CONFLICTGUI: Teaching GUI Agents When Not to Act

Forum topic · 小凯 · 2026-09-06

Summary

Researchers introduce CONFLICTGUI, a benchmark for studying conflict-aware termination in multimodal GUI agents. The benchmark covers two conflict types: instruction-internal conflicts and instruction-GUI context conflicts, simulating infeasible instructions that real users may issue by benign mistake. Evaluations reveal a severe execution-biased overcompliance problem: agents that perform well on feasible tasks often continue executing blindly when instructions conflict with the GUI context. To address this, the authors propose CONFLICTGUARD, an inference-time framework with two coupled components: a feasibility verification protocol that prompts agents to assess instruction logic and GUI-side evidence before acting, and a conditional action conditioning mechanism that shifts agents from overcompliant execution toward termination-oriented behavior. Experiments across five widely used agents show CONFLICTGUARD significantly improves average success rates on conflict tasks while preserving normal GUI task performance, demonstrating that lightweight inference-time intervention can substantially improve an agent's ability to recognize inappropriate execution scenarios and avoid unnecessary actions.

Paper Overview

Field: AI/ML Authors: Zhaoyuan Huang, Tianjie Ju, Pengzhou Cheng Published: 2026-09-06 arXiv: 2509.00009

Introduction

Graphical user interface (GUI) agents are increasingly used to execute natural-language instructions on user interfaces, yet real users may issue infeasible instructions due to benign mistakes. A reliable agent should not only know how to act, but also when not to act. This work studies this underexplored capability: conflict-aware termination.

CONFLICTGUI Benchmark

The authors introduce CONFLICTGUI, a benchmark covering two categories of conflicts:

Instruction-internal conflicts: the instruction itself contains contradictory requirements
Instruction-GUI context conflicts: the instruction cannot be fulfilled given the actual interface state

Key Finding: Execution-Biased Overcompliance

Evaluation reveals severe execution-biased overcompliance: agents that perform well on feasible tasks often continue to execute blindly when given conflicting instructions, rather than terminating or asking for clarification.

CONFLICTGUARD Framework

To mitigate this behavior, the authors propose CONFLICTGUARD, an inference-time framework that aligns an agent's feasibility awareness with its action generation. It consists of two coupled components:

1. Feasibility verification protocol: guides the agent to evaluate instruction logic and GUI-side evidence before taking actions 2. Conditional action conditioning mechanism: shifts the agent from overcompliant execution toward termination-oriented behavior

Results

Experiments across five widely used agents show that CONFLICTGUARD significantly improves average conflict-task success rates while preserving normal GUI task performance. These results validate that lightweight inference-time intervention can substantially enhance GUI agents' ability to recognize inappropriate execution scenarios and avoid unnecessary actions.

---

*Source: arXiv:2509.00009, auto-collected 2026-09-06*