DeepWideSearch: Integrated Deep & Wide Retrieval

• DeepWideSearch is an agentic information seeking framework that integrates multi-hop reasoning to identify core entities with wide-scale collection of structured attributes.
• The regime is benchmarked using precision, recall, and F1 metrics over schema-constrained tables, revealing significant performance gaps in end-to-end retrieval and synthesis.
• It leverages strategies like branch-and-return control, parallel tool invocation, and reinforcement learning to manage depth, breadth, and verification in complex search tasks.

DeepWideSearch denotes a regime of agentic information seeking in which an agent must combine deep, multi-hop retrieval and reasoning with wide, large-scale evidence collection and structured synthesis. In its benchmarked form, the task is not merely to locate a single obscure answer, but to identify and verify core entities through multi-step search and then populate a large, schema-constrained table over many entities and attributes. This joint requirement was formalized as the missing quadrant between deep-only evaluations such as BrowseComp-style tasks and wide-only evaluations such as WideSearch-style table construction, and the benchmark introduced for it reported that current agents remain far from robust end-to-end performance (Lan et al., 23 Oct 2025).

1. Formal definition and task structure

In the canonical formulation, each task is a tuple $(Q, C)$, where $Q$ is a complex natural-language prompt and $C = \{c_i\}_{i=1}^N$ specifies the schema of the required table. The agent must output a structured response $R$ that first identifies the target entities through multi-hop retrieval and reasoning, and then fills all required columns for each row. “Depth” refers to the ability to identify target entities via multi-hop retrieval and explicit constraint satisfaction; “width” refers to comprehensive collection of attributes across many rows with high row- and cell-level fidelity (Lan et al., 23 Oct 2025).

The standard metrics are precision–recall–F1 style measures at multiple granularities. With ground-truth table $G$ and prediction $R$, exact success is $SR = \mathbf{1}[R = G]$. Row-level precision and recall are defined over the row sets, item-level precision and recall over the cell sets, and Column-F1 over the recovered schema. Core Entity Accuracy measures whether the central entity or entity set has been correctly identified. The benchmark reports Avg@4 and Pass@4 by running each system four times per question, while later evaluations also report Max@4 for F1-type metrics (Lan et al., 23 Oct 2025).

The original DeepWideSearch benchmark was built through two conversion pipelines. Deep2Wide converts deep-search instances into deep-plus-wide table tasks, and Wide2Deep converts wide-search instances by synthesizing an additional deep sub-question that must be solved before the table can be completed. The released benchmark contains 220 curated questions across 15 domains in English and Chinese, split into 85 Deep2Wide and 135 Wide2Deep instances. Its average table volume is 414.10 information units per task, and the average number of steps required to identify core entities is 4.21; the Deep2Wide subset has average table volume 247.74 and average steps 3.22, whereas Wide2Deep has average table volume 518.84 and average steps 4.55 (Lan et al., 23 Oct 2025).

2. Benchmark ecology beyond the original DeepWideSearch dataset

The broader benchmark ecosystem around DeepWideSearch separates and recombines depth, breadth, language coverage, and source heterogeneity in different ways. WideSearch evaluates broad information-seeking over 200 manually curated tasks, split evenly between English and Chinese, spanning 18 topics. Each task requires exhaustive collection of atomic facts into a schema-aligned table, and the benchmark’s five-stage quality-control pipeline enforces difficulty, completeness, and verifiability. The average human completion time is 2.33 hours overall, with 44.10 unique source pages consulted per task on average. Under the strict exact-match Success Rate criterion, most systems remain near 0%, the best multi-agent configuration reaches 5.1% Avg@4 SR, and human single-player performance reaches 20.0% SR, while cross-validation by multiple human testers approaches near 100% success (Wong et al., 11 Aug 2025).

Ko-WideSearch extends breadth-oriented evaluation to Korean web sources and makes structural hardness explicit through table width and a $2$-D composite key. It contains 228 tables, 4,262 gold rows, and 14,560 attribute cells over 190 set-parent entities and 16 categories, organized into Easy, Medium, and Hard tiers. The benchmark shows a consistent gap between set recovery and row completion: for GPT-5.5, Item-F1 is 92.8, Column-F1 74.3, Row-F1 53.7, and table success 19.3%. The paper’s diagnosis is that agents often recover the member set but fail on per-cell extraction, normalization, and cross-source verification, especially for open-ended free-text cells (Jeong, 25 Jun 2026).

WideSeekBench targets what its authors call General Broad Information Seeking and emphasizes logical compositionality at scale. It contains 5,156 tasks, with 4,436 train and 720 test instances, covering 18 top-level domains and 200 sub-domains. Constraints are synthesized with logical operators such as AND, OR, and NOT, target entity set sizes are bounded up to 1,024, and target table sizes span from 8 to 8,192 cells. This benchmark is primarily width-centric, but its clause composition and attribute aggregation make it directly relevant to the “wide” component of DeepWideSearch systems (Huang et al., 2 Feb 2026).

