PranayPasula
commented
11 months ago My (quick) overview of Lilian Weng's blog post "LLM-powered Autonomous Agents" at https://lilianweng.github.io/posts/2023-06-23-agent/
Agent System Overview
- Introduction to Agents: The post begins with an introduction to agent systems, emphasizing their importance in AI and their ability to perform complex tasks.
- Agent Architecture: An explanation of the high-level architecture, including planning, memory, and tool use components, is provided.
Component One: Planning
Task Decomposition
- Hierarchical Planning: Lilian discusses how tasks can be broken down into sub-tasks, making them more manageable.
- Reinforcement Learning Integration: The integration of RL techniques for planning is highlighted.
- References
- Chain of thought (CoT;
- Wei et al. 2022)
- Tree of Thoughts (Yao et al. 2023)
- LLM+P (Liu et al. 2023)
Self-Reflection
- Meta-Cognition: The importance of self-reflection in agents, allowing them to learn from their experiences and adapt, is discussed.
- Self-Improvement: The agent's ability to improve itself through continuous learning is emphasized.
- References
- ReAct (Yao et al. 2023)
- Reflexion (Shinn & Labash 2023)
- Chain of Hindsight (CoH; Liu et al. 2023)
- Algorithm Distillation (AD; Laskin et al. 2023)
Component Two: Memory
Types of Memory
- Sensory Memory: I.e., learning embedded representations
- Short-Term and Long-Term Memory: Compared to in-context learning (ICL) and external storage (e.g., vector databases), respectively
- Memory Augmentation: Techniques for enhancing memory capabilities are discussed.
Maximum Inner Product Search (MIPS)
- Efficiency in Retrieval: MIPS is explained as a method for efficient retrieval of information from memory.
- Applications: The various applications and importance of MIPS in agent systems are highlighted.
- Some References (I selected)
- LSH (Locality-Sensitive Hashing)
- ANNOY (Approximate Nearest Neighbors Oh Yeah)
- ScaNN (Scalable Nearest Neighbors)
- More Approximate Nearest Neighbors (ANN) algorithms and performance here: ann-benchmarks.com
Component Three: Tool Use
Case Studies
- Real-World Applications: Several case studies are presented to illustrate how agents can be applied in different domains.
- References:
- MRKL (Karpas et al. 2022): A MRKL system is proposed to contain a collection of “expert” modules and the general-purpose LLM works as a router to route inquiries to the best suitable expert module.
- TALM (Tool Augmented Language Models; Parisi et al. 2022)
- Toolformer (Schick et al. 2023)
- Both TALM and Toolformer fine-tune an LM to learn to use external tool APIs
- HuggingGPT (Shen et al. 2023): uses ChatGPT as the task planner to select models available in HuggingFace platform according to the model descriptions
- Stages of References broken down
- Task planning
- Model selection
- Task execution
- Response generation
- Benchmarking Tool-based LLMs: API-Bank (Li et al. 2023) is a benchmark for evaluating the performance of tool-augmented LLMs.
Scientific Discovery Agent
- Automating Discovery: The potential for agents to automate scientific discovery is explored, with examples provided.
- References
- ChemCrow (Bran et al. 2023)
- Emergent autonomous scientific research capabilities of large language models Boiko et al. (2023)
Generative Agents Simulation
- Simulation Capabilities: Generative Agents (Park, et al. 2023): 25 virtual characters, each controlled by a LLM-powered agent, are living and interacting in a sandbox environment, inspired by The Sims. Generative agents create believable simulacra of human behavior for interactive applications.
Proof-of-Concept Examples
- Demonstrating Feasibility: Several proof-of-concept examples are provided to demonstrate the practicality and potential of the concepts discussed.
- References
Challenges
- Finite context length: The challenges related to scaling agent systems are discussed.
- Challenges in long-term planning and task decomposition::
- Ethical challenges: Resulting from the other limitations, lack of governance around tools used, which can lead to rebellious or harmful behavior
- Reliability of natural language interface: Something we've mentioned before. How the interface between LLMs and external is restrictive, which can be positive, but is also restrictive
Irony
ChatGPT-4 + plugins did a horrendous job of summarizing this even with guidance, so I wrote everything (and copied and pasted certain parts from Lilian's post) except for the ### and #### headers (which I explicitly provided in my prompt to ChatGPT-4 anyways, so I don't give it credit for this either).
My goodness! I don't want to have to use ReAct or even CoT just to get a decent quality answer from GPT-4 on something as simple as this. Lilian states that ChatGPT Plugins is a good example of LLMs augmented with tool use capability working practice, which I intentionally omitted from the overview. I guess it is a good example, but it has poor performance without significant prompt engineering, which I don't have the time to do at 8:40 AM.
Note: Some references don't have descriptions, but they can be summarized in one sentence. Check the abstract or Lilian's post.
For You, the reader
Limitations provided by Lilian are very limited, and I encourage folks to consider other patterns in limitations with the approaches, and more importantly, the paradigms mentioned, and to share these in the comments before this issue is closed out. I have several in mind, which I'll share after some others do.
bhavul
For team review: https://lilianweng.github.io/posts/2023-06-23-agent/