What is DynaSaur
DynaSaur is a large language model (LLM) agent framework launched by Adobe Research, breaking through the limitations of traditional LLM agent systems that are constrained by predefined action sets. The framework supports agents in dynamically creating and composing actions, interacting with the environment based on generating and executing Python code, enabling more flexible problem-solving. DynaSaur can accumulate generated actions and build a reusable function library, improving the efficiency and adaptability of future tasks. In the GAIA benchmark test, DynaSaur demonstrates significant flexibility, especially performing well when handling complex and long-term tasks.
DynaSaur's Main Features
- Dynamic Action Creation: Based on the environment and task requirements, dynamically generate new Python functions as actions, without relying on predefined action sets.
- Action Accumulation and Reuse: The framework accumulates generated actions, builds a reusable function library, uses them in future tasks, and improves the efficiency of problem-solving.
- Environment Interaction: Interact with the environment based on generating and executing Python code; the code defines new actions or calls existing actions.
- Flexibility and Adaptability: When facing complex and long-term tasks, it can adapt flexibly, especially recovering and completing tasks when predefined actions are not applicable or fail.
DynaSaur's Technical Principles
- Action Representation: Represent each action as a Python function, using Python's generality and LLM's code generation capabilities.
- Action Retrieval: Based on the action retrieval function, retrieve the most relevant previously generated actions from the accumulated action library according to the query and context.
- Action Accumulation: During task execution, add newly generated actions to the action library, and the actions are used for future tasks.
- Partially Observable Markov Decision Process (POMDP): Model the agent's behavior as a POMDP, which includes the task space, action space, state space, observation space, state transition function, and observation function.
- Dynamic Action Space: Introduce a potentially infinite action set A*, supporting the agent in proposing any action at each time step to solve the task.
DynaSaur's Project Address
- GitHub Repository：https://github.com/adobe-research/dynasaur
- arXiv Technical Paper：https://arxiv.org/pdf/2411.01747
DynaSaur's Application Scenarios
- Automated Customer Service: As a chatbot, dynamically generate solutions to handle customer inquiries and problems, providing more flexible and personalized service.
- Intelligent Personal Assistant: As a personal assistant, help users manage schedules, execute tasks, search for information, and dynamically learn new skills according to users' needs.
- Software Development Assistance: Help software developers automatically generate code, solve programming problems, or serve as an automated testing agent in software testing.
- Education and Training: In the education field, as a teaching assistance tool, dynamically adjust teaching content and difficulty according to students' learning progress and needs.
- Data Science and Analysis: Handle complex data analysis tasks, dynamically generate data processing workflows, and help researchers and analysts conduct data exploration and pattern recognition more efficiently.