DiscoBench finds AI search agents fail by not asking, not by bad searching
A new benchmark from Tencent Hunyuan and Tsinghua, DiscoBench, tests 211 tasks with 463 ambiguous points and finds the best model, Doubao Seed 2.0 Pro, tops out at 43.1% end-to-end — failing not at retrieval but at asking the user to disambiguate. The behavioral split is the buildable signal: agents that search then ask a follow-up succeed 93.4% of the time, while searching harder and guessing drops to 51.9%, so the fix is wiring detected ambiguity into a clarifying question rather than more tool calls.
Source: the-decoder.com ↗
Models that search first and then ask a follow-up average a 93.4 percent success rate. Guessing without asking drops to 56.5 percent.
DiscoBench researchers (Tencent Hunyuan & Tsinghua University)
Why this matters
- → AI search agents fail at clarification, not retrieval—fixing this gap is simpler than improving search.
- → SearchThenAsk strategy yields 93.4% success vs. 51.9% for repeated searching without asking.
- → Ambiguity detection and question quality are decoupled skills; agents need both to work.
Ask, don't search harder