Source-linked AI summary
Explore Before Committing: Hypothesis-Guided Search for Deep Research Agents
Ruochen Zhou, Zhengyu Chen, Luan Zhang, Siyang Gao, Yee Whye Teh, Shiqi Chen
TL;DR
Deep-research agents can commit prematurely when several search directions remain plausible, leaving later retrieval and synthesis within a potentially misleading path. HYPOSEARCH addresses this by generating lightweight hypotheses, exploring bounded independent branches, and comparing their evidence before commitment. Across four benchmarks and three backbones, it consistently outperforms the evaluated baselines, including improving Qwen3.5-122B from 46.7 to 60.0 on BC-small while using fewer tool calls than full independent sampling.
Problem
Single-trajectory agents must choose one direction before collecting comparative evidence when multiple candidates or perspectives are plausible, making early search allocation a central challenge.
Method
HYPOSEARCH generates lightweight hypotheses as soft search hints, explores them through bounded independent branches, and aggregates evidence to decide whether to answer or refine the hypotheses.
Results
HYPOSEARCH achieves the strongest overall performance across four deep-research benchmarks and three model backbones, improving Qwen3.5-122B from 46.7 to 60.0 on BC-small.
Takeaways & Limitations
Search scaling is useful for keeping several concrete paths active around search uncertainty, and the identified behaviors can also guide compact supervised-fine-tuning data selection.
Takeaways & Limitations
The implementation uses fixed branching and branch budgets, while its prompt-based detector may misclassify ambiguous states and fail to match exploration effort to uncertainty.
Abstract
from arXiv · showhide
Deep-research agents answer complex questions by interacting with search and browsing tools, yet they often search along a single evolving trajectory. Our trajectory-level analysis reveals a common failure mode in which the agent may encounter an early search state with several plausible directions, but follow one direction before collecting enough comparative evidence. Once this happens, subsequent tool calls tend to reinforce the same path, increasing the chance of failure when the initial direction is misleading. We further find that successful trajectories reduce this risk through two behaviors: grounding vague exploration in concrete candidates and shifting directions when the current path is weak or incomplete. Based on these findings, we propose HypoSearch, which generates lightweight hypotheses as soft search hints, explores them through bounded independent branches, and compares branch-level evidence before commitment. Across four deep-research benchmarks and three backbone models, HypoSearch consistently outperforms single-trajectory search and standard parallel baselines, improving Qwen3.5-122B from 46.7 to 60.0 on BC-small while using fewer tool calls than five independent trajectories. A pilot supervised fine-tuning study further shows that these behavioral signals can curate compact training trajectories and reduce degradation from unfiltered data.
1 Introduction
Deep-research agents often commit to one search direction even when multiple perspectives remain plausible, motivating hypothesis-guided branching that compares evidence before commitment.
- Problem: Single-trajectory search must choose one direction when multiple candidates or evidence perspectives are plausible, creating an early allocation bottleneck.Later retrieval and synthesis remain within the channel established by that initial choice.
- HYPOSEARCH: HYPOSEARCH generates lightweight hypotheses, explores them through bounded independent branches, and aggregates comparative evidence before answering or refining hypotheses.The framework intervenes before commitment rather than waiting for a complete trajectory.
- Search-state analysis: Divergent search states require collecting and comparing evidence across alternatives, unlike direct states where the next lookup step is clear.The trajectory analysis distinguishes direct states from divergent states based on whether the next search step is determined.
- Behavioral signals: Successful trajectories ground vague exploration in concrete candidates and shift directions when the current path is weak or incomplete.These behaviors motivate making plausible directions explicit before aggregation.
- Evaluation: HYPOSEARCH achieves the strongest overall performance in evaluated settings across four deep-research benchmarks and three model backbones.On BC-small, Qwen3.5-122B improves from 46.7 to 60.0, while the method uses fewer tool calls than full independent parallel sampling.
2 Analysis of Divergent Search
Single-trajectory search is especially fragile when multiple directions remain plausible, because early choices can make later evidence increasingly narrow and misleading. Successful divergent exploration instead combines concrete candidate grounding with movement across search directions, motivating pre-commit computation allocation.
- 300 problems per benchmark were analyzed with gpt-4.1 and human experts to label failure categories and extract behavioral signals.
- Exploration Failure accounts for approximately 78% of failed BrowseComp trajectories, while exploration failures account for around 6% of failures on GAIA.Failures are labeled by the earliest stage where the error becomes unrecoverable.
- BrowseComp query similarity drops from 0.53 in Round 1 to 0.16 in Round 3, whereas FRAMES and GAIA remain above 0.54 in Round 3.The comparison uses average pairwise Jaccard similarity across five independently sampled trajectories.
- Higher Candidate-targeted Investment Rate means more candidate-level investigation, while higher Search-direction Switch Rate indicates more frequent movement across directions.CIR measures concrete candidate grounding; SSR measures switches between consecutive direction segments.
- The highest success rate occurs in the high-CIR and high-SSR region, while high CIR with low SSR reaches 25% and low CIR with high SSR reaches 26%.Successful divergent exploration is associated with the combination of candidate grounding and controlled direction movement rather than either behavior alone.