HERB shifts the setting from open-web search to heterogeneous enterprise data. Its retrieval pool contains 39,190 artifacts across Slack, meetings, meeting chats, documents, GitHub pull requests, URLs, and customer metadata, with 815 answerable and 699 unanswerable queries. The best agentic RAG method reaches an average score of 32.96, and the analysis concludes that retrieval coverage across heterogeneous sources is the primary bottleneck: models often reason over partial context because they fail to retrieve all necessary evidence chains (Choubey et al., 29 Jun 2025).

3. Inference-time control: balancing depth, breadth, and return

One line of DeepWideSearch research focuses on inference-time control rather than purely on offline training. TreeSeeker formalizes deep search as long-horizon information seeking over dependency-structured sub-goals and organizes search as branch-and-return over sub-goal trees $\mathcal{T}_i$. Its controller, TreeSearch, selects among Exploit, Explore, and Prune at the operation level rather than ranking full paths. For each operation $a$, it predicts ordinal Value, Uncertainty, and Risk signals, $Q$0 with each component in $Q$1, maps them to $Q$2, and scores actions by the parameter-free rule $Q$3. TreeMem stores branch-local evidence, uncertainty, conflicts, progress indicators, failure cues, and the latest leaf trace, with periodic summarization every $Q$4 decision rounds. In the reported implementation, paths selected at least 8 times without escaping a loop should be considered for backtracking, and paths selected at least 12 times must be backtracked. On XBench-DeepSearch, BrowseComp, and BrowseComp-ZH, TreeSeeker with gpt-5.2 scores 56.3, 47.0, and 43.0 respectively, compared with 50.7, 43.0, and 40.3 for Flash-Searcher, while using 1,415.8K tokens and 71.92 tool calls versus 1,474.6K and 90.20 for Flash-Searcher. Ablations show 56.3 for the full system, 52.0 without textual UCB, 48.0 without Explore and Prune, and 51.3 without leaf trace (Shi et al., 10 Jun 2026).

A second control strategy is distribution-aware query probing. WeDas models the search engine as a black-box retrieval function over a web content distribution and introduces the Query-Result Alignment Score, a proxy for how well a query matches the local content landscape. For a probe query $Q$5 and textualized snippets $Q$6, the score is $Q$7, where the three components evaluate topical relevance, information density, and noise robustness on a 0–10 scale. A few-shot probing loop generates derived queries, scores them, and returns distribution-aware guidance for whether to deepen or widen. On GAIA with GPT-5-mini + Miroflow, pass@1 increases from 51.46 to 57.28; on XBench-DS it rises from 38.00 to 44.00; and an ablation on probe iterations shows that $Q$8 gives pass@1 57.28 and pass@3 74.76, compared with 49.51 and 72.82 at $Q$9 (Yu et al., 7 Mar 2026).

A third control perspective is Search Intensity Scaling. DeepDiver defines SIS as the ability to increase search frequency and depth when ambiguity, conflict, or noise is present rather than stopping early with an overconfident answer. Its RL training environment permits up to 7 search rounds, 1–5 queries per round, and top-2 results per query, while the reward includes an additional tool bonus when search-enabled rollouts succeed and no-search rollouts fail. On WebPuzzle, DeepDiver-Pangu-7B reaches 38.1 accuracy with 2.89 average rounds, compared with 37.1 and 1.48 for DeepSeek-R1 with iterative RAG; on the Hard/Outliers subset, it scores 27.9 versus 24.2. This suggests that adaptive search intensity is especially beneficial when ambiguity and conflict cannot be resolved by shallow retrieval (Shi et al., 30 May 2025).

4. Width scaling and multi-agent execution

Where TreeSeeker and related systems regulate branching within a single search process, width-scaling research explicitly parallelizes retrieval. W&D formulates width as the number of concurrent tool calls per reasoning step and depth as the number of sequential turns. In its parallel trace, each step produces $C = \{c_i\}_{i=1}^N$0 and receives $C = \{c_i\}_{i=1}^N$1. The work studies several schedulers, including constant width, ascending width, descending width, and an automatic policy. On BrowseComp scheduler ablations, Constant 1 Tool achieves 66% with 45.7 average turns, Constant 3 Tools 68% with 23.8 turns, Automatic 72% with 26.6 turns, and Descending performs best at 74% with 23.5 turns. On the full BrowseComp dataset, GPT-5-Medium with parallel tool calling reaches 62.2% accuracy without context management. On the first 100 tasks, moving from single-tool to 3-tool parallel calling reduces cost from $C = \{c_i\}_{i=1}^N$265.7 per 100 tasks and wall time from 1522.6 s to 904.2 s while reaching 68% accuracy (Lin et al., 7 Feb 2026).

