Runaway is Ashamed, But Helpful: On the Early-Exit Behavior of Large Language Model-based Agents in Embodied Environments

Qingyu Lu, Liang Ding, Siyi Cao, Xuebo Liu, Kanjian Zhang, Jinxia Zhang, Dacheng Tao

Abstract

Agents powered by large language models (LLMs) have demonstrated strong planning and decision-making capabilities in complex embodied environments. However, such agents often suffer from inefficiencies in multi-turn interactions, frequently trapped in repetitive loops or issuing ineffective commands, leading to redundant computational overhead. Instead of relying solely on learning from trajectories, we take a first step toward exploring the early-exit behavior for LLM-based agents. We propose two complementary approaches:

  1. an intrinsic method that injects exit instructions during generation, and 2. an extrinsic method that verifies task completion to determine when to halt an agent's trial. To evaluate early-exit mechanisms, we introduce two metrics: one measures the reduction of redundant steps as a positive effect, and the other evaluates progress degradation as a negative effect. Experiments with 4 different LLMs across 5 embodied environments show significant efficiency improvements, with only minor drops in agent performance. We also validate a practical strategy where a stronger agent assists after an early-exit agent, achieving better performance with the same total steps. We will release our code to support further research.

Introduction

LLM-based agents are increasingly deployed in complex "embodied" environments—virtual or physical spaces where they must plan and execute actions to achieve specific goals. However, these agents often suffer from inefficiencies due to a lack of self-awareness regarding their own progress. When faced with difficult tasks or errors, they frequently fall into "unproductive loops," repeating ineffective commands until a hard step limit is reached. In cloud-based systems, this wastes tokens; in physical robotics, it wastes energy and risks mechanical damage. This research addresses these inefficiencies by introducing "early-exit" behaviors, enabling agents to recognize when they are stuck or have completed a task.

Methodology

The authors propose two complementary strategies for early-exit behavior:

  1. Intrinsic Early Exit: This "plug-and-play" method injects specific instructions into the agent's prompt (e.g., "You can exit if you think the task is complete"). This encourages the LLM to generate an "EXIT" action as part of its natural reasoning process.
  2. Extrinsic Early Exit: This involves an external verification module that monitors the interaction history. It periodically evaluates whether the agent should stop based on perceived lack of progress or goal reachability.
Comparison of Intrinsic and Extrinsic Early-Exit strategies.

To evaluate these mechanisms, the study introduces two novel metrics: Redundant Steps (RS), measuring unnecessary steps taken after progress has ceased, and Progress Degradation (PD), measuring the loss in task completion caused by premature halting. The methods were tested across five environments (AlfWorld, ScienceWorld, BabyAI, Jericho, and PDDL) using four LLM backbones: Llama 3.1 (8B/70B) and Mistral (7B/24B).

Results and Analysis

Experimental results demonstrate that early-exit mechanisms can reduce redundant steps by 50% to 70% with minimal impact on overall success rates.

Success Rate vs. Average Steps across different LLM backbones.

Key insights from the evaluation include:

  • The Impact of Model Scale: Larger models, such as Llama 3.1-70B, exhibit better judgment in self-termination. Smaller models tend to be more "trigger-happy," leading to higher progress degradation.
  • Efficiency Gains: In the AlfWorld environment, Llama 3.1-8B reduced its average interaction steps from 33.4 to 15.9 using intrinsic exit. For Llama 3.1-70B, progress rates remained stable (dropping only ~2%) while redundant steps were nearly halved.
  • Weak-to-Strong Handoff: A practical application of this work is the "weak-to-strong" handoff. By using a smaller agent with early exit enabled to attempt a task first, and then handing off to a larger model only when the first agent gives up, the researchers achieved a 10% higher success rate than the baseline within the same total step budget.

Conclusion

This work demonstrates that teaching LLM agents "when to quit" significantly improves computational efficiency in embodied tasks. While current methods do not eliminate all redundancy and face challenges in exploratory scenarios, they provide a robust foundation for more self-aware AI agents. Future research may explore fine-tuning agents specifically for early-exit trajectories to further refine their self-assessment capabilities.

Summary updated

Abstract

Agents powered by large language models (LLMs) have demonstrated strong planning and decision-making capabilities in complex embodied environments. However, such agents often suffer from inefficiencies in multi-turn interactions, frequently trapped in repetitive loops or issuing ineffective commands, leading to redundant computational overhead. Instead of relying solely on learning from trajectories, we take a first step toward exploring the early-exit behavior for LLM-based agents. We propose two complementary approaches: 1. an $\textbf{intrinsic}$ method that injects exit instructions during generation, and 2. an $\textbf{extrinsic}$ method that verifies task completion to determine when to halt an agent's trial. To evaluate early-exit mechanisms, we introduce two metrics: one measures the reduction of $\textbf{redundant steps}$ as a positive effect, and the other evaluates $\textbf{progress degradation}$ as a negative effect. Experiments with 4 different LLMs across 5 embodied environments show significant efficiency improvements, with only minor drops in agent performance. We also validate a practical strategy where a stronger agent assists after an early-exit agent, achieving better performance with the same total steps. We will release our code to support further research.
BibTeX
@inproceedings{lu2025runaway,
  author = {Qingyu Lu and Liang Ding and Siyi Cao and Xuebo Liu and Kanjian Zhang and Jinxia Zhang and Dacheng Tao},
  title = {{Runaway is Ashamed, But Helpful: On the Early-Exit Behavior of Large Language Model-based Agents in Embodied Environments}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=5sRxW5GCYX},
}
Figure