Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLM Agents Beyond Utility: An Open-Ended Perspective

Asen Nachkov, Xi Wang, Luc Van Gool

TL;DR

The paper investigates whether pretrained LLM agents can transcend narrow task-solving to become open-ended entities that autonomously generate goals and leave persistent artifacts. It extends the ReAct framework with autonomous task generation, a simple memory hierarchy (short-term run buffers and long-term persistent files), and basic file I/O tools to support continuity and self-reference. Qualitative results show the agent can follow complex instructions, solve self-generated tasks, and accumulate knowledge across runs, but it remains sensitive to prompt design, can loop on task generation, and cannot form an explicit self-representation of its own state. The work highlights both the promise of open-ended LLM agents and the current limitations, outlining concrete directions for memory management, productive exploration, and long-horizon goal pursuit in future training and system design.

Abstract

Recent LLM agents have made great use of chain of thought reasoning and function calling. As their capabilities grow, an important question arises: can this software represent not only a smart problem-solving tool, but an entity in its own right, that can plan, design immediate tasks, and reason toward broader, more ambiguous goals? To study this question, we adopt an open-ended experimental setting where we augment a pretrained LLM agent with the ability to generate its own tasks, accumulate knowledge, and interact extensively with its environment. We study the resulting open-ended agent qualitatively. It can reliably follow complex multi-step instructions, store and reuse information across runs, and propose and solve its own tasks, though it remains sensitive to prompt design, prone to repetitive task generation, and unable to form self-representations. These findings illustrate both the promise and current limits of adapting pretrained LLMs toward open-endedness, and point to future directions for training agents to manage memory, explore productively, and pursue abstract long-term goals.

LLM Agents Beyond Utility: An Open-Ended Perspective

TL;DR

The paper investigates whether pretrained LLM agents can transcend narrow task-solving to become open-ended entities that autonomously generate goals and leave persistent artifacts. It extends the ReAct framework with autonomous task generation, a simple memory hierarchy (short-term run buffers and long-term persistent files), and basic file I/O tools to support continuity and self-reference. Qualitative results show the agent can follow complex instructions, solve self-generated tasks, and accumulate knowledge across runs, but it remains sensitive to prompt design, can loop on task generation, and cannot form an explicit self-representation of its own state. The work highlights both the promise of open-ended LLM agents and the current limitations, outlining concrete directions for memory management, productive exploration, and long-horizon goal pursuit in future training and system design.

Abstract

Recent LLM agents have made great use of chain of thought reasoning and function calling. As their capabilities grow, an important question arises: can this software represent not only a smart problem-solving tool, but an entity in its own right, that can plan, design immediate tasks, and reason toward broader, more ambiguous goals? To study this question, we adopt an open-ended experimental setting where we augment a pretrained LLM agent with the ability to generate its own tasks, accumulate knowledge, and interact extensively with its environment. We study the resulting open-ended agent qualitatively. It can reliably follow complex multi-step instructions, store and reuse information across runs, and propose and solve its own tasks, though it remains sensitive to prompt design, prone to repetitive task generation, and unable to form self-representations. These findings illustrate both the promise and current limits of adapting pretrained LLMs toward open-endedness, and point to future directions for training agents to manage memory, explore productively, and pursue abstract long-term goals.
Paper Structure (4 sections, 2 figures)

This paper contains 4 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Method -- Elements of an Open-Ended Agent
  3. Qualitative Results
  4. Conclusion & Future Work

Figures (2)

  • Figure 1: The open-ended LLM agent loop. The agent takes in user prompts and feedback (step 2.) but the task it attempts to solve is set by itself (step 3.). The loop around steps 4. to 6. represents the standard ReAct design pattern for solving the task. Afterwards, the agent stores a summary of this run into a storage medium (step 7.), which can be accessed using the right tools. Within a single full run (steps 1. to 7.) all interaction messages are stored in a buffer, representing the agent's short term memory, from which the input prompts to the LLM in each step are constructed (colored boxes represent tokens). The storage in step 7. represents the agent's long-term memory.
  • Figure 2: A trajectory of sampled states. The open-ended agent can generate tasks while incorporating user feedback. Its impact can be assessed through its artifacts in the environment.