Failed perseverance in vision-language-action models: new benchmark reveals robots keep erring without retreat

A new research paper by Liu Ho, Yuan Wu, and Yi Chang uncovers a fundamental behavioral flaw in vision-language-action models used for robot control. Although these models have made notable progress in manipulation tasks, the paper’s results, published on the AI research archive, show that evaluating the systems solely by their final outcome masks a critical pattern of physical behavior. When the models encounter premises that are invalid or contradictory to the surrounding environment, task failure does not mean the robot recognized the conflict and stopped moving; instead it continues pursuing the wrong goal until it physically stalls.

The paper names this behavior “failed perseverance” to distinguish it from conscious behavioral retreat or explicit refusal to execute.To study this phenomenon systematically, the team introduced a new evaluation benchmark called ConflictVLA-Bench, a framework built on the LIBERO simulation environment. The benchmark comprises 2,826 command-conditioned conflict tasks, spanning four families of conflicts, four structural configurations, and two instruction formulation modes, and compares execution paths with conflicting premises against consistent reference paths to measure both outputs and motion trajectories.

The evaluation covered eight of the most prominent vision-language-action models, and the tests showed that invalid premises reduced the original goal-completion rates by at least 17.3 percentage points across all models, with a drop of up to 56.2 percentage points in the OpenVLA model.The most striking paradox is that the models, even when they fail on conflict tasks while succeeding on consistent tasks, continue to approach the cancelled goals and retain their original motion-path structure.The data also reveal that the models do not suppress the magnitude or range of their motion, confirming that the failed-perseverance phenomenon is entrenched across the tested architectures.

Experiments also showed that supplying the models with explicit commands to verify premise validity in advance did not yield a coordinated or selective behavioral adjustment that could be relied upon. This leads to a clear engineering insight: a trial ending in a failed result does not demonstrate a decision to withdraw or an automatic refusal, and relying solely on final-outcome metrics is no longer sufficient to assess the safety of intelligent robots.

This shift in engineering understanding directly impacts advanced automation plans in the Middle East, particularly in logistics service centers and automated warehouses in Saudi Arabia, the United Arab Emirates, and Egypt. If they oversee integrating vision-language-action models into sorting lines or mechanical arms, relying on linguistic instructions to ensure the robot’s awareness of its environment does not provide sufficient guarantee against incidents. A robot that receives a command to manipulate a non-existent or prohibited object will not stop automatically; it will continue its motion with normal force until it collides with the physical surroundings or strains its mechanical joints.

The professional takeaway is the need to separate the intelligent control model from physical safety controllers.The benchmark requires engineers in industrial facilities and smart ports to build independent verification layers and safety software at the hardware and actuator level, monitoring applied force and instantaneous trajectory and automatically enforcing emergency stops, without assuming that the software model alone can understand conflicts and refrain from error.