PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.14412v1cs.AIcs.CLcs.IR

TL;DR

Reasoning-intensive questions often combine heterogeneous clues that must be found across documents, making the first retrieval move consequential. Question’s Gambit prepares a clue-aware opening context before iterative search begins. On BrowseComp-Plus, it raises GPT-5.5 answer accuracy from 83.1% to 90.5% over Pi-Serini, with reported transfer to MultiHop-RAG.

  • Problem

    Reasoning-intensive questions combine heterogeneous clues that jointly constrain answers, requiring broad clue coverage and cross-document verification.

    Source: §1 · p. 1

  • Method

    Question’s Gambit decomposes a question into clues, retrieves and expands them, then consolidates and reranks candidates into an opening context before the agent’s iterative search.

    Source: §1 · p. 2

  • Results

    90.5% GPT-5.5 answer accuracy versus 83.1% with Pi-Serini on BrowseComp-Plus; the paper also reports robust and stable transfer to MultiHop-RAG.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Trajectory analyses show that stronger openings help relevant evidence persist through subsequent agent selection and reasoning stages.

    Source: §6 · p. 8

    Evaluation focuses on fixed-corpus deep research, with generalization to other corpora and live web search still to be studied.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Deep research agents answer complex questions through iterative loops of searching, reading, and reasoning. Recent work on reasoning-intensive benchmarks such as BrowseComp-Plus shows that well-configured lexical retrieval can surface high-quality evidence, yet agents may still fail to connect documents carrying evidence to the gold documents. We identify a deep research agent's first retrieval move as an important design decision for this setting. We introduce Question's Gambit, a first-move retrieval module that decomposes the question into a set of clues, reformulates them into complementary searches, consolidates the retrieved results, and reranks the candidate pool before the agent begins its iterative search-and-reasoning process. This produces an opening context designed to support both clue aggregation and final-answer verification. We further evaluate on MultiHop-RAG to test whether these benefits transfer beyond BrowseComp-Plus to a more conventional multi-hop question structure. Experiments on BrowseComp-Plus show that Question's Gambit improves retrieval recall and downstream agent accuracy over strong baselines, improving answer accuracy from 83.1% to 90.5% with gpt-5.5 over Pi-Serini, the strongest reported agentic baseline. Our results confirm that effective agentic deep research depends not only on the tools available inside the loop, but also on the quality of the first move. We published our implementation publicly at https://github.com/radinhamidi/Question-s-Gambit.

1 Introduction

Question’s Gambit addresses cold starts in reasoning-intensive search, where heterogeneous clues require broad evidence coverage and cross-document verification. It prepares a clue-aware opening context before iterative search, improving BrowseComp-Plus accuracy and transferring to MultiHop-RAG.

1 Introduction

  • Reasoning-intensive questions combine heterogeneous clues that jointly constrain answers without forming a linear reasoning path.Answering requires broad clue coverage, cross-document evidence connection, and verification that candidate answers satisfy all constraints.Source: §1 · p. 1
  • Weak initial retrieval can create a cold start, sending agents toward irrelevant paths and making recovery more costly as noisy context accumulates.Prior analyses describe missed answer-relevant documents, repeated queries, cluttered contexts, and overly broad searches despite evidence being retrievable.Source: §1 · pp. 1–2
  • Question’s Gambit decomposes questions into clues, retrieves and expands each clue, then merges and reranks candidates into a compact opening context.It requires no agent or retriever retraining and leaves the subsequent search loop unchanged.Source: §1 · p. 2
  • Unlike in-loop planning or one-shot decompose-and-retrieve pipelines, Question’s Gambit exposes question structure before the loop while preserving iterative search.The first-move intervention is a one-time cost and changes only the agent’s first observation.Source: §1 · p. 2
  • GPT-5.5 answer accuracy rises from 83.1% to 90.5% over Pi-Serini, with higher retrieval recall and comparable search-call counts.The paper also reports consistent accuracy improvements for DeepSeek-v4-pro and GPT-5.4-mini.Source: §1 · p. 2

2 Related Work

Prior agentic-search work often adds retrieval tools within the loop, while Question’s Gambit prepares a clue-aware candidate pool before iterative search. It also changes the reranker’s candidate pool rather than proposing a new reranker.

2 Related Work

  • Prior agentic workflows improve search by adding stronger retrievers, rerankers, or tools inside the agent loop.Question’s Gambit instead prepares a clue-aware candidate pool before the iterative process unfolds.Source: §2 · pp. 2–3
  • Retrieve-and-rerank methods for reasoning-intensive queries are sensitive to candidate-pool construction, query formulation, and ranking configuration.The paper positions its approach as complementary to this work rather than as a new reranker.Source: §2 · p. 3
  • Question’s Gambit unions candidates retrieved from individual clues and reranks that clue-diverse pool against the full question.The paper reports that this produces a stronger opening context than question-level or reformulation-only retrieval.Source: §2 · p. 3

3 Methodology

Question’s Gambit replaces the agent’s cold-start opening with a one-time, clue-aware retrieval context. The agent then continues its usual iterative loop to resolve remaining evidence needs and produce an answer.

