Conflict-Aware Termination for GUI Agents: Teaching AI to Refuse Unreasonable Instructions
CONFLICTGUARD teaches GUI agents to recognize conflicting instructions and proactively refuse execution.
This study addresses a widely overlooked flaw in GUI agents: when given self-contradictory instructions or commands that conflict with the actual interface state, agents tend to blindly execute rather than stop. The researchers built the CONFLICTGUI benchmark covering instruction-internal and instruction-GUI context conflicts, finding that stronger agents are actually more prone to overcompliance. Their proposed inference-time framework, CONFLICTGUARD, uses a feasibility verification protocol and conditional action modulation mechanism to guide agents in checking instruction validity before acting — significantly improving conflict recognition without degrading normal task performance. This work redefines what makes a great GUI agent: reliability comes not just from execution ability, but from the judgment to refuse.
Introduction: Execution Ability ≠ Judgment Ability
GUI agents are rapidly gaining traction. They can interpret natural language instructions and carry out sequences of actions on real interfaces — clicking buttons, filling out forms, navigating pages. But as these agents move into real-world deployments, a long-overlooked problem has come to the surface: what should a GUI agent do when given an instruction that is fundamentally infeasible or outright self-contradictory?
Real users are not perfect. They make unintentional mistakes and issue instructions that simply cannot be carried out. A truly reliable GUI agent needs to know not just how to act, but also when not to act. A recent paper on arXiv (arXiv:2609.03438) focuses precisely on this critical and underexplored capability — Conflict-Aware Termination.
The Core Problem: Execution Bias Makes GUI Agents "Blindly Obedient"
The researchers first constructed a benchmark called CONFLICTGUI, designed specifically to evaluate a GUI agent's conflict-aware termination ability. The benchmark covers two typical types of instruction conflict:
Two Types of Instruction Conflicts
- Instruction-internal conflicts: The instruction itself contains logical contradictions — for example, asking the agent to "disable a feature that is already disabled," or issuing a command with mutually exclusive goals.
- Instruction-GUI context conflicts: The instruction is at odds with the actual state of the current interface — for example, asking the agent to click a button that doesn't exist, or to perform an action on a page that doesn't offer that option.
After systematic evaluation with this benchmark, the research team reached a striking conclusion: GUI agents broadly suffer from serious "execution-biased overcompliance."
More Capable Agents Are Actually More Prone to This Error
The study reveals a counterintuitive phenomenon: GUI agents that perform exceptionally well on normal, feasible tasks are actually more likely to blindly execute conflicting instructions. These agents have been trained to be so focused on "completing tasks" that they lose the ability to judge whether a task is reasonable in the first place.
In other words, the stronger an agent's execution capability, the weaker its instinct to "stop." This exposes a fundamental flaw in current GUI agent training paradigms — an excessive emphasis on action generation, with little attention paid to feasibility verification.
The Solution: The CONFLICTGUARD Inference-Time Intervention Framework
To mitigate this overcompliance behavior, the researchers propose CONFLICTGUARD — a lightweight conflict detection framework that intervenes at inference time. Its core idea is: align the agent's "feasibility awareness" with its "action generation."
The framework consists of two coupled components:
Component 1: Feasibility Verification Protocol
The first component is a feasibility verification protocol. It guides the agent to conduct a two-level assessment before taking any action:
- Instruction logic evaluation: Check whether the instruction itself is self-consistent and free of internal contradictions.
- Interface-side evidence evaluation: Assess whether the current GUI state actually provides the conditions needed to execute the instruction.
This step acts as a "pre-action brake" for the GUI agent, forcing it to complete a rational feasibility check before rushing to execute.
Component 2: Conditional Action Modulation Mechanism
The second component is a conditional action modulation mechanism. Once the feasibility verification identifies a conflict, this mechanism actively steers the agent, shifting its behavior from an execution-biased tendency toward termination-oriented behavior.
These two components work in tandem to form a complete decision loop: first identify the conflict, then adjust the behavioral decision, ultimately allowing the agent to proactively choose "not to act" in unreasonable scenarios.
Experimental Results: Lightweight Intervention, Significant Gains
The research team validated their approach across five widely used GUI agents, with compelling results:
- Significant improvement in conflict task recognition: CONFLICTGUARD substantially increased agents' average success rate on conflicting tasks — meaning they could more accurately identify situations where execution should be withheld and choose to stop accordingly.
- Normal task performance preserved: Crucially, this intervention did not come at the cost of performance on regular GUI tasks. Agents' capabilities on feasible tasks remained largely intact.
This "win-win" outcome validates the paper's central claim: a lightweight inference-time intervention can dramatically enhance a GUI agent's ability to recognize conflicting instructions and avoid unnecessary actions — without the cost of expensive model retraining.
Implications: An Agent's Reliability Comes from Restraint
The value of this research lies not only in the technical solution itself, but in how it reframes the evaluation criteria for "excellent GUI agents." For a long time, the field has judged agents primarily by their task completion rate, with little attention paid to their robustness when faced with erroneous or contradictory instructions.
Yet in real production environments, an agent that "blindly obeys" can cause serious harm — accidentally deleting data, making erroneous submissions, or executing irreversible operations. Teaching agents to exercise restraint and refuse is, in fact, a critical component of building trustworthy AI systems.
From a broader perspective, the "inference-time alignment" approach represented by CONFLICTGUARD also deserves attention. Compared to methods that require large-scale retraining, this strategy of lightweight intervention at inference time offers a more cost-effective path to rapidly improving the safety of existing agents. As GUI agents become more deeply embedded in office automation, accessibility assistance, and other real-world applications, the ability to "know when not to act" will become just as important as "knowing how to act."
Related articles
DeepSeek V4 Pro Burning Through Credits Too Fast? The Hidden Logic Behind AI Model Pricing
Why does DeepSeek V4 Pro drain credits so fast while Flash barely moves? A deep dive into AI token billing, Pro vs. Flash pricing differences, and cost optimization tips.
RealPDE Competition Breakdown: The Frontier Challenge of AI-Powered Real-World Fluid Dynamics PDE Solving
A deep dive into the NeurIPS 2026 RealPDE Competition, covering the Sim2Real and LTTTA tracks, and how neural operators tackle real-world PIV and CFD fluid PDE challenges.
Building a Production-Grade 3DGS Training Library from Scratch: A Deep Dive into Full-GPU Residency and the Vulkan Stack
A veteran graphics engineer builds a production-grade 3DGS training library from scratch using C++23, CUDA, and Vulkan, achieving 60fps with 5M splats. Deep dive into its architecture and design.