WideSeek-R1 treats width scaling as a learned organizational problem. Its architecture consists of a lead agent and parallel subagents instantiated from a shared LLM with isolated contexts. The lead agent can spawn up to 10 subagents per turn, has up to 10 turns, and each subagent has up to 20 turns with up to 5 parallel tool calls per turn. Training uses a multi-agent GRPO-style objective with rollout-level shared rewards and dual-level advantage reweighting. On WideSearch, WideSeek-R1-4B achieves Item F1 Avg@4 = 40.0% and Max@4 = 51.8%, Row F1 Avg@4 = 15.3% and Max@4 = 24.4%, and Success Rate Avg@4 = 0.4% with Pass@4 = 1.0%. Its Item F1 Avg@4 is comparable to single-agent DeepSeek-R1-671B at 41.3%, and it improves over the single-agent SingleSeek-R1-4B from 28.1% to 40.0% (Xu et al., 4 Feb 2026).

A-MapReduce reinterprets wide and deep-wide search as a horizontally structured retrieval problem and makes breadth an explicit execution object. Its planning decision is $C = \{c_i\}_{i=1}^N$3, where $C = \{c_i\}_{i=1}^N$4 is a schema-aligned task matrix, $C = \{c_i\}_{i=1}^N$5 is a row-grounded prompt template, and $C = \{c_i\}_{i=1}^N$6 is a batching strategy such as per_atom, by_attr, or open/adaptive. On the DeepWideSearch benchmark, A-MapReduce with GPT-5-mini reports CE Acc. 79.09, Column-F1 51.78, Item-F1 42.11, Row-F1 26.44, and Success Rate 4.43 under Avg@4, with corresponding Pass@4 or Max@4 improvements over strong baselines. Relative to Flash-Searcher with GPT-5-mini, Item-F1 rises from 34.97 to 42.11 and Row-F1 from 19.36 to 26.44, while the framework remains on the reported cost–performance Pareto frontier (Chen et al., 1 Feb 2026).

Web2BigTable addresses both wide and deep regimes through a bi-level architecture: an upper-level orchestrator decomposes tasks and lower-level worker agents solve subtasks in parallel while coordinating through a shared workboard. Long-term strategy learning occurs through a run–verify–reflect loop that updates human-readable SKILL.md memory rather than fine-tuning model weights. On WideSearch, the system reaches Avg@4 Success Rate 38.50, Row F1 63.53, and Item F1 80.12; on XBench-DeepSearch it reaches 73.0 accuracy. Ablations show that removing learned orchestrator skills drops WideSearch SR to 7.00 and XBench accuracy to 41.0, removing the workboard lowers Row F1 to 54.81 and XBench accuracy to 60.0, and removing worker skill evolution lowers Row F1 to 59.67 and XBench accuracy to 64.0 (Huang et al., 29 Apr 2026).

5. Data synthesis, reinforcement learning, and test-time scaling

A second major axis of DeepWideSearch research concerns how agents are trained. DeepDive frames deep search as a POMDP with actions in $C = \{c_i\}_{i=1}^N$7 and combines knowledge-graph-driven data synthesis with end-to-end multi-turn RL. Its KG pipeline uses KILT and AMiner, random walks of length $C = \{c_i\}_{i=1}^N$8, neighbor degree filtering with $C = \{c_i\}_{i=1}^N$9 and $R$0, LLM obfuscation, and adversarial filtering. The resulting dataset contains 3,250 QA pairs, with 1,016 for SFT and 2,234 for RL. DeepDive-32B reports 14.8 on BrowseComp, 25.6 on BrowseComp-ZH, 50.5 on Xbench-DeepSearch, and 29.3 on SEAL-0, improving over its SFT-only version on all four benchmarks. The same work also studies test-time scaling: on a BrowseComp-266 subset, moving from a single sample to 8 parallel samples raises accuracy from 12.0 to 18.8 with majority vote and to 24.8 with minimal-tool-call selection (Lu et al., 12 Sep 2025).

OpenSeeker approaches the problem from a data-centric SFT perspective. It represents the web as a graph $R$1, expands seed pages to local subgraphs $R$2 with controllable size $R$3, synthesizes topology-grounded, entity-obfuscated questions, and uses retrospective summarization to generate denoised teacher trajectories. The released training set contains 10.3k English and 1.4k Chinese samples, for 11.7k total. Trained with simple SFT on Qwen3-30B-A3B-Thinking-2507 and a 256k context window, the system reaches 29.5% on BrowseComp, 48.4% on BrowseComp-ZH, 74.0% on xbench-DeepSearch, and 59.4% item F1 on WideSearch (Du et al., 16 Mar 2026).

