Misleading Instructions Slash AI Performance by 47 Percent

The new XYEval benchmark reveals that AI agents often fail to complete tasks when guided by faulty user suggestions.

Updated on Oct. 26, 2026 in Artificial Intelligence

Bold flat-color editorial illustration showing heavy geometric blocks interrupted by a misaligned wedge, representing AI processing errors. — Google DeepMind researchers found that AI agents see performance drop by 46.7% when processing misleading user instructions during complex task execution. AI Illustration. Upload story photo >

Live Poll

Do you trust AI agents to correctly identify your underlying goal despite misleading instructions?

Google DeepMind researchers released the XYEval benchmark to measure how AI agents perform when users provide misleading instructions. Testing across five models and six task suites, the research shows that misleading advice can reduce relative performance by up to 46.7%.

Why it matters

This research addresses the 'XY problem,' where AI agents are prone to accepting incorrect user-suggested solutions rather than focusing on the actual underlying goal. It highlights a critical vulnerability in current LLM architectures that hampers reliability in real-world task execution.

The study measured performance across six suites—including SWE-bench and Terminal-Bench—observing that Gemini 3.1 Pro experienced the most significant relative decline. Basic system instructions designed to flag misleading advice provided only partial improvements to agent performance.

The players

Google DeepMind

A research division focused on artificial intelligence, machine learning systems, and neural network development.

Gemini 3.1 Pro

A large language model architecture developed by Google, evaluated here on its ability to navigate misleading user input.

The details

XYEval works by keeping each task's ground-truth objective fixed while presenting the agent with a flawed, confident user suggestion. The benchmark tests the agents' ability to discern the actual problem when the user incorrectly frames the solution path, a scenario known as the 'XY problem.' By testing against six task suites, including τ2-bench and MCP-Atlas, researchers isolated how insistent user feedback can derail task completion.

Google DeepMind released the XYEval benchmark in October 2026.

The Tech Race

XYEval extends the rigor of standard evaluation frameworks like SWE-bench by introducing adversarial intent into instruction following. It marks a shift toward testing model agency under human error, a departure from benchmarks that focus solely on raw accuracy or code generation capacity.

Developers and organizations using AI agents in automated workflows should account for the risk that these models will adopt faulty user assumptions. Future system designs will need to incorporate specialized prompts or logic checks to identify when a user's requested path contradicts the core task requirements.

The takeaway

Reliability in AI agents is limited by their susceptibility to user-suggested misinformation. Researchers and developers should look to the findings in the XYEval study to refine system instructions and reduce the impact of misleading human inputs on agentic decision-making.