Generative Agents: Interactive Simulacra of Human Behavior

Motivation

  • ​LLMs  are everywhere and are being used for everything.

  • Idea is to develop "generalized" agents where "generalized" captures the following:

    • Long-term coherence

    • Retrieval of relevant events

    • Reflections of events to draw high-level inferences

    • Planning and reacting in a non sub-optimal way (actions make sense now and in the future)

Experiment Setup

  • ​25 Agents are initialized where Agent -> LLM

 

  • Each agent is initialized with a natural language paragraph with following information:

    • Identity (Name, Occupation, Nature)

    • Information about social and physical environment

    • Information about relationships

    • Initial guided directive to their actions

Experiment Setup

  • Agents communicate with the environment and with other agents through Natural Language.

  • With Environment:

    • "Isabella Rodriguez is checking her emails"

    • "Isabella Rodriguez is talking with her family on the phone"

  • With other Agents:

Generative Agent Architecture

  • ​Due to finite context length, we cannot just put all the experience of the agent in an LLM Prompt.

  • Challenge is to extract relevant pieces of memory when needed.

  • Their approach is the creation of a "Memory Stream" alongside a retrieval mechanism.

Memory Stream

  • Observations are events directly perceived by an agent.

  • Can be behaviors performed by the agent themselves or behaviors perceived from other agents.

  • Stored in the Memory Stream with the following information:

    • Natural Language description

    • Creation time-stamp

    • Most Recently Accessed time-stamp

Observations

Memory Stream

  • Reflections are higher-level and more abstract thoughts generated by the agent.

  • Generated periodically by prompting the model with insight-extracting prompts.

  • Example

    1. Klaus Mueller is writing a research paper.

    2. Klaus Mueller enjoys reading a book on gentrification [.....]

    • ​​​​​​​What 5 high-level insights can you infer from the above statements?

Reflections

Memory Stream

  • Plans describe a future sequence of actions for the agent.

  • A Plan includes the following:

    • Location

    • Starting Time

    • Duration

  • Plans are created top-down and recursively

    • Prompt the model with agent summary and their previous day.

    • LLM gives a brief plan for the entire day.

    • Recursively break down the plan for the day into granular time scales.

Plans

Retrieval

  • How to retrieve relevant experiences from the agent's memory stream?

  • They consider 3 parameters while designing the retrieval function

    • Recency: Higher score to memory stream objects that were recently accessed (Exponential decay function)

    • Importance: Simple! Just ask the LLM.

      • Brushing Teeth : 1 (Mundane Task)

      • Breakup : 10 (Big Event)

    • Relevance: Context relevance. Use LLM to generate embeddings for all objects in memory stream and every incoming query.

Retrieval

Emergent Social Behaviors

  • Observed 3 prominent social behaviors.

    • Information Diffusion

      • Sam tells Tom about his candidacy in the local elections  which soon becomes the talk of the town.

    • Relationship Memory

      • Sam and Latoya do not know each other but in their initial meet, Latoya tells Sam she's working on a photography project.  In a later interaction between them,  Sam asks Latoya about her project progress.

    • Coordination

      • Isabella is initialized with an intent to organize a Valentines Day party. She asks her friend Maria for help  to which Maria agrees and they both successfully organize the party inviting many guests.

QUESTIONS ?

BabootaRahul

rahulbaboota

Generative Agents

By Rahul Baboota

Made with Slides.com

Generative Agents

An Introduction to the different methods of Feature Selection in Python .

  • 119

More from Rahul Baboota