Question's Gambit: The First Move Matters in Agentic Deep Search
Radin Hamidi Rad, Amin Bigdeli, Negar Arabzadeh, Sajad Ebrahimi, Charles L. A. Clarke, Benjamin C. M. Fung, Ebrahim Bagheri
cs.AI, cs.CL, cs.IR
2026-09-13
Question's Gambit splits a query into clues, retrieves each, then reranks them as the first observation. On Pi-Serini, GPT-5.5 rises from 83.1% to 90.5% on BrowseComp-Plus.
Deep-research agents interleave search, reading, and reasoning in a ReAct loop. A single BrowseComp-Plus question can take dozens of tool calls. These questions do not follow the linear HotpotQA pattern of "look up A, then B." They hide the answer behind a bundle of heterogeneous clues: dates, places, job titles, coauthors, episode counts. Supporting evidence often lives in different documents.
Retrieval coverage is already high. Pi-Serini with BM25 surfaces gold documents at 94.4% recall for GPT-5.5. The failure mode is the cold start. From empty history the agent packs the whole question into one overloaded query, clues compete for terms, the first hits cover only part of the constraint set, and later turns chase an incomplete hypothesis.
Question's Gambit takes over only the first retrieval. The rest of the loop stays as-is. Neither the agent nor the retriever is retrained.
Clues buy coverage. The reranker buys question-level order. Expansion model and reranker are locked across DeepSeek-v4-pro, GPT-5.4-mini, and GPT-5.5, so a stronger agent cannot be credited for a better expander. After the opening, Pi-Serini continues with Preview, Read, Paginate, and Retrieve on plain BM25, wall-clock budget 900 seconds.
BrowseComp-Plus has 830 queries over 100,195 documents, 6.1 evidence and 2.9 gold documents per query on average.
| Setup | Acc. | Calib. | Gold recall |
| Pi-Serini + DeepSeek-v4-pro | 71.4% | 7.0 | 92.5 |
| Gambit | 76.9% | 3.99 | 95.7 |
| Pi-Serini + GPT-5.4-mini | 68.1% | 13.7 | 94.1 |
| Gambit | 79.0% | 7.90 | 94.6 |
| Pi-Serini + GPT-5.5 | 83.1% | 15.7 | 94.4 |
| Gambit | 90.5% | 7.58 | 98.1 |
GPT-5.5 gains 7.4 points (p = 9.7×10⁻⁶). The mini model gains 10.9, the largest jump. Other published systems sit lower, gpt-5 with Qwen3-Embedding-8B at 73.0% and Tongyi-DR with AgentIR-4B at 68.1%, but those are not matched-agent comparisons.
Trajectory recall moves more than first-stage recall. For GPT-5.5 gold documents, surfaced recall goes from 94.4 to 98.1, previewed from 72.9 to 85.9, and behavior recall (opened or cited) from 56.1 to 75.9. The opening set helps evidence survive later selection.
Search volume barely changes. Total tool calls rise by 2.3 to 5.3, mostly the one first-move call plus extra reads. GPT-5.5 search calls go from 13.5 to 14.5. Median first-move cost is 28.5 seconds and about $0.05, against a 98-second, $0.55 median trajectory.
On a 200-query MultiHop-RAG slice, accuracy moves from 76.0% to 77.0% for GPT-5.4-mini and from 81.0% to 81.5% for GPT-5.5. No regression, almost no gain.
Of 79 GPT-5.5 errors, gold never surfaces in 3 cases (3.8%). 67.1% fail after gold is retrieved but before it is opened. 29.1% open gold and still answer wrong. Coverage is not the remaining bottleneck.
Most agentic-search papers add tools inside the loop: denser retrievers, memory modules, reasoning-aware indexes. This one changes only the first observation and still moves GPT-5.5 from 83.1% to 90.5% on BrowseComp-Plus, with no retraining.
The fit is multi-clue inverted identification. Conventional multi-hop barely budges, so this is not a universal preprocessor. Residual errors also name the next bottleneck: preview ranking, evidence selection, and synthesis, not another retrieval pass.
The implementation is public. Extra spend is one decomposition call, about ten parallel expansions, and a Cohere rerank, cheap relative to the agent loop.
The authors flag the obvious constraints. Evaluation is on fixed corpora, not live web search. Three agent models on BrowseComp-Plus, two on MultiHop-RAG. Candidate generation is BM25 only; dense or hybrid first-stage retrieval is untested. Anchoring on a misleading opening set is not fully measured.
A few gaps they leave open. There is no ablation that removes expansion or reranking, so the 9.6 clues, k=5, and Cohere Rerank4 Pro arrive as a bundled bet. The MultiHop-RAG deltas of 0.5 to 1 point support "does not hurt" more than "transfers robustly." The opening context is still a title plus a roughly 300-character excerpt; the agent must still call readdocument. The module replaces the first candidate list. It does not skip reading.