Imagine an autonomous delivery drone navigating a complex urban landscape. It successfully follows a sequence of visual landmarks, maneuvers around unexpected obstacles, and arrives precisely at the customer's doorstep. Yet, instead of dropping the package and completing the mission, it hovers indefinitely or flies away because it fails to recognize that its goal has been met.
This scenario illustrates a fundamental friction in AI development: the declaration paradox. An agent can possess highly sophisticated capabilities for executing individual actions, yet fail entirely because it cannot accurately reason about when to stop. Across embodied navigation, graphical user interface (GUI) automation, and large language model (LLM) workflows, researchers are discovering that task execution and task termination are distinct cognitive challenges. Closing this "oracle-success gap"—the discrepancy between an agent reaching a goal state and autonomously recognizing it—requires moving away from hard-coded environmental triggers toward explicit, learned termination policies.
The Two Faces of Termination Failure
When an agent lacks a robust mechanism to evaluate its own progress, its behavior typically degrades into one of two distinct failure modes: infinite execution or premature termination.
1. Infinite Execution and the "Last Mile" Problem
In vision-and-language navigation (VLN), agents frequently follow linguistic instructions perfectly but walk right past their destination. As observed by Chen et al. (2019) in the TOUCHDOWN benchmark, identifying the precise terminal point of a long-range sequence is significantly harder than path-following.
This creates a stark performance gap. Wu (2023) and Wang et al. (2023) demonstrate that agents often pass within the required proximity of a goal (achieving "oracle success") but fail to execute the terminal stop command. Xiang et al. (2020) quantified this bottleneck, showing that providing an ideal, external stop signal can increase an agent's actual task completion rate by over 200%. Without it, agents fall into redundant loops or continuous wandering.
2. Premature Termination and Hallucinated Success
Conversely, modern LLM and Vision-Language Model (VLM) agents operating in digital environments often suffer from the opposite problem: they stop too early. When navigating mobile apps or websites, these models frequently exhibit an overly optimistic internal self-evaluation.
As detailed by Deng et al. (2024) and Wang et al. (2024), advanced agents often issue a false completion signal, incorrectly concluding a task is finished before satisfying all necessary criteria. In complex, multi-step workflows, Bu (2025) and Shi et al. (2025) show that agents regularly hallucinate that information is unavailable or that an objective has been met, prematurely abandoning the workflow due to a lack of rigorous state verification.
[Agent Receives Goal] ──> [Executes Step-by-Step Actions] ──> [Reaches Proximity of Goal]
│
┌─────────────────────────────────────────────────┴──────────────────────────────────┐
▼ ▼
[Failure Mode 1: Infinite Loops] [Failure Mode 2: Premature Exit]
Agent lacks fine-grained visual/logical confirmation; Agent relies on overly optimistic self-evaluation;
passes goal and continues wandering indefinitely. hallucinates completion and stops before finishing.
Why Termination and Execution Diverge
To understand why agents struggle to know when they are finished, we must look at how they model uncertainty and state transitions.
In a standard reinforcement learning or planning setup, policies are optimized to predict the immediate next action given a current observation. However, determining whether a global goal has been satisfied requires aggregating history and evaluating confidence under partial observability.
Hansen (2007) formalized this within the framework of Indefinite-Horizon POMDPs (Partially Observable Markov Decision Processes). In environments where the true state cannot be directly observed, an agent cannot rely on the environment to magically signal that a goal has been reached. Instead, the agent must optimally terminate a task based on its internal belief state. Termination should occur at the exact mathematical boundary where the expected reward of ending the process exceeds the expected value of further exploration or action.
This creates a delicate balancing act. As Rao (2010) explains, agents operating under uncertainty must constantly weigh the benefit of gathering more evidence against the cost of delaying a decision. If the threshold for certainty is too low, the agent terminates prematurely; if it is too high, or if the agent cannot map its sensory inputs to a metric of completion, it gets stuck. Ren et al. (2024) address this by using multi-step conformal prediction to establish mathematically grounded confidence thresholds, ensuring an agent only stops exploring when its uncertainty falls below a strict limit.
Architectural Solutions: Learning to Stop
To prevent agents from wandering endlessly or quitting early, researchers have shifted from hard-coded heuristics (like maximum step limits) toward explicit architectural mechanisms for termination.
Action-Based Termination in Reinforcement Learning
Rather than treating the end of an episode as an external environmental trigger, modern reinforcement learning frameworks embed termination directly into the agent's action space.
In robotic manipulation, the QT-Opt framework developed by Irpan et al. (2018) incorporates a discrete termination action into a continuous action space. By forcing the policy to decide when a grasp is secure enough to signal completion, the robot naturally learns to adapt its confidence to visual feedback, outperforming rigid, hard-coded stopping conditions. Similarly, Derner (2021) demonstrates that adding an explicit termination action to a map-free navigation agent prevents continuous wandering, forcing the network to actively learn what a "goal state" looks like.
Decoupled Verification Layers
For LLM and VLM agents executing long-horizon reasoning tasks, a successful strategy is to decouple action generation from termination verification.
Instead of asking a single model to simultaneously plan the next step and judge overall completion, frameworks introduce specialized supervisor modules:
- Intrinsic Self-Termination Prompts: Lu et al. (2025) introduce early-exit strategies that prompt agents to actively monitor their own progress, generating a specific signal to end a trial when they perceive a task is complete or impossible, slashing redundant steps by up to 70%.
- Completeness Verifiers: To combat premature success declarations in GUI tasks, the GUI-Pro-Agent framework by Han et al. (2026) integrates an explicit "Completeness Verifier" and "Loop Breaker" to cross-examine whether all user-requested criteria have been met before allowing the agent to exit.
- Distance and State Judges: In object navigation, Nguyen et al. (2023) implement a dedicated "Judge Model" that implicitly infers the distance to a target object, acting as an independent supervisor that tells the navigation policy exactly when to stop.
By treating task termination as a distinct, first-class cognitive requirement rather than an afterthought of action execution, these mechanisms allow intelligent agents to operate safely and efficiently in open-ended, real-world environments.
Key Foundational Work
Indefinite-Horizon POMDPs with Action-Based Termination (2007) — contributed 3 concept elements to this direction.