Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Affordable Generative Agents

Yangbin Yu, Qin Zhang, Junyou Li, Qiang Fu, Deheng Ye

TL;DR

Affordable Generative Agents (AGA) address the high cost of sustaining believable LLM-based agents in open-world settings by decoupling agent-environment inference from inter-agent dialogue. The framework introduces Lifestyle Policy to reuse plans and reduce redundant LLM reasoning for environment interactions, and Social Memory to compress and structure social context for inter-agent conversations, enabling cheaper yet coherent multi-agent simulations. Key contributions include Plan Decomposition, Policy Reuse, and Social Memory with Summary Events plus Relationship & Feeling, plus systematic evaluation in Generative Agents town and VirtualHome showing substantial token-cost reductions with maintained believability. The work also analyzes limits of emergent behavior in fixed environments and proposes mind wandering and related strategies to broaden behavioral richness, highlighting practical implications for scalable, believable agent-based simulations.

Abstract

The emergence of large language models (LLMs) has significantly advanced the simulation of believable interactive agents. However, the substantial cost on maintaining the prolonged agent interactions poses challenge over the deployment of believable LLM-based agents. Therefore, in this paper, we develop Affordable Generative Agents (AGA), a framework for enabling the generation of believable and low-cost interactions on both agent-environment and inter-agents levels. Specifically, for agent-environment interactions, we substitute repetitive LLM inferences with learned policies; while for inter-agent interactions, we model the social relationships between agents and compress auxiliary dialogue information. Extensive experiments on multiple environments show the effectiveness and efficiency of our proposed framework. Also, we delve into the mechanisms of emergent believable behaviors lying in LLM agents, demonstrating that agents can only generate finite behaviors in fixed environments, based upon which, we understand ways to facilitate emergent interaction behaviors. Our code is publicly available at: https://github.com/AffordableGenerativeAgents/Affordable-Generative-Agents.

Affordable Generative Agents

TL;DR

Affordable Generative Agents (AGA) address the high cost of sustaining believable LLM-based agents in open-world settings by decoupling agent-environment inference from inter-agent dialogue. The framework introduces Lifestyle Policy to reuse plans and reduce redundant LLM reasoning for environment interactions, and Social Memory to compress and structure social context for inter-agent conversations, enabling cheaper yet coherent multi-agent simulations. Key contributions include Plan Decomposition, Policy Reuse, and Social Memory with Summary Events plus Relationship & Feeling, plus systematic evaluation in Generative Agents town and VirtualHome showing substantial token-cost reductions with maintained believability. The work also analyzes limits of emergent behavior in fixed environments and proposes mind wandering and related strategies to broaden behavioral richness, highlighting practical implications for scalable, believable agent-based simulations.

Abstract

The emergence of large language models (LLMs) has significantly advanced the simulation of believable interactive agents. However, the substantial cost on maintaining the prolonged agent interactions poses challenge over the deployment of believable LLM-based agents. Therefore, in this paper, we develop Affordable Generative Agents (AGA), a framework for enabling the generation of believable and low-cost interactions on both agent-environment and inter-agents levels. Specifically, for agent-environment interactions, we substitute repetitive LLM inferences with learned policies; while for inter-agent interactions, we model the social relationships between agents and compress auxiliary dialogue information. Extensive experiments on multiple environments show the effectiveness and efficiency of our proposed framework. Also, we delve into the mechanisms of emergent believable behaviors lying in LLM agents, demonstrating that agents can only generate finite behaviors in fixed environments, based upon which, we understand ways to facilitate emergent interaction behaviors. Our code is publicly available at: https://github.com/AffordableGenerativeAgents/Affordable-Generative-Agents.
Paper Structure (32 sections, 1 equation, 11 figures, 6 tables)

This paper contains 32 sections, 1 equation, 11 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM Agents
  4. The Cost of LLM Interaction
  5. Method
  6. Agent-environment interactions
  7. Plan Decomposition
  8. Policy Reuse
  9. Inter-agent interactions
  10. Experimental Setup
  11. Environment
  12. Language Models Adopted
  13. Evaluation Protocol
  14. Evaluations
  15. Results on Generative Agents
  16. ...and 17 more sections

Figures (11)

  • Figure 1: Our proposed method for optimizing (a) Agent-environment interactions and (b) Inter-agent interactions. The gray arrows represent the baseline (Generative Agents) implementation, while the blue arrows represent our approach.
  • Figure 2: An example of relational evolution driven by the updating of Social Memory following conversational interactions.
  • Figure 3: Ablation study on token consumption. Baseline means Generative Agents. Lifestyle Policy and Social Memory mean only using the corresponding module.
  • Figure 4: Accumulative token consumption over game time for different methods
  • Figure 5: Average relationship score map between agents. The x and y axes represent the initials of the names of 25 agents. When identical abbreviations occur, the full names are retained.
  • ...and 6 more figures