- HYPOSEARCH allocates computation before commitment by maintaining multiple concrete directions and preserving their evidence for later comparison.This principle complements parallel aggregation, reactive recovery, and verifier-based selection by targeting the underdetermined search state before one path dominates.
3 Method
HYPOSEARCH intervenes at local search states that may require divergent exploration, turning plausible directions into lightweight hypotheses and bounded independent branches. It then compares branch evidence to answer when sufficient or refines the hypotheses for another search round.
- HYPOSEARCH operates at local search states initialized by the original query and updated with observations from previous tool calls.
- The method first determines whether divergent exploration is needed, then applies hypothesis generation, parallel search, and comparative aggregation.Direct states continue with single-path search, while divergent states are expanded into multiple hypotheses.
- 3.1 Hypothesis Generation: At a divergent state, HYPOSEARCH generates K natural-language hypotheses describing plausible candidate entities, explanatory paths, or evidence perspectives.The state is represented as s_t = (x, m_t), where x is the original question and m_t is local context.
- Each hypothesis initializes a bounded independent branch as a soft search hint rather than a claim to be proven.Branches may revise their initial direction when new evidence indicates a better one and collect support, contradictions, constraints, and uncertainties.
- Branch isolation preserves alternatives by allowing different hypotheses to develop separate evidence states before comparison.Completed branches return structured summaries containing the hypothesis, key evidence, verified constraints, unresolved issues, and a branch-level conclusion.
- The aggregation agent compares constraint satisfaction, evidence reliability, conflicts, and remaining uncertainty before producing an answer or refined hypotheses.It returns a final answer with supporting evidence and reasons for rejecting weaker alternatives when the evidence is sufficient.
4 Experiments
The experiments evaluate HYPOSEARCH across four benchmarks and three backbones against single-trajectory and independent-trajectory baselines, finding stronger accuracy with lower search cost than full independent sampling. Performance improves as the hypothesis budget grows, especially when questions support divergent search directions.
- Experimental Setup: HYPOSEARCH is evaluated on four deep-research benchmarks across three backbone models using common tool and evaluation protocols.The benchmarks are BC-small, BrowseComp-zh, FutureX, and ResearchRubrics; the backbones are Kimi-K25, Qwen3.5-122B, and DeepSeek-V3.2.
- Experimental Setup: HYPOSEARCH generates up to five hypothesis-conditioned branches, while Majority Voting and Best-of-N each use five independent trajectories.Pass@1 denotes standard single-trajectory ReAct search.
- Main Results: HYPOSEARCH achieves the best results across the evaluated backbones and benchmarks.On BC-small, Qwen3.5-122B improves from 46.7 with Pass@1 to 60.0, while Kimi-K25 improves from 58.3 to 66.7.
- Tool-Call Efficiency Analysis: 132.5 versus around 205 tool calls per question: HYPOSEARCH reduces Kimi-K25’s cost relative to Majority Voting by 35.4%.The corresponding reduction on Qwen3.5-122B is about 29.1%, while HYPOSEARCH still uses more calls than Pass@1.
- Effect of the Number of Hypotheses: At K = 5, Kimi-K25 reaches 71.6, Qwen3.5-122B reaches 66.1, and DeepSeek-V3.2 reaches 70.9 on BrowseComp-zh.These values are higher than the corresponding Majority Voting and Best-of-N settings; smaller hypothesis sets are less stable.
5 A Pilot Study on Training Search Behavior
The pilot study tests whether candidate-grounded exploration and controlled direction movement can filter training trajectories. Filtering 80K trajectories to 10K preserves search performance while reducing degradation on non-search reasoning.
- Data Filtering: The study selects trajectories using candidate-grounded exploration measured by CIR and controlled direction movement measured by SSR.These behavioral metrics are intended to encourage concrete candidates while avoiding premature lock-in.
- Data Filtering: Q = 0.2 · SSR + 0.8 · CIR ranks trajectories for selecting the top 10K as Filtered-10K.CIR receives the larger weight because candidate formation is treated as the primary grounded-exploration signal.
- Experimental Setup: The pilot compares Full-80K and Filtered-10K SFT models based on 80K raw trajectories and evaluates them on BC-small and AIME25.Qwen3-32B is used as the base model, with identical training configurations across settings.
- Results: Filtered-10K matches or slightly exceeds Full-80K on BC-small while using only one eighth of the data.The result suggests that the selected subset preserves search patterns relevant to early direction-selection tasks.
- Results: 43.3 versus 26.7 on AIME25: Filtered-10K retains a higher score than Full-80K.The study interprets this contrast as evidence that filtered data retain more of the base model’s non-search reasoning ability.
6 Related Works
Related work equips agents to retrieve and use external evidence and allocates additional computation through multiple reasoning paths or structured exploration. These approaches nonetheless often leave deep-research search organized as a single evolving trajectory.
