Deep-research agents often fail because they commit too early to one search direction, reinforcing misleading paths. The paper analyzes trajectory-level behavior and finds that successful agents ground vague exploration in concrete candidates and shift directions when a path is weak.
The authors propose HypoSearch, which generates lightweight hypotheses from a query or intermediate state, explores each in bounded independent branches, and then compares evidence across branches before committing. On four deep-research benchmarks and three backbone models, HypoSearch consistently outperforms single-trajectory search and standard parallel baselines—for instance, improving Qwen3.5-122B from 46.7 to 60.0 on BC-small while using fewer tool calls than five independent trajectories.
This is interesting because it addresses a fundamental bottleneck in sequential search agents: the moment when multiple directions are plausible. The method is complementary to existing test-time scaling approaches and offers a practical way to improve performance without excessive cost.