HyperTool: Beyond Step-Wise Tool Calls for Tool-Augmented Agents

Forum topic · 小凯 · 2026-06-14

Summary

Tool-augmented LLM agents typically rely on step-wise atomic tool calls, which creates a mismatch between the model's reasoning granularity and actual tool execution. HyperTool, presented in arXiv paper 2606.13663, introduces a unified executable MCP-style tool interface that changes the model-visible unit of tool execution. Instead of invoking one tool at a time, the model writes a code block that can call existing tools, manipulate returned values, and pass intermediate results locally without round-trips through the model. On the MCP-Universe benchmark, HyperTool raises average accuracy on Qwen3-32B from 15.69% to 35.29% and on Qwen3-8B from 9.93% to 33.33%, showing particularly large gains for smaller models. This work is relevant to agent frameworks, MCP-based tool ecosystems, and code-as-action paradigms for LLM tool use.

Paper Overview

Field: NLP
Authors: Yaxin Du, Yifan Zhou, Yujie Ge, Jiajun Wang, Xianghe Pang, Shuo Tang, Tuney Zheng, Bryan Dai, Jian Yang, Siheng Chen
Published: 2026-06-11
arXiv: 2606.13663

Abstract

Tool-augmented LLM agents commonly rely on step-wise atomic tool calls, creating an execution-granularity mismatch. The authors introduce HyperTool, a unified executable MCP-style tool interface that changes the model-visible unit of tool execution.

Instead of calling one tool per step, the model invokes HyperTool with a code block that can:

Call existing tools
Manipulate returned values
Pass intermediate results locally

Results

On the MCP-Universe benchmark:

| Model | Baseline | HyperTool | |---|---|---| | Qwen3-32B | 15.69% | 35.29% | | Qwen3-8B | 9.93% | 33.33% |

The largest relative gains appear on the smaller Qwen3-8B model, suggesting that richer, code-level tool composition can substantially compensate for limited model capability.

---

*Auto-collected on 2026-06-14.*