---
id: 20260906-T0-06
title: "提升智能体视觉语言模型工具调用准确率，新奖励机制见效"
title_en: "Tool-Evidence Path Rewards Boost Agentic VLM Accuracy"
url: https://ai.daily.yangsir.net/daily/20260906-T0-06
issue_date: 2026-09-06
publish_date: 2026-09-05T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2609.03493
---

# 提升智能体视觉语言模型工具调用准确率，新奖励机制见效

一篇新论文提出了一种名为“必要工具证据路径奖励”的机制，旨在提升智能体视觉语言模型（VLM）在回答需要细粒度视觉细节或外部知识的复杂问题时的表现。该方法通过奖励模型在调用工具时遵循必要的证据路径，减少无效调用。实验表明，该机制能有效引导模型获取关键缺失信息，从而提升复杂图像问答的准确性。对于开发者而言，该研究提供了一种优化智能体推理与工具使用策略的新思路，有助于构建更可靠的视觉问答系统。

## English Version

**Tool-Evidence Path Rewards Boost Agentic VLM Accuracy**

A new paper introduces Necessary Tool-Evidence Path Rewards, a mechanism to improve agentic vision-language models (VLMs) on complex image-grounded questions. By rewarding models for following evidence paths during tool calls, the method reduces ineffective invocations and enhances accuracy. Findings indicate this approach effectively guides models to retrieve critical missing information, offering developers a new strategy to build more reliable visual question-answering systems.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2609.03493)

**详情页**：https://ai.daily.yangsir.net/daily/20260906-T0-06

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*