PaperlayerView exact revision on arXiv

Source-linked AI summary

Table-as-Search: Formulate Long-Horizon Agentic Information Seeking as Table Completion

Tian Lan, Felix Henry, Bin Zhu, Qianghuai Jia, Junyang Ren, Qihang Pu, Haijun Li, Longyue Wang, Zhao Xu, Weihua Luo

arXiv:2602.06724v1cs.CL

TL;DR

Long-horizon agentic information seeking strains unstructured state tracking as search plans and results accumulate. Table-as-Search reformulates the task as structured table completion and significantly outperforms state-of-the-art baselines across Deep, Wide, and DeepWide Search benchmarks, with improved robustness, efficiency, scalability, and flexibility.

  • Problem

    Current InfoSeeking agents rely on fragile unstructured contexts to track planning and massive search states during long-horizon exploration.

  • Method

    Table-as-Search maps queries to external structured tables whose rows are candidates, columns are constraints or information, and empty cells encode pending searches.

  • Results

    +14.0% on GAIA (52.4% vs. 38.4%) over Multi-Agent ReAct, while TaS significantly outperforms state-of-the-art baselines across Deep, Wide, and DeepWide Search benchmarks.

  • Takeaways & Limitations

    TaS provides a robust, efficient, scalable, and flexible framework for long-horizon agentic information seeking.

  • Takeaways & Limitations

    TaS remains dependent on optimizing the Planner Model, with specialized planners needed to reduce reliance on proprietary models.

Abstract

from arXiv · show

Current Information Seeking (InfoSeeking) agents struggle to maintain focus and coherence during long-horizon exploration, as tracking search states, including planning procedure and massive search results, within one plain-text context is inherently fragile. To address this, we introduce \textbf{Table-as-Search (TaS)}, a structured planning framework that reformulates the InfoSeeking task as a Table Completion task. TaS maps each query into a structured table schema maintained in an external database, where rows represent search candidates and columns denote constraints or required information. This table precisely manages the search states: filled cells strictly record the history and search results, while empty cells serve as an explicit search plan. Crucially, TaS unifies three distinct InfoSeeking tasks: Deep Search, Wide Search, and the challenging DeepWide Search. Extensive experiments demonstrate that TaS significantly outperforms numerous state-of-the-art baselines across three kinds of benchmarks, including multi-agent framework and commercial systems. Furthermore, our analysis validates the TaS's superior robustness in long-horizon InfoSeeking, alongside its efficiency, scalability and flexibility. Code and datasets are publicly released at https://github.com/AIDC-AI/Marco-Search-Agent.

1 Introduction

Agentic InfoSeeking requires multi-step navigation and synthesis across massive web environments, yet plain-text state management is inherently fragile. Table-as-Search addresses this by reformulating InfoSeeking as structured table completion with rows for candidates and columns for constraints or required information.

  • Motivation: Agentic InfoSeeking requires agents to navigate massive web environments and synthesize answers through multi-step reasoning.This capability is central to next-generation Deep Research Systems.
  • Motivation: Current LLM-based agents often use fragile unstructured plain text to manage planning procedures and massive search results.Recent work on context management and procedural planning attempts to mitigate this limitation.
  • Table-as-Search: TaS reformulates InfoSeeking as Table Completion by mapping each query into a schema with candidate-entity rows and constraint or required-information columns.The structured table explicitly manages search states, with filled cells representing search history and results.

A Structured Table-as-Search (TaS) Framework Unified Representation InfoSeeking Challenge

TaS reformulates long-horizon InfoSeeking as structured Table Completion, replacing fragile unstructured planning with iterative row expansion and cell population. This unified representation supports Deep Search, Wide Search, and DeepWide Search.

