Research Papers 论文研究 6h ago Updated 1h ago 更新于 1小时前 48

Making Every Tool Call Count: Necessary Tool-Evidence Path Rewards for Agentic Vision-Language Models 让每次工具调用都算数:面向智能体视觉语言模型的工具-证据路径必要奖励

NTEP (Necessary Tool-Evidence Path) is a novel annotation scheme that explicitly specifies essential external evidence and corresponding tool calls for each query in agentic VLMs NTEP-R introduces a supervision mechanism that rewards agents for aligning pre-call intent with necessary evidence-seeking goals and post-call observations with necessary evidence A non-repeated-goal regularizer penalizes redundant tool calls that revisit already satisfied NTEP goals The 8B-parameter NTEP-8B model signi 提出NTEP(必要工具-证据路径)标注方案,为每个查询显式指定所需外部证据及对应工具调用 设计NTEP-R奖励机制,从"调用前意图对齐"和"调用后信息提取对齐"两个维度监督工具使用过程 引入非重复目标正则化器,惩罚冗余调用已满足NTEP目标的行为,提升工具调用效率 8B参数模型NTEP-8B在7个图像 grounding 基准上显著改善搜索准确率和工具使用效率

65
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • NTEP (Necessary Tool-Evidence Path) is a novel annotation scheme that explicitly specifies essential external evidence and corresponding tool calls for each query in agentic VLMs
  • NTEP-R introduces a supervision mechanism that rewards agents for aligning pre-call intent with necessary evidence-seeking goals and post-call observations with necessary evidence
  • A non-repeated-goal regularizer penalizes redundant tool calls that revisit already satisfied NTEP goals
  • The 8B-parameter NTEP-8B model significantly improves both search-oriented accuracy and tool-use efficiency across seven image-grounded benchmarks
  • Fine-grained tool-evidence path supervision addresses two key shortcomings: redundant/off-target tool calls and failure to extract necessary information from observations

Why It Matters

This work addresses a critical gap in agentic VLM training where existing paradigms only evaluate tool-use based on final answer correctness, neglecting the evidence acquisition and utilization pipeline. For AI practitioners building vision-language agents, this demonstrates that intermediate-step supervision—rewarding not just outcomes but the quality of each tool invocation and observation processing—can substantially improve both accuracy and efficiency. The approach offers a practical framework for reducing wasteful tool calls while ensuring models extract actionable information from external evidence.

Technical Details

  • NTEP Annotation Scheme: Explicitly annotates each query with the essential external evidence required and the specific tool calls needed to obtain it, creating a ground-truth path for tool-evidence alignment
  • NTEP-R Reward Mechanism: Provides dual rewards—(1) pre-call reward for aligning the agent's intent with a necessary evidence-seeking goal, and (2) post-call reward for ensuring the summarized information from tool observations aligns with the annotated necessary evidence
  • Non-Repeated-Goal Regularizer: Penalizes redundant tool calls that revisit NTEP goals already satisfied, encouraging efficient evidence gathering without repetition
  • Unified Three-Tool Framework: Integrates image cropping, image search, and text search as the tool set for the NTEP-8B model
  • Evaluation: Tested on seven image-grounded benchmarks, with the 8B-parameter NTEP-8B instantiation showing significant improvements in both search-oriented accuracy and tool-use efficiency

Industry Insight

  • The emphasis on intermediate-step supervision for tool use signals a shift from outcome-only training toward process-aware reinforcement, which could become a standard practice for agentic VLM development
  • Redundant tool calls represent both a computational cost and a latency bottleneck; methods like the non-repeated-goal regularizer offer actionable strategies for improving real-world agent efficiency
  • The NTEP annotation scheme could serve as a template for other domains requiring multi-step evidence gathering, suggesting broader applicability beyond vision-language tasks

TL;DR

  • 提出NTEP(必要工具-证据路径)标注方案,为每个查询显式指定所需外部证据及对应工具调用
  • 设计NTEP-R奖励机制,从"调用前意图对齐"和"调用后信息提取对齐"两个维度监督工具使用过程
  • 引入非重复目标正则化器,惩罚冗余调用已满足NTEP目标的行为,提升工具调用效率
  • 8B参数模型NTEP-8B在7个图像 grounding 基准上显著改善搜索准确率和工具使用效率

为什么值得看

本文针对Agentic VLM工具调用训练中的核心痛点——证据获取与利用监督不足——提出了细粒度的过程级奖励机制,为训练高效、精准的工具使用能力提供了新思路。

技术解析

  • NTEP标注方案:为每个复杂查询显式标注所需的外部证据(如图像细节、文本信息)及对应的工具调用路径,建立"查询→证据→工具调用"的映射关系。
  • NTEP-R双阶段奖励:在工具调用前,奖励模型意图与必要证据目标对齐;在工具调用后,奖励模型从观察结果中提取的信息与必要证据对齐,形成闭环监督。
  • 非重复目标正则化:设计惩罚项,对重复调用已满足NTEP目标的行为施加惩罚,有效减少冗余工具调用。
  • 统一三工具框架:整合图像裁剪、图像搜索、文本搜索三种工具,在7个图像 grounding 基准测试上验证NTEP-8B的性能提升。

行业启示

  • Agentic VLM的训练范式需从"结果监督"转向"过程监督",工具调用的中间环节(意图、证据提取)应纳入奖励设计。
  • 工具调用效率是Agentic系统落地的关键瓶颈,通过正则化手段抑制冗余调用可显著降低推理成本。
  • 细粒度证据路径标注可作为高质量训练数据构建的新范式,值得在更多多模态Agent任务中推广。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 Agent Agent Multimodal 多模态 LLM 大模型 Fine-tuning 微调