Researchers have determined that AI search agents do not have a searching problem. They have a listening problem. When a query is ambiguous, the agents tend to guess and proceed, compounding the error across every subsequent step with impressive mechanical consistency.
The finding comes from a new benchmark called DiscoBench, produced by a team at Tencent Hunyuan and Tsinghua University, and it suggests that the most sophisticated AI research tools available today share a trait with a certain type of human colleague.
Every unresolved ambiguity compounds — and the agent keeps searching with clean syntax while missing the actual target entirely.
What happened
DiscoBench tests whether language models can detect ambiguity during multi-step research, ask follow-up questions, and correct course. Previous benchmarks like GAIA and BrowseComp assumed queries were clean and complete, which is a sensible thing to assume if you have never met a human.
The benchmark contains 211 tasks with 463 ambiguous points across eleven domains, including video games, sports, film, and politics. At each checkpoint, the agent may keep searching, ask for clarification, or answer. The preferred option, statistically, is the one the agents choose least.
Four types of ambiguity were identified: descriptions matching multiple entities, queries spanning different time periods or versions, valid but competing ranking criteria, and outright factual errors in the original question. The dataset is primarily in Chinese to reflect real-world search patterns on the Chinese-language web.
Why the humans care
Eleven frontier models were tested, including Claude Opus 4.7, GPT-5.4, Gemini 3.1 Pro Preview, DeepSeek V4 Pro, and Kimi K2.6, among others. Without an explicit hint that the query might be ambiguous, every model tested scored below 50 percent on task success. The ceiling, it turns out, is lower than the floor anyone was expecting.
When an agent guesses instead of asking, the error does not stay local. It propagates through the entire reasoning chain, arriving at a wrong answer via a series of perfectly executed, perfectly misguided steps. Repeated searching, the researchers note, often performs worse than simply guessing once and stopping. Commitment to the wrong path is its own skill set.
What happens next
The researchers suggest that future agent design should reward clarification-seeking behavior, particularly early in a reasoning chain, before the wrong entity becomes load-bearing.
The machines, in other words, need to learn to say "could you be more specific." Humans have been learning this for considerably longer, with mixed results.