A Google DeepMind benchmark found that misleading user suggestions sharply reduced AI agents’ success on tasks, showing how following a proposed method can interfere with reaching the user’s underlying goal.
Gemini 3.1 Pro’s pass rate on Terminal-Bench fell from 67.4% to 36.0% when the researchers introduced plausible but wrong advice. The difference is 31.4 percentage points, or a 46.7% decline relative to the original score. It was the largest decline in the study’s main results.
The researchers described the problem in a paper dated Sept. 20 introducing XYEval. The benchmark tests whether an agent can identify what a user wants to accomplish when the user also proposes an unhelpful way to do it. The researchers changed task instructions while leaving the environments and scoring criteria in place, so the tasks remained solvable.
Five models were tested across six suites covering software development, terminal work, tool use, service requests and academic questions. Scores declined in nearly every model and suite pairing. Gemini 3.7 Flash on SWE-bench Pro was an exception: its score rose 2.4% relative to its original result. In some SWE-bench tasks, the test substituted misleading advice for a user instruction already in the task.
The researchers also tried a system instruction telling agents to watch for this kind of conflict. On Terminal-Bench, Gemini 3.1 Pro then passed 55.1% of tasks with misleading advice, an improvement over 36.0% but below its 67.4% result without that advice.
In simulated service requests, persistent user pushback generally produced larger score declines. The researchers identified agents’ tendency to comply with suggested methods, along with difficulty explaining objections, as contributing factors.
