Research Papers 论文研究 19h ago Updated 1h ago 更新于 1小时前 48

From Fixed Keys to Readable Schemas: Small Language Models for Vehicle Agent Function Calls 从固定键到可读模式:小型语言模型在车载智能体函数调用中的应用

The paper compares two function-surface representation approaches for SLM-based vehicle agents: Functional Tokens (FT) and Schema-in-Prompt (SIP) A new benchmark of 9,822 single-turn examples across 79 Android Automotive functions was introduced, including held-out functions and refusal cases On seen functions, model scaling provides limited benefit (270M matches 1.7B; peak performance at 0.6B), but SIP generalizes to unseen functions while FT achieves zero accuracy by construction FT risks invo 车载助手在内存和延迟约束下使用小语言模型(SLM)进行功能调用,核心设计选择是Functional Token(FT)与Schema-in-Prompt(SIP)两种方案 创建包含9,822个单轮示例的基准测试,涵盖79个Android Automotive车辆功能,包括未见功能和要求拒绝的请求 训练内功能上270M模型可匹敌1.7B模型,最优性能出现在0.6B规模;未见功能上FT准确率为零,SIP可泛化且随规模显著提升 超出范围请求时FT可能调用未训练功能,SIP更可靠地基于可用函数拒绝;SIP灵活性以更高内存和延迟为代价 功能表面表示方式而非模型规模 alone 决定了SLM车辆功能调用的

65
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • The paper compares two function-surface representation approaches for SLM-based vehicle agents: Functional Tokens (FT) and Schema-in-Prompt (SIP)
  • A new benchmark of 9,822 single-turn examples across 79 Android Automotive functions was introduced, including held-out functions and refusal cases
  • On seen functions, model scaling provides limited benefit (270M matches 1.7B; peak performance at 0.6B), but SIP generalizes to unseen functions while FT achieves zero accuracy by construction
  • FT risks invoking unavailable functions on out-of-scope requests, whereas SIP more reliably refuses based on offered functions, at the cost of higher memory and latency
  • The key finding is that function-surface representation, not model scale alone, determines the capabilities and failure modes of SLM-based vehicle function calling

Why It Matters

This research directly addresses a critical deployment challenge for in-vehicle AI assistants: how to effectively represent function catalogs for resource-constrained small language models running on-device. For AI practitioners building agent systems, the findings provide actionable guidance on the trade-offs between compact token-based representations and flexible schema-based prompting, with implications for reliability, generalization, and system design in safety-critical domains like automotive.

Technical Details

  • Functional Tokens (FT): Each vehicle function is represented by a dedicated token, enabling compact inference but restricting the model to functions seen during training only
  • Schema-in-Prompt (SIP): Function schemas are provided directly in the prompt, enabling generalization to unseen functions at the cost of longer prompts and higher inference overhead
  • Benchmark: 9,822 single-turn examples spanning 79 vehicle functions derived from Android Automotive, with held-out functions and out-of-scope refusal requests
  • Models evaluated: Four SLMs ranging from 270M to 1.7B parameters, all under matched fine-tuning conditions
  • Theoretical analysis: Explains how SIP enables generalization to unseen functions and why longer schema contexts increase inference cost

Industry Insight

  • For on-device automotive AI, a ~0.6B parameter model with SIP may offer the best accuracy-latency trade-off for known functions, while SIP is essential when the function catalog needs to evolve without retraining
  • System designers should prioritize function-surface representation strategy over simply scaling model size when deploying SLMs for tool-use agents in constrained environments
  • The zero generalization of FT to unseen functions highlights a hard limitation of token-based function encoding, suggesting that any production vehicle assistant requiring dynamic function updates should adopt SIP or a hybrid approach despite the latency cost

TL;DR

  • 车载助手在内存和延迟约束下使用小语言模型(SLM)进行功能调用,核心设计选择是Functional Token(FT)与Schema-in-Prompt(SIP)两种方案
  • 创建包含9,822个单轮示例的基准测试,涵盖79个Android Automotive车辆功能,包括未见功能和要求拒绝的请求
  • 训练内功能上270M模型可匹敌1.7B模型,最优性能出现在0.6B规模;未见功能上FT准确率为零,SIP可泛化且随规模显著提升
  • 超出范围请求时FT可能调用未训练功能,SIP更可靠地基于可用函数拒绝;SIP灵活性以更高内存和延迟为代价
  • 功能表面表示方式而非模型规模 alone 决定了SLM车辆功能调用的能力和失败模式

为什么值得看

本文为车载AI助手提供了关键的设计决策依据:在资源受限的端侧部署场景下,如何选择功能表示方式直接影响模型的泛化能力和可靠性。研究结果对智能座舱、车载语音助手等垂直领域的AI系统架构设计具有重要参考价值。

技术解析

  • 两种功能表示方案:Functional Token(FT)为每个功能分配专用token,推理紧凑但仅能调用训练期间学到的功能;Schema-in-Prompt(SIP)将功能schema直接放入prompt,可泛化到未见功能但需要更长prompt和更高推理开销
  • 基准测试设计:9,822个单轮示例覆盖79个Android Automotive车辆功能,包含held-out功能和需要拒绝的out-of-scope请求,确保全面评估泛化和拒绝能力
  • 模型规模实验:在270M至1.7B参数的四种SLM上进行匹配微调,训练内功能上270M可匹敌1.7B,最优在0.6B;未见功能上SIP随规模显著提升而FT始终为零
  • 理论分析:论文提供理论解释说明SIP如何启用泛化能力,以及更长schema上下文为何增加推理成本

行业启示

  • 车载AI系统设计中,功能表示策略的选择比单纯扩大模型规模更重要;对于需要频繁更新功能的场景,SIP方案更具长期价值
  • 端侧部署需权衡泛化能力与推理效率:FT适合功能固定的场景,SIP适合需要支持新功能的场景
  • 车辆助手的安全性和可靠性依赖于正确的拒绝机制设计,SIP在out-of-scope请求处理上表现更优,可降低误触发风险

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。

Agent Agent Autonomous Driving 自动驾驶 Inference 推理 Deployment 部署 Research 科学研究