CAMPHOR - Apple's on-device small language model multi-agent framework
CAMPHOR is a small language model (SLM) multi-agent framework developed by Apple, designed to improve privacy and responsiveness on mobile devices. The framework is based on processing multiple user inputs locally on the device and performing personalized contextual inference...
What is CAMPHOR?
CAMPHOR is a multi-agent framework for Small Language Models (SLM) on mobile devices, developed by Apple, that improves privacy and responsiveness. The framework is based on processing multiple user inputs locally on the device and performing personalized contextual reasoning, ensuring user privacy and security. CAMPHOR uses a hierarchical architecture, where higher-order inference agents decompose complex tasks and coordinate expert agents to perform personalized contextual retrieval, tool interaction, and dynamic plan generation. Through parameter sharing and prompt compression techniques among agents, CAMPHOR significantly reduces model size, latency, and memory footprint.
CAMPHOR's main functions
- Multi-user input processingIt can process input from multiple users simultaneously.
- Local context reasoningPerform personal context reasoning locally on the device to protect privacy.
- Complex task decompositionBreaking down complex tasks into smaller subtasks makes them easier to manage and execute.
- Tool InteractionInteract with tools and applications on the device to perform specific tasks.
- Dynamic Programming: Dynamically generate execution plans based on user needs and context.
- Parameter sharing: Sharing parameters among different agents reduces model size and improves efficiency.
CAMPHOR's technical principles
- Layered architectureIt uses a hierarchical agent architecture, including a high-level reasoning agent and multiple expert agents.
- Advanced ReasoningThe higher-order reasoning agent is responsible for planning and coordinating the entire task execution process.
- Expert Intelligent AgentExpert agents are responsible for specific tasks, such as personal context retrieval and tool interaction.
- Parameter sharingIt enables parameter sharing among agents, reducing the storage and computational requirements of the model.
- Prompt to compressBased on compressing the function definition into a single token, reducing the length of the prompt, it can work with limited resources.
- Execute locallyAll processing is done on the user's device, without needing to communicate with the server, thus 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 assistantOn smartphones, it acts as a personalized mobile assistant, handling users' daily tasks such as schedule management, reminder settings, and information retrieval.
- Privacy-protected data processingSuitable for scenarios that require protection of user privacy, such as health data management and financial information processing.
- Multi-task parallel processingSuitable for environments that require parallel processing of multiple requests, such as home automation systems.
- Localization servicesIn the absence of a stable network connection, it provides services based on local data, such as navigation and local information retrieval.
- Smart device interactionIt interacts with various smart devices, such as smart home devices, to enable collaborative work between devices.