Please note: This PhD seminar will take place in DC 3301.
Amin Bigdeli, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Charles Clarke, Ebrahim Bagheri
Deep search agents have emerged as a promising approach to complex information-seeking tasks, extending traditional retrieval-augmented generation through iterative cycles of searching, reading, and reasoning. By progressively gathering and combining evidence from multiple sources, these agents can answer questions that require connecting information scattered across different documents. However, their effectiveness is often limited by the cold-start problem. At the beginning of the search process, agents have no retrieved evidence to guide their initial queries and must rely on the question alone. As a result, they may overlook important clues or retrieve insufficient or irrelevant information. Since subsequent searches and reasoning steps depend on previously collected evidence, a poor initial search can lead to inefficient exploration and ultimately affect answer quality.
This work addresses the cold-start problem through Question’s Gambit, a first-move retrieval framework that provides deep search agents with a stronger starting point before iterative search begins. Question’s Gambit decomposes complex questions into individual clues, expands each clue using corpus-grounded feedback, and retrieves supporting documents independently. The retrieved candidates are then consolidated and reranked against the original question to construct an informative initial context. This approach requires no retraining and leaves the subsequent agent workflow unchanged. Experiments on reasoning-intensive benchmarks including BrowseComp-Plus demonstrate improvements in evidence recall and answer accuracy, highlighting the importance of effective initialization in agentic search.