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.

Introduction

While Large Language Model (LLM) agents excel at high-level planning, they often lack sufficient "self-awareness" regarding their own operational failures. In multi-turn embodied tasks—such as searching for an object in a complex household environment—agents frequently become trapped in unproductive loops or continue to issue commands long after a task has become impossible to complete. This behavior results in significant redundant computational overhead and, in physical robotics, unnecessary energy consumption and mechanical wear. This paper addresses this gap by exploring "early-exit" strategies, enabling agents to proactively terminate trials when progress stalls.

Methodology: Intrinsic vs. Extrinsic Early Exit

The researchers propose two "plug-and-play" strategies that do not require additional model training:

  1. Intrinsic Early Exit: This method modifies the agent's prompt to include specific termination instructions (e.g., "You can exit if you think the task is complete"). This encourages the agent to generate a specialized EXIT action when it perceives it has either succeeded or reached an impasse.
  2. Extrinsic Early Exit: An external verification module, powered by the same LLM backbone, monitors the interaction history. After each step, this module performs a binary check to decide whether the agent should continue or halt based on the likelihood of success or loop detection.
Illustration of Intrinsic and Extrinsic Early Exit mechanisms compared to the standard ReAct baseline.

New Metrics for Agent Efficiency

To rigorously evaluate these mechanisms, the authors introduce two novel metrics:

  • Redundant Steps (RS): Quantifies the number of steps taken after the final successful subgoal was achieved. Reducing RS is the primary goal of early exit.
  • Progress Degradation (PD): Measures the loss in task completion progress caused by terminating a trial prematurely. This serves as a check against overly aggressive exiting.

Experimental Results and Insights

The methods were tested across five diverse environments (ALFWorld, ScienceWorld, BabyAI, Jericho, and PDDL) using four models: LLaMA 3.1 (8B/70B) and Mistral (7B/24B).

Efficiency Gains

Across all environments, early-exit mechanisms reduced redundant steps by 50% to 70%. As shown in the performance analysis, larger models like LLaMA 3.1-70B demonstrate a more favorable trade-off, maintaining high success rates while significantly lowering the average number of steps per task.

Success Rate vs. Average Steps for various LLMs. The red stars indicate the performance shift when Early Exit is applied to the ReAct baseline.

The "Relay" Strategy

A key practical finding is the effectiveness of a "weak-to-strong" relay. A smaller, cheaper agent (e.g., Mistral-24B) can handle the initial phases of a task and exit early if it encounters difficulty. A stronger agent (e.g., LLaMA-70B) then takes over the remaining step budget. In ALFWorld, this hybrid approach achieved a 69.4% success rate, outperforming the Mistral-24B baseline (58.2%) within the same total step budget.

Conclusion

This study demonstrates that "knowing when to quit" is a vital component of agentic efficiency. By implementing simple intrinsic or extrinsic exit signals, LLM agents can drastically reduce wasted computation with minimal impact on task success. Future work will look toward fine-tuning models specifically for self-termination and applying these strategies to real-world robotic systems where efficiency is paramount.

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 **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.
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