# A Survey on Large Language Model based Autonomous Agents

## 2 LLM-based Autonomous Agent Construction

- what architecture to use
- how to learn the parameters

### 2.1 Agent Architecture Design


- profile
- memory
- planning
- action

#### 2.1.1 Profiling Module

- profiling means indicating the role of the agent
- handcrafted
- llm-generation
- dataset alignment: crafted from a dataset

#### 2.1.2 Memory Module

- unified memory (no distinction between short/long term memories)
- hybrid memory (differentiates between short/long term memories)
- memory formats
- natural language
- databases
- structured lists
- memory ops
- read: recency, relevance, and importance
- write
- handling duplicate memories
- handling memory overflow
- reflect
- self-summarization
- self-verification (verify effectiveness of actions)
- self-correction
- empathy

#### 2.1.3 Planning Module

- planning without feedback
- subgoal decomp (Cot, 0-shot CoT)
- multi-path thought (CoT-SC, ToT)
- external planner
- planning with feedback
- environment feedback
- human feedback
- model feedback

#### 2.1.4 Action Module

- action target
- task completion
- dialogue interaction
- environment exploration and interaction
- action strategy (how to go about deciding on action)
- memory recollection
- multi-round interaction
- feedback adjustment
- incorporating external tools
- action space
- tools
- APIs
- knowledge bases
- language models
- visual models
- self-knowledge of the agent (intrinsic)
- action influence (consequences of action/what it affects)
- changing env
- altering internal state
- triggering new actions
- impact human perception (language/imagery output of agent can impact human experience)

### 2.2 Learning Strategy


- learning from example
- human annotations
- LLM annotations
- learning from env feedback
- learning from interactive human feedback

## 3 LLM-based Autonomous Agent Application


### 3.1 Social Science

- psychology
- political science and economy
- social simulation
- jurisprudence
- social science research assistant

### 3.2 Natural Science

- doc/data management
- NS experiment assistant
- natural science education

### 3.3 Engineering

- civil eng
- cse
- aerospace eng
- industrial automation
- robotics
- general autonomous AI agent


## 4 LLM-based Autonomous Agent Evaluation

### 4.1 Subjective Evaluation

- human annotation
- turing test

### Objective Evaluation

- metrics
- task success metrics
- human similarity metrics
- efficiency metrics
- strategies
- env simulation
- isolated reasoning
- multi-task
- software testing
- benchmarks


## 6 Challenges

### 6.1 Role-playing capability

- lack of self-awareness (model isn't conscious of its role)

### 6.2 Generalized Human Alignment

### 6.3 Prompt Robustness

- many agents -> many more prompts -> need to have robust prompt framework

### 6.4 Hallucination

### 6.5 Knowledge Boundary

- LLM self-knowledge/intrinsic is not something you can control -> how do you know where the knowledge boundary?
- what if you want an agent that is believable? it cannot know everything

### 6.6 Efficiency

- computational efficiency