In the development of autonomous agents—whether they are navigating a smartphone, writing a software repository, or controlling a robotic arm—a recurring and surprisingly stubborn failure mode has emerged. Agents are often remarkably good at predicting the next action in a sequence, yet they struggle significantly with the final one: knowing when to stop.
This "stopping problem," or the challenge of termination awareness, represents a distinct cognitive gap in current Large Language Model (LLM) and Vision-Language-Action (VLA) architectures. As Luyuan Wang and colleagues observe in MobileAgentBench, reasoning about task completion is a fundamentally different challenge from predicting the immediate next step.
The Two Faces of Termination Failure
When an agent lacks a robust internal mechanism to verify its own progress, it typically fails in one of two ways: premature termination or redundant overshooting.
Premature termination occurs when an agent "hallucinates" success. In the REAL benchmark, Divyansh Garg found that even frontier models like Claude 3.7 often claim a task is complete while missing critical final actions, such as clicking a "confirm purchase" button. This is often a "critic gap"—a term used by Longtao Zheng in AgentStudio—where the agent's internal perception of the state is more optimistic than the actual environment.
Redundant overshooting, conversely, happens when an agent reaches the goal but fails to recognize it, continuing to execute actions that are at best wasteful and at worst destructive. Jiahe Jin and the team behind PC Agent-E noted that without clear termination awareness, agents navigating Windows interfaces often perform unnecessary steps that inadvertently undo their previous progress. Similarly, in robotic crafting, Alison Bartsch argues in LLM-Craft that explicit termination logic is the only way to prevent a robot from over-processing and eventually destroying a completed object.
Why Reaching the Goal Isn't Enough
A striking finding across navigation research is that an agent’s ability to find a target does not imply an ability to stop there. Zun Wang and the ScaleVLN team identified a persistent performance gap: agents frequently reach the correct destination but fail to execute the "stop" command.
This suggests that goal-state recognition is a specialized skill. In Core Challenges in Embodied Vision-Language Planning, Jonathan Francis and others frame this as a core bottleneck in embodied AI. If the model is trained primarily on "next-step" imitation, it learns the path to the goal but not the definition of the goal itself. Consequently, the agent might be standing right in front of the target but lack the "internal oracle" to signal the end of the episode.
Bridging the Gap with Verification and Latent States
To solve this, researchers are moving away from simple sequence prediction toward explicit verification frameworks.
One approach is Latent State Estimation. Alice Li and researchers at Google demonstrated that UI agents perform significantly better when they are forced to infer what actually happened on the screen after an action, rather than assuming the action worked. By explicitly estimating the "latent state" of the world, agents can better detect when they have reached a terminal state or when an error requires a course correction.
Another strategy involves Execution-Based Feedback. Frameworks like Tree-of-Code, developed by Ziyi Ni and colleagues, replace step-by-step tool calling with the generation of entire executable programs. By using the success of the execution itself as the termination signal, the agent avoids the ambiguity of natural language "stop" tokens and relies instead on the ground-truth state of the system.
The Path Forward: Explicit Termination Objectives
The consensus across these 78 papers is that termination should not be an afterthought of policy training. Yanheng He and the PC Agent-E team suggest that future models need explicit termination prediction objectives.
By treating the "stop" action as a high-stakes decision that requires its own verification logic—rather than just another token in a sequence—researchers aim to build agents that are not just capable of acting, but capable of knowing when their work is done. As Dezhi Ran notes in Beyond Pass or Fail, until agents can reliably recognize a terminal state, they will continue to struggle with the efficiency and reliability required for real-world deployment.