3.1 Problem Formulation

  • Multi-clue questions require identifying heterogeneous constraints, retrieving their supporting evidence, and integrating information across documents rather than finding one answer-bearing passage.The agent’s earliest retrievals shape its trajectory while much of the relevant evidence remains unseen.Source: §3.1 · pp. 3–4

3.2 Question’s Gambit

  • Question’s Gambit retrieves per-clue candidates, consolidates and ranks them against the full question, then supplies a compact context as the agent’s first observation.The module runs once without retraining and leaves the subsequent search loop unchanged.Source: §3.2 · p. 4

3.2.1 Question Decomposition

  • The agent decomposes the question into unordered clues, each expressing one answer constraint that can be searched independently.Using the agent’s own model keeps clue extraction aligned with its later interpretation of the question.Source: §3.2.1 · p. 4

3.2.2 Per-Clue Retrieval

  • Each clue’s raw retrieval results ground an expansion in corpus terminology; the clue, expansion, and question are combined for retrieval.The raw clue preserves the specific constraint, the expansion widens lexical coverage, and the full question keeps answer-bearing documents accessible.Source: §3.2.2 · pp. 4–5

3.2.3 Question-Conditioned Reranking

  • The per-clue lists are deduplicated into a pool, then each document is scored against the full question to select the top-k opening context.Candidates are scored using document text rather than short retrieval snippets.Source: §3.2.3 · p. 5

3.3 Iterative Deep Search

  • After receiving the opening context, the agent previews, reads, paginates, or retrieves documents to fill remaining gaps before synthesizing its answer.The loop ends when the agent emits an answer or exhausts its wall-clock budget.Source: §3.3 · p. 5

4 Experimental Setup

The evaluation uses BrowseComp-Plus as its primary benchmark and MultiHop-RAG as a control for transfer to conventional multi-hop questions. Comparisons include agentic baselines, while implementation details hold key retrieval components fixed.

4.1 Datasets

  • BrowseComp-Plus contains 830 queries over 100,195 documents, while the MultiHop-RAG evaluation samples 200 queries across four categories.BrowseComp-Plus averages 6.1 evidence documents and 2.9 gold documents per query; MultiHop-RAG includes two-to-four-document questions and null queries.Source: §4.1 · p. 5

4.2 Baselines

  • The study compares against sparse and dense reference configurations, recent agentic systems, and PI-SERINI as its closest baseline.PI-SERINI and Question’s Gambit share the BM25 backend and subsequent search, result-reading, and document-reading loop.Source: §4.2 · pp. 5–6

4.3 Implementation Details

  • Question’s Gambit uses BM25 for first-stage retrieval, expands clues with MUGI, reranks the merged pool, and gives the agent five opening documents.The system is integrated as a single first-move tool; subsequent searches use plain BM25, and three agent models are evaluated under 900-second budgets.Source: §4.3 · p. 6

5 Results

Question’s Gambit improves BrowseComp-Plus accuracy and recall, and its gains transfer to MultiHop-RAG. Its tool-use, cost, and error analyses show where the opening context changes the agent’s work and where residual errors remain.

5.1 BrowseComp-Plus Results

  • Accuracy rises against matched PI-SERINI configurations from 71.4% to 76.9% for DeepSeekv4-pro, 68.1% to 79.0% for GPT-5.4-mini, and 83.1% to 90.5% for GPT-5.5.The configurations use the same BM25 retriever and downstream agent loop.Source: §5.1 · p. 7
  • For GPT-5.5, gold recall rises from 94.4 to 98.1 when surfaced, 72.9 to 85.9 when previewed, and 56.1 to 75.9 in behavior.Across all three models, evidence and gold recall improve at every stage.Source: §5.1 · p. 7, Table 2 · p. 7

5.2 Transfer to MultiHop-RAG

  • On MultiHop-RAG, accuracy increases from 76.0% to 77.0% for GPT-5.4-mini and from 81.0% to 81.5% for GPT-5.5.The gains are smaller than on BrowseComp-Plus, but performance improves for both models.Source: §5.2 · p. 7, Table 3 · p. 7

5.3 Tool-Use Analysis

  • On BrowseComp-Plus, search calls change from 14.2 to 16.6 for DeepSeek-v4-pro, 20.2 to 19.9 for GPT-5.4-mini, and 13.5 to 14.5 for GPT-5.5, while accuracy gains are 5.5–10.9 percentage points.Total calls increase by 2.3–5.3, while browse calls remain unchanged or decrease for every model.Source: §5.3 · p. 7, Table 4 · p. 7
  • On MultiHop-RAG, search calls fall from 8.34 to 8.15 for GPT-5.4-mini and 5.57 to 4.25 for GPT-5.5; GPT-5.5 total tool use falls from 10.23 to 9.34 while accuracy improves.GPT-5.4-mini total calls increase from 11.71 to 12.74, alongside the first-move action and additional document reading.Source: §5.3 · pp. 7–8, Table 5 · p. 8

