Skip to main content
Back

Alibaba Researchers Catch AI Agents Googling Answers During Benchmark Tests

Researchers from Alibaba and NTU show that web-searching AI agents can accidentally look up their own test answers, inflating benchmark scores by up to 4%

Paper, Full Access
  • Researchers from Alibaba and NTU published a paper formally studying Search-Time Contamination (STC) in deep research agents.
  • STC occurs when agents browsing the web during evaluation accidentally retrieve benchmark questions or ground-truth answers from public sources.
  • Three contamination tiers are defined: Benchmark Metadata Leakage, Question-Context Leakage, and Explicit Answer Leakage, with increasing severity.
  • Across six medical benchmarks (MMLU, MedQA, MedMCQA, MedXpertQA, HLE, Medbullets5op), STC inflates scores by up to 4%.
  • MedMCQA is the most exposed: nearly one quarter of its questions can have answers retrieved directly from the web.
  • The paper recommends isolated sandboxes, transparent search logs, and controlled dataset access as mitigation strategies.

Benchmarks are the scorecards of AI progress. When a new deep research agent posts a record accuracy on MMLU or Humanity's Last Exam, the field takes notice. But a new paper from researchers at Alibaba's Tongyi Lab and Nanyang Technological University raises an uncomfortable question: what if the agent is quietly looking up the answers while being tested?

The study, led by Yongjie Wang and colleagues, introduces a formal framework for what they call Search-Time Contamination (STC) -- a phenomenon where deep research agents, which browse the web during inference, inadvertently retrieve benchmark questions, context, or even ground-truth answers from publicly accessible sources. The result: inflated scores that overstate how well these systems actually reason.

A New Kind of Cheating Sheet

Traditional data contamination happens at training time -- a model memorizes test answers because they appeared in its pretraining corpus. STC is different and trickier. Public benchmarks become fragile for deep research agents that actively search the web during inference, because such agents may retrieve public benchmark metadata, question context, or even ground-truth answers via web search -- bypassing intended reasoning and inflating measured performance.

This is not a hypothetical edge case. Because many public medical benchmarks are widely reproduced across repositories, examination platforms, question banks, educational websites, and forums, they provide a high-stakes and highly exposed setting for studying STC. All evaluated benchmark datasets exhibit varying degrees of STC, and the issue is particularly severe for early benchmark MedMCQA, where nearly one quarter of the questions can have their answers retrieved from the web.