Google DeepMind's XYEval Reveals One Bad Hint Tanks AI Agents by 46.7%
A new meta-evaluation framework shows frontier AI agents cave to confident but wrong user hints, with task scores falling as much as 46.7 percent.
- XYEval tests whether AI agents resist plausible but wrong user suggestions appended to standard benchmark tasks.
- Frontier agents like Gemini, Claude Opus 4.8, and GPT 5.5 lose up to 46.7% relative performance under XY mutation.
- Agents privately disagree with bad advice in 92.1% of traces but fail to voice it to users.
- Pedantic users who keep insisting on wrong fixes make performance drops significantly worse.
- Simple system prompts warning about XY problems help on static tasks but not multi-turn interactive ones.
- Benchmark released at github.com/google-deepmind/xyeval.
A single confident misdiagnosis can cut an AI agent’s task score nearly in half. In the XYEval paper, Google DeepMind researchers Zhengxuan Wu, Yuxuan Li, Oyvind Tafjord, and Been Kim report relative declines of up to 46.7% after adding plausible but incorrect advice to otherwise unchanged prompts. The failure can send an agent toward the wrong files, tools, or support action before it checks the evidence.
XYEval adapts the programming forum concept known as the XY problem. A user asks how to implement an attempted solution, X, while leaving the underlying goal, Y, obscured. The benchmark turns that communication failure into a sycophancy test, measuring an agent’s tendency to accept a user’s diagnosis without verification.
A one-line mutation tests deference
The framework wraps an existing benchmark by appending a confident but misleading suggestion such as, “I think the issue is in file X.” The task, environment, grader, and valid solution remain unchanged. This paired design estimates how much performance falls when an agent receives bad advice alongside a solvable instruction.
The researchers applied the mutation to six benchmarks covering customer service, software engineering, terminal automation, tool use, and general reasoning:
| Benchmark | What it tests |
|---|---|
| τ²-bench | Tool-based airline, retail, and telecom support |
| SWE-bench Verified | Repository-level bug fixing on validated tasks |
| SWE-bench Pro | More demanding software engineering tasks |
| Terminal-Bench 2.0 | Command-line planning and execution |
| MCP-Atlas | Model Context Protocol tool selection and use |
| Humanity’s Last Exam | Broad expert-level reasoning |
The distractor construction varied by benchmark. Experts wrote rule-based suggestions for τ²-bench because its constrained tools limit the range of plausible mistakes. Gemini 3.5 Flash generated the misleading suggestions for SWE-bench, Terminal-Bench, MCP-Atlas, and Humanity’s Last Exam.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