5.4 Cost and Latency

  • The opening stage has a 28.5-second median latency, 29% of the 98-second median trajectory, and an estimated $0.05 cost, less than 10% of the approximately $0.55 per-question total.Across a 50-question matched-budget development comparison, median latency increases from 59 to 88 seconds, with 35 seconds spent on the opening; subsequent-loop median API cost remains $0.23 in both systems.Source: §5.4 · p. 8, Table 6 · p. 8

5.5 Error Analysis

  • Among 79 incorrect GPT-5.5 answers, 53 (67.1%) fail after gold evidence is surfaced but before it is opened or cited, while only 3 (3.8%) lack surfaced gold.Another 23 errors (29.1%) occur after the agent opens or cites a gold document, pointing to verification or evidence-synthesis failures.Source: §5.5 · p. 8, Table 7 · p. 8

6 Concluding Remarks

Question’s Gambit is a clue-aware first-move retrieval module for deep research agents. It improves results across three LLM agents on BrowseComp-Plus and transfers to MultiHop-RAG, while trajectory analyses show relevant evidence persisting through later stages.

6 Concluding Remarks

  • Question’s Gambit is a clue-aware first-move retrieval module for deep research agents.The paper presents the first move as a design direction and motivates adaptive openings tailored to question complexity and information needs.Source: §6 · p. 8
  • The module delivers gains across three LLM agents on BrowseComp-Plus and transfers reliably to MultiHop-RAG.Trajectory analyses show that stronger openings help relevant evidence persist through subsequent selection and reasoning stages.Source: §6 · p. 8
  • The authors propose adaptive openings tailored to each question’s complexity and information needs as a future direction.Source: §6 · p. 8

Limitations

The evaluation and method leave several boundaries open, including corpus and model coverage, retriever choice, and possible anchoring on misleading opening evidence.

Limitations

  • Generalization beyond fixed-corpus deep research, including to other corpora and live web search, remains unstudied.The evaluation uses the full BrowseComp-Plus benchmark and a 200-question MultiHop-RAG control set.Source: Limitations · p. 9
  • Evaluation covers three agent models on BrowseComp-Plus and two on MultiHop-RAG, leaving broader model coverage for future work.Source: Limitations · p. 9
  • Question’s Gambit currently uses BM25 candidate generation, and its behavior with dense, hybrid, or learned retrievers may differ.Source: Limitations · p. 9
  • Matched comparisons isolate the main first-move component but do not fully characterize anchoring on misleading opening evidence.Future work could adaptively calibrate or skip the first move based on question difficulty, ambiguity, and retrieval confidence.Source: Limitations · p. 9

A Question Structure Across Benchmarks

The benchmarks both require evidence from multiple documents but differ in how explicitly they identify the relevant evidence and entities.

A Question Structure Across Benchmarks

B Statistical Significance Analysis

Question’s Gambit significantly improves accuracy across all three agents, while repeated development-set runs provide complementary evidence of stability.

B Statistical Significance Analysis

C Prompts

The appendix supplies prompts for the deep research agent, covering the first retrieval move, clue decomposition, iteration, stopping, and answer format.

C Prompts

  • The default first move for multi-clue questions decomposes clues and returns a top-5 list with title and excerpt previews.The result is keyed to the original question; a distinctive quoted phrase may instead warrant a quoted search.Source: C Prompts · p. 11
  • When plausible candidates appear in the top five, inspect deeper ranks or read the strongest candidate before rewriting the query.Source: C Prompts · p. 12
  • Refinements should use a specific new clue from reading a document, not merely a different phrasing.Source: C Prompts · p. 12
  • The prompt says to stop when a document explicitly contains the answer, evidence is sufficient to conclude it is absent, or a submit-now steer arrives.Source: C Prompts · p. 12
  • The final response must include an explanation with supporting document citations, a concise exact answer, and a confidence score.Source: C Prompts · p. 12
  • After reading first-move hits, use tactical searches for newly discovered entities, dates, or specific phrases, or fall back when results are weak.Source: C Prompts · p. 12
  • First-move previews are not verified content; the agent must open at least one document before answering.The prompt treats read_document as the only ground-truth tool.Source: C Prompts · p. 12
  • Clues should be atomic, independently verifiable factual constraints, phrased as short BM25-searchable claims.The prompt calls for 5–20 words and concrete dates, ranges, names, numbers, or quoted phrases where relevant.Source: C Prompts · p. 12
Loading 2609.14412v1…