each one against the query, fixing the ordering that fast vector search got approximately right.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Key quotes

“We introduce the state path, a pre-execution route from the observable request state to the desired outcome, and propose State-Path Tool Menu to learn it.”

“On ToolBench, our menu raises online success from 0.737 to 0.898 and outperforms retrieval, reranking, generation, and routing baselines without changing the agent.”

“The State-Path menu also covers more complete chains with 32 tools than the official list covers with 128, and its success gain persists across executor families with different model capacities.”

Read the source arxiv.org