Content
summary Summary

Apple's AI research team has introduced CAMPHOR, a new AI framework designed to process complex user queries locally on mobile devices while preserving user privacy.

Ad

CAMPHOR (Collaborative Agents for Multi-input Planning and High-Order Reasoning On Device) uses a hierarchical structure with specialized agents coordinated by a higher-level reasoning agent. This agent breaks down complex tasks into sub-steps and assigns them to expert agents.

Illustration: Smartphone screen with app icons and function calls, alongside a user figure with speech bubble for queries.
The CAMPHOR dataset simulates a realistic smartphone environment with various functions and personal information. This helps develop and test AI assistants that can interact with different apps and services. | Image: Fu et al.

Specialized agents cover areas such as personal context, device information, user perception, external knowledge, and task execution. Apple says this approach improves privacy and reduces latency compared to server-based solutions, which often require multiple exchanges between server and device.

Flowchart: CAMPHOR multi-agent system with Device, User, Personal Context, External Knowledge, and Task Completion agents.
The CAMPHOR system combines different agents to process and execute complex user queries. | Image: Fu et al.

Squeezing AI into your pocket

CAMPHOR uses prompt compression to make the most of limited mobile device resources. This technique reduces memory requirements by compressing function definitions into individual tokens. The study found that for certain agents, the number of static prompt tokens was reduced by up to 96 percent.

Ad
Ad
Diagram: SLM prompt compression, function and prompt tokens processed by SLM and integrated into the prompt.
Prompt compression makes it faster and easier to process function descriptions and conversation histories, which helps SLMs work better in different situations. | Image: Fu et al.

This substantial reduction in memory needs comes with only minor changes in accuracy. The Plan-F1 value, which measures task performance accuracy, only dropped from 39.89% to 38.45%. In tests, CAMPHOR's fine-tuned small language models outperformed large language models on personalized tasks, scoring up to 35% higher in task completion performance.

Two tables: Comparison of prompting strategies and model performances for various metrics such as Tool F1 and Plan F1.
According to Apple, SLMs fine-tuned for specific tasks can outperform larger cloud AI models at those tasks. | Image: Fu et al.

The team notes that the current CAMPHOR framework is limited to single interactions. Future work will focus on extending CAMPHOR for multiple interactions and integrating more complex runtime feedback and error handling.

AI could one day control your iPhone - and more

Apple's CAMPHOR represents a vision of AI agents that can interact with users in natural language while understanding and interacting with their environment, such as the smartphone interface. If successful, this principle could extend to many areas of work and life. OpenAI CEO Sam Altman is investing in hardware that integrates agentic, assistant-like AI into everyday life.

OpenAI recently released an open-source framework called "Swarm" on GitHub, a tool for creating, orchestrating, and deploying multi-agent systems. Swarm aims to make agent coordination and execution lightweight, controllable, and easy to test. Swarm uses routines and handoffs. Routines contain instructions and tools. Handoffs allow agents to route conversations. According to OpenAI, Swarm is experimental and not yet ready for production.

Ad
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • Apple's AI research team has developed an AI framework called CAMPHOR, which is designed to process complex user requests locally on mobile devices using different SMLs (Small Language Models) while maintaining privacy.
  • CAMPHOR uses a hierarchical structure of specialized agents coordinated by a higher-level reasoning agent. It breaks down complex tasks into sub-steps and assigns them to the specialized agents.
  • According to Apple, CAMPHOR's small language models, fine-tuned for personalized tasks, sometimes outperform large cloud AI models.
Sources
Online journalist Matthias is the co-founder and publisher of THE DECODER. He believes that artificial intelligence will fundamentally change the relationship between humans and computers.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.