TL;DR
Deep research agents often fail on complex questions because an initial poorly formulated search sends them down a rabbit hole of wasted tool calls. Question's Gambit introduces a "warm start" strategy that decomposes questions into atomic clues, runs parallel corpus-grounded searches, and reranks the results into a high-quality opening context. This simple first-move intervention boosts downstream agent accuracy from 83.1% to 90.5% on reasoning-intensive benchmarks.
What Problem Does This Paper Address?
Large Language Model (LLM) agents are increasingly used for "deep research" tasks, where they iteratively search, read, and reason over large document collections. However, complex real-world questions are rarely straightforward; they are often "multi-clue," meaning they pack multiple heterogeneous constraints (dates, locations, partial descriptions) that jointly point to an answer.
When an agent starts with a single, overloaded query, these constraints compete, leading to a "cold start." The agent retrieves noisy or incomplete documents, gets sidetracked by incomplete hypotheses, and wastes its limited interaction budget trying to recover. The authors identify the agent's very first retrieval move as a critical, yet overlooked, bottleneck in agentic search.
Methods
The authors propose Question's Gambit, a plug-and-play, first-move retrieval module that runs entirely before the agent's standard iterative loop begins. It requires no retraining of the agent or the underlying retriever and operates in three distinct stages:
- Question Decomposition: The agent's LLM breaks the complex question down into a set of atomic, independent factual constraints (clues).
- Per-Clue Retrieval & Expansion: To avoid short, underspecified queries, each raw clue is sent to the retriever to gather a small feedback set. An expansion model uses this feedback to rewrite the clue using actual terminology found in the corpus. The clue, its expansion, and the original question are concatenated to retrieve a deep candidate pool (top 1,000 documents per clue).
- Question-Conditioned Reranking: The candidate lists from all clues are unioned and deduplicated. A pointwise cross-encoder reranker then scores each document against the full original question. The top-$k$ (e.g., $k=5$) highest-scoring documents are handed to the agent as its initial observation.
Datasets
- BrowseComp-Plus: The primary benchmark used for evaluation. It contains 830 complex, reasoning-intensive queries over a corpus of ~100,000 documents, featuring "inverted" questions where the target entity is hidden behind multiple clues.
- MultiHop-RAG: Used as a control benchmark to test robustness. It consists of 2,556 queries over 609 news articles, representing a more conventional multi-hop structure where questions follow a clear chain of explicit claims sharing an entity.
Key Insights and Ideas
- The First Move is a Trajectory Shaper: In iterative agentic workflows, early observations heavily condition all subsequent actions. A weak opening context creates a compounding error loop where the agent spends its budget fixing its trajectory rather than verifying the answer.
- Divide and Conquer for Retrieval Constraints: Forcing an LLM or a lexical retriever to handle five constraints in one query causes them to compete and dilute relevance. Searching them independently ensures broad clue coverage.
- Corpus-Grounded Expansion Beats Parametric Guessing: Instead of letting an LLM blindly expand a query using its internal weights, using a quick initial retrieval to ground the expansion in the corpus's actual vocabulary prevents hallucinations and aligns lexical matching.
- Union Locally, Judge Globally: While documents are surfaced based on individual clues, they must be reranked against the entire question to ensure the final opening context is cohesive.
Key Results
- Substantial Accuracy Gains: When integrated into Pi-Serini (a state-of-the-art deep research framework), Question's Gambit boosted GPT-5.5 answer accuracy from 83.1% to 90.5% on BrowseComp-Plus, outperforming all dense and sparse retrieval baselines.
- Better Confidence Calibration: It drastically reduced calibration error (e.g., from 15.7 down to 7.58 for GPT-5.5), meaning the agents became much better at knowing when they were actually correct.
- Robust Transfer: On the MultiHop-RAG control benchmark, performance remained stable or slightly improved, proving that the method doesn't degrade performance on simpler, linear multi-hop tasks.
- Low Overhead: The entire process adds a median latency of 28.5 seconds and an estimated API cost of just $0.05 per question—less than 10% of the total deep research trajectory cost.
Limitations and Future Work
- Fixed-Corpus Constraint: The evaluation was restricted to fixed document collections; how this approach scales to open, live web searches remains to be studied.
- Retrieval Backend: The current implementation relies strictly on BM25 for initial candidate generation; its behavior with dense, hybrid, or learned retrievers has not yet been explored.
- Vulnerability to Misleading Evidence: The system does not explicitly handle scenarios where the initial opening context contains highly convincing but misleading "honeypot" documents.
- Future Directions: The authors suggest developing adaptive strategies that can dynamically decide whether to trigger or skip the first-move module based on estimated question difficulty, ambiguity, and retrieval confidence.
Summary updated