What is CAMPHOR
CAMPHOR is an on-device small language model (SLM) multi-agent framework introduced by the Apple team, capable of enhancing privacy protection and response speed on mobile devices. The framework is based on processing multiple user inputs locally on the device and performing personal context reasoning, ensuring user privacy and security. CAMPHOR is based on a hierarchical architecture, in which a high-level reasoning agent is responsible for decomposing complex tasks and coordinating expert agents to perform personal context retrieval, tool interaction, and dynamic plan generation. Based on parameter sharing and prompt compression techniques among agents, CAMPHOR significantly reduces model size, latency, and memory usage.
Main Features of CAMPHOR
- Multi-User Input Processing: Capable of processing inputs from multiple users simultaneously.
- Local Context Reasoning: Performs reasoning on personal context locally on the device, protecting privacy.
- Complex Task Decomposition: Decomposes complex tasks into smaller subtasks for easier management and execution.
- Tool Interaction: Interacts with tools and applications on the device to perform specific tasks.
- Dynamic Plan Generation: Dynamically generates execution plans based on user needs and context.
- Parameter Sharing: Shares parameters among different agents, reducing model size and improving efficiency.
Technical Principles of CAMPHOR
- Hierarchical Architecture: Uses a hierarchical agent architecture, including a high-level reasoning agent and multiple expert agents.
- High-Level Reasoning: The high-level reasoning agent is responsible for planning and coordinating the execution flow of the entire task.
- Expert Agents: Expert agents are responsible for specific tasks, such as personal context retrieval, tool interaction, etc.
- Parameter Sharing: Enables parameter sharing among agents, reducing the model's storage and computational requirements.
- Prompt Compression: Based on compressing function definitions into a single token, reducing prompt length and enabling operation under limited resources.
- Local Execution: All processing is completed on the user's device without communicating with servers, protecting privacy and reducing latency.
CAMPHOR Project Address
- Project Official Website：machinelearning.apple.com/research/collaborative-agents
- arXiv Technical Paper：https://arxiv.org/pdf/2410.09407
Application Scenarios of CAMPHOR
- Personalized Mobile Assistant: On smartphones, serves as a personalized mobile assistant to handle users' daily tasks, such as schedule management, reminder setting, information retrieval, etc.
- Privacy-Preserving Data Processing: Suitable for scenarios that require protecting user privacy, such as health data management, financial information processing, etc.
- Multi-Task Parallel Processing: Suitable for environments that need to process multiple requests in parallel, such as home automation systems.
- Localized Services: Provides services based on local data when there is no stable network connection, such as navigation, local information retrieval, etc.
- Smart Device Interaction: Interacts with various smart devices, such as smart home devices, to enable collaborative work among devices.