What they did
The authors examined tool-use trajectories to compare successful and unsuccessful deep-research searches. They found that stronger trajectories turn vague possibilities into concrete candidates and change direction when the available evidence is weak or incomplete.
Based on this analysis, they developed HypoSearch. It generates hypotheses as nonbinding search hints, investigates them through bounded independent branches, and compares the evidence collected by each branch before committing to a path. The system was evaluated on four deep-research benchmarks with three backbone models, alongside single-trajectory and parallel-search baselines.
Key findings
- HypoSearch consistently outperformed both single-trajectory search and standard parallel baselines across all four benchmarks and three models tested.
- On BC-small, it raised Qwen3.5-122B's score from 46.7 to 60.0, a gain of 13.3 points.
- It used fewer tool calls than running five fully independent search trajectories, suggesting that structured branching can be more efficient than brute-force parallelism.
- In a pilot supervised fine-tuning study, behavioral signals identified by the trajectory analysis helped select a compact training set and reduced the performance degradation associated with unfiltered trajectory data.
Why it matters
The results suggest that deep-research agents benefit not merely from more searches, but from explicit mechanisms for preserving and comparing plausible alternatives. Hypothesis-guided branching offers a practical form of search orchestration that may reduce confirmation-like feedback loops without requiring many complete, independent runs.
Caveats
The reported evidence covers four benchmarks and three models, but broader robustness remains uncertain, particularly for tasks where useful hypotheses are difficult to formulate or evidence cannot be cleanly compared. The fine-tuning result is described as a pilot, and further work is needed to establish its scale, generality, and full compute and latency trade-offs.