- Tool-Augmented Agents: WebGPT, ReAct, Self-RAG, Search-o1, and Search-R1 support web interaction, external actions, or adaptive retrieval during reasoning.These methods improve agents’ ability to acquire and use external evidence.
- Open Gap: Despite these advances, the search process is still often organized as a single evolving trajectory.This limitation connects tool-augmented retrieval work with the paper’s focus on search allocation.
- Test-Time Scaling: Self-consistency, Best-of-N, verifier-based methods, process supervision, and Tree of Thoughts allocate or structure additional test-time exploration.The cited approaches sample, generate, evaluate, or organize multiple reasoning paths and candidate solutions.
7 Conclusion
The paper argues that deep-research scaling should allocate computation around search uncertainty rather than only final answers. HYPOSEARCH operationalizes this through hypothesis-guided branching and evidence-level aggregation, with benchmark gains, lower tool use than independent sampling, and supportive training-filtering results.
- Conclusion: For queries with multiple plausible directions, keeping several concrete search paths active before commitment is useful.HYPOSEARCH implements this strategy through hypothesis-guided branching and evidence-level aggregation.
- Conclusion: HYPOSEARCH achieves consistent gains across benchmarks and backbones while using fewer tool calls than full independent sampling.The conclusion also reports that the pilot fine-tuning study supports using these search behaviors for trajectory selection.
8 Limitations
The implementation uses fixed branching rather than adapting hypotheses and branch budgets to each search state's uncertainty, limiting exploration granularity.
- The branching configuration is fixed, including the maximum number of hypotheses and branch budgets.The detector is prompt-based and may misclassify ambiguous states.
- Fixed allocation can produce redundant hypotheses in some divergent states and insufficiently fine-grained exploration in others.
- Future work could investigate confidence-aware branch detection and adaptive hypothesis and branch-budget allocation.
A Benchmark Details
The evaluation covers complementary deep-research settings, including multilingual browsing, future-event prediction, and long-form evidence synthesis, under a controlled protocol.
- BC-small preserves complex web-browsing difficulty while reducing large-scale evaluation cost.Questions require combining clues, identifying candidate entities, and verifying them against external evidence.
- The benchmark suite spans fact-seeking web search, Chinese-language browsing, future-event prediction, and long-form research synthesis.Table 3 describes these as complementary deep-research settings.
- BrowseComp-zh tests query formulation, source selection, and cross-source evidence integration in Chinese information environments.
- FutureX evaluates evidence comparison for future-event prediction under time-filtered search access.The February 2026 Week 3 split is used, with retrieval restricted before the evaluation cutoff.
- ResearchRubrics evaluates evidence collection, cross-source integration, and completeness in long-horizon research tasks.
- All methods use the same tools, search and browsing interfaces, judging protocol, judge model, and evaluation prompt where applicable.This controls differences in search strategy rather than evaluation conditions.
B Baseline Implementation Details
The baselines compare single-trajectory search, five independent trajectories, confidence-based selection, and hypothesis branching under shared tools and controlled budgets.
- All methods use the same ReAct-style agent and external tools; they differ only in trajectory generation, allocation, and aggregation.
- Pass@1 follows one evolving trajectory and represents the default linear search process.Its single path is vulnerable to early commitment when multiple directions are plausible.
- Majority Voting runs N = 5 independent Pass@1 trajectories and aggregates their final answers.
- Best-of-N runs N = 5 independent Pass@1 trajectories and selects the trajectory with the highest self-reported confidence.
- The benchmark suite covers fact-seeking, low-verifiability future-event prediction, and long-form research synthesis.
- MV@5 latency assumes ideal parallelism and equals the slowest completion time among five trajectories.
- Budgets grow with independent trajectories for the baselines and are distributed across bounded, potentially early-terminating branches for HYPOSEARCH.
C Latency Analysis
HypoSearch reduces latency and tool use by branching selectively, guiding independent searches with lightweight hypotheses, and comparing evidence across branches before final aggregation.
- Latency Analysis: HypoSearch achieves lower latency than MV@5 across all backbones and four benchmarks.MV@5 latency is the maximum completion time among five parallel trajectories.
- Latency Analysis: Structured parallel branches are more compute-efficient than independent parallel sampling when combined with the reported tool-call reduction.
- Ablation on Branch Detection: On BC-small with Kimi-K25, HYPOSEARCH achieves the best accuracy while using fewer tool calls and lower latency than Always Branch.
- Ablation on Branch Detection: Always Branch improves accuracy over Direct Search but substantially increases tool calls and latency.
- Ablation on Branch Detection: Divergence detection allocates branching to states where multi-direction exploration is more useful instead of expanding every query uniformly.
- Branch Detection: The detector chooses divergent states when multiple plausible directions require separate investigation and chooses direct states when the next step is clear or branching would duplicate effort.When uncertain, the prompt instructs the detector to choose direct.
- Hypothesis Generation: The hypothesis proposer generates diverse, searchable, concrete directions that can suggest a first query or lookup.
- Hypothesis Generation: Hypotheses are soft starting hints that independent agents may follow, revise, or abandon rather than final answers or hard constraints.