An agent's short search query can hide the intent already expressed in its preceding reasoning. Reusing that reasoning gives the retriever task context without another query-expansion model call; adding the entire interaction history can instead introduce noise.
Key Findings
- DR-Synth converts question-answering examples into supervision for intermediate searches. The reported training run uses 5,238 examples to adapt Qwen3-Embedding-4B to reasoning and query inputs.
- In the paper's search-only Tongyi-DeepResearch evaluation on BrowseComp-Plus, accuracy reaches 66.27%, versus 50.72% for Qwen3-Embedding-8B. Improving the retriever changes the agent's final answers without enlarging the agent itself.
- Adding more previous reasoning turns does not improve over the current turn alone. This suggests that selective retention of relevant clues matters more than supplying the retriever with the longest possible history.
Who Should Read It
Great fit if you build research agents that expose reasoning before search and can adapt the retrieval component. Look elsewhere if you need a drop-in improvement for a query-only interface: the method needs reasoning access, and its benchmark gains do not establish the same benefit for ordinary single-turn search.