OpenSeeker-v2 pushes the same line further by enlarging the knowledge-graph expansion budget, expanding the tool set, and filtering out low-step trajectories with the rule $R$4. Its 10.6k SFT samples have an average of 64.67 tool-call steps per trajectory, compared with 46.97 for OpenSeeker-v1 and 36.01 for RedSearcher. On 30B-scale ReAct evaluation, it reports 46.0% on BrowseComp, 58.1% on BrowseComp-ZH, 34.6% on Humanity’s Last Exam, and 78.0% on xbench-DeepSearch, surpassing Tongyi DeepResearch on all four listed benchmarks (Du et al., 5 May 2026).

Fathom-DeepResearch explicitly makes breadth, depth, and horizon into RL-controlled dimensions. Fathom-Search-4B is trained on DUETQA, a 4,988-sample dataset generated through multi-agent self-play with enforced live-web dependence and heterogeneous source grounding, and optimized with RAPO plus a step-level reward that distinguishes UniqueSearch from RedundantSearch and Exploration from Verification. The reward uses saturation thresholds $R$5 and $R$6 and a verification budget $R$7; the reported Stage-2 setting is $R$8, $R$9, $G$0. The method is reported to enable reliable extension beyond 20 tool calls when warranted. Fathom-Search-4B Stage-2 scores 90.0 on SimpleQA, 64.8 on FRAMES, 50.0 on WebWalker, 22.5 on Seal0, and 33.2 on MuSiQue, while the full Fathom-DeepResearch pipeline reaches 45.47 overall on DeepResearch-Bench with Citation Accuracy 56.1 and Effective Citation Count 38.3 (Singh et al., 28 Sep 2025).

6. Failure modes, limits, and ongoing extensions

The original DeepWideSearch benchmark makes the core difficulty explicit: even strong tool-augmented systems achieve very low exact-match success under joint depth-and-width constraints. The best Avg@4 Success Rate reported there is 2.39% for WebSailor with Claude Sonnet 4, with Avg@4 Row F1 16.88%, Item F1 32.90%, Column F1 42.01%, and Core Entity Accuracy 70.91%. The benchmark’s error analysis identifies four recurring failure modes: lack of reflection, overreliance on internal knowledge, insufficient retrieval, and context overflow, with context overflow reported in 24.96% of cases (Lan et al., 23 Oct 2025).

WideSearch and Ko-WideSearch show that this is not only a problem of obscure multi-hop reasoning, but also of exhaustive completion under schema constraints. WideSearch reports that agents often fail through incomplete query decomposition, lack of reflection and iterative refinement, evidence-utilization mistakes, hallucinated fills, tool invocation failures, formatting errors, and context-length exceedance; even when Item-level F1 becomes high under test-time scaling, exact Success Rate remains low because a single missing or incorrect cell invalidates the table (Wong et al., 11 Aug 2025). Ko-WideSearch sharpens the point by showing that neither more search nor more spend closes the set-versus-row gap: pooled Row-F1 drops from 60.4 on exhaustive-only tables to 32.3 on cross-source tables, and the most difficult cells are open-ended free-text fields rather than dates or names (Jeong, 25 Jun 2026).

A plausible implication is that DeepWideSearch cannot be reduced to either stronger chain-of-thought or more tool calls in isolation. The field’s recent multimodal extension reinforces this view. SimpleSearch-VL, although not framed by its authors as “DeepWideSearch,” improves multimodal agentic search by using Factorized Adaptive Rollout, evidence-verified reasoning with thumbnail checks for reverse image search, and in-agent webpage self-summary rather than external summarizers. With 5,193 supervised trajectories and 1,995 RL prompts, it improves Qwen3-VL agentic baselines by 15.8 average points for the 8B model and 16.0 for the 30B-A3B model, suggesting that the same design pressures that shape text-only DeepWideSearch—verification, efficient rollout allocation, and strict evidence control—also extend to visual and multimodal settings (Dai et al., 30 Jun 2026).

Across these lines of work, DeepWideSearch has emerged less as a single algorithm than as a systems problem with several interacting axes: decomposition of global objectives into dependency-respecting sub-goals, explicit management of branch-local evidence and failure cues, parallelization of width without losing schema consistency, training data that forces live retrieval rather than parametric recall, and evaluation protocols that separate entity identification, cell correctness, row completeness, and exact success. The persistent gaps between Item-F1 and Row-F1, between long-context and retrieval-constrained performance, and between benchmark success and human cross-validated success indicate that future progress will likely depend on tighter retrieval planning, stronger verification loops, and more explicit memory structures than on larger backbones alone.