AI research agents don't fail at searching — they fail at asking you questions
A new benchmark from Tencent and Tsinghua finds that when questions are ambiguous, AI agents that search harder do worse than ones that simply ask the user — and even the best model stays below 50 percent.
If an AI assistant has ever confidently researched the wrong thing for you, a new study explains why — and it’s probably not what you’d guess. Researchers at Tencent Hunyuan and Tsinghua University built DiscoBench, a benchmark that tests how AI search agents handle a very human problem: questions that are vague, incomplete, or contain a mistake. Their finding: agents rarely fail because they can’t search. They fail because they won’t stop and ask you what you actually meant.
The benchmark contains 211 multi-step research tasks with 463 deliberately ambiguous points — a name that matches several people, an unclear version or time period, a built-in factual error. At each step, the agent can keep searching, ask the user a clarifying question, or answer. Eleven recent models were tested, and none reached 50 percent: the best, ByteDance’s Doubao Seed 2.0 Pro, hit 43.1 percent end-to-end accuracy, with Gemini 3.1 Pro at 40.8 and Claude Opus 4.7 at 39.8 percent. Strip the ambiguity out of the same questions, and accuracy jumps by up to 40 points. The problem isn’t the research — it’s the misunderstanding.
What’s behind this? The behavior analysis is the interesting part. Agents that searched and then asked a follow-up question succeeded 93.4 percent of the time at ambiguous steps. Agents that guessed without asking: 56.5 percent. And agents that searched over and over but never asked did worst of all, at 51.9 percent — worse than just guessing. The researchers’ reading: the repeated searching shows the model sensed something was off but never turned that doubt into a question. In a multi-step research chain, one wrong guess early on quietly poisons everything after it. Even telling models explicitly to watch for ambiguity only helped a little — they got better at spotting unclear questions, not at resolving them. One caveat: the tasks are mostly in Chinese, reflecting Chinese-language web search, so exact numbers may shift in other languages.
What this means for you: The practical lesson is refreshingly simple: the biggest lever for better AI research results is you being specific. Which person, which version, which time period, judged by what standard — spelling that out up front can be worth more than any model upgrade. And when an AI presents deep-research findings with total confidence, remember that it almost never asked itself whether it understood the assignment. If a result smells wrong, check the first step of its reasoning, not the last — that’s where wrong guesses live. Watch for assistants that ask clarifying questions; on this evidence, it’s a feature worth actively wanting.
Sources
Faster homework, worse exams: a 26,000-student study puts numbers on AI's hidden learning cost
A 30-month study of Chinese secondary students finds AI users finished homework faster and scored higher — but dropped up to 24 percent on exams, with the full damage taking two years to surface.