  • Framework: TaS reformulates InfoSeeking as Table Completion, using row expansion for candidate discovery and cell population for constraint verification or information collection.A central planner orchestrates sub-agents through these iterative operations.
  • Unified Representation: TaS unifies three long-horizon InfoSeeking paradigms: Deep Search, Wide Search, and the challenging DeepWide Search.Deep Search performs precise target filtering, Wide Search aggregates broad information, and DeepWide Search combines broad exploration with deep verification.
  • Empirical Results: Extensive experiments show that TaS significantly outperforms state-of-the-art baselines across Deep Search, Wide Search, and DeepWide Search.The cited baselines include Yao et al. (2022), Wong et al. (2025), and Zhu et al. (2025).

2 Related Work

Related work frames agentic information seeking around Deep, Wide, and hybrid DeepWide Search, while existing planning and context-management methods remain limited by unstructured or lossy state tracking. TaS addresses these limitations with structured search-state management and a new benchmark for DeepWide InfoSeeking.

  • Agentic Information Seeking: Agentic information seeking comprises Deep Search for single-target multi-step reasoning, Wide Search for broad source aggregation, and hybrid DeepWide Search.Benchmarks include BrowseComp for Deep Search and WideSearch for Wide Search, but public high-quality DeepWide evaluations remain lacking.
  • Agentic Information Seeking: The E-commerce Business Development benchmark fills the DeepWide evaluation gap by stress-testing agents on challenging real-world information seeking.The benchmark is explicitly designed for DeepWide InfoSeeking.
  • Agent Frameworks: ReAct-based agent frameworks add procedural planning, but Routine, ADaPT, ReCode, and ReCAP remain bound by unstructured plain-text planning in long-horizon InfoSeeking.ReAct is described as the cornerstone of current agentic systems, while these improvements retain its shared limitation.
  • Context Management: Context summarization, folding, and multi-agent isolation mitigate context overflow but still incur lossy compression and imprecise unstructured search-state recording.TaS is orthogonal to these strategies: it imposes a structured schema rather than compressing text and can incorporate the strategies.

3 Task Formulation

The paper formulates InfoSeeking as web interaction for complex queries and reformulates it as Table Completion for precise long-horizon search-state management. TaS maps queries to schemas and represents Deep, Wide, and DeepWide Search through a unified tabular framework.

  • 3.1 Problem Definition: An InfoSeeking task is defined as T = ⟨q, W⟩, where an agent interacts with web environment W to fulfill complex query q.The interaction unfolds over T steps as trajectory τT = (o1, r1, a1, . . . , oT, rT, aT).
  • 3.2 Table-as-Search (TaS) Framework: The framework reformulates InfoSeeking as a Table Completion problem for precise search-state management.
  • 3.2 Table-as-Search (TaS) Framework: TaS maps query q into schema S = ϕ(q) → S, comprising key candidates K, constraints C, and information to collect I.Varying these attribute-set configurations generalizes the formulation across distinct InfoSeeking paradigms.
  • 3.2 Table-as-Search (TaS) Framework: TaS maintains long-horizon InfoSeeking as table Tt, with rows for discovered or potential candidates and columns for schema attributes.Each cell Tt[i, j] takes values from V ∪ {∅, N/A}; ∅ marks pending information and N/A marks information that need not be retrieved.
  • 3.2 Table-as-Search (TaS) Framework: Deep Search identifies a unique candidate row satisfying all constraints, often through complex multi-hop verification to filter false positives.Its defining condition is |C| > 0.
  • 3.2 Table-as-Search (TaS) Framework: Wide Search gathers required information for many candidates under minimal constraints, while DeepWide Search combines strict constraint satisfaction with dense information collection.Wide Search has |I| > 0; DeepWide Search has |C| > 0 and |I| > 0.

4 Implementation of TaS Framework

TaS is implemented as a multi-agent system centered on a shared structured database table and organized into table initialization, dynamic orchestration, and answer synthesis. The Planner and Sub-Agents manipulate the table through database operations, enabling parallel candidate discovery and cell population before synthesizing a response from verified evidence.

  • Table Initialization: The Planner parses the user query and initializes the database table structure.This occurs through the ConstructSchema operation.
  • Dynamic Orchestration: During dynamic orchestration, the Planner selects between parallel row expansion and cell population based on table state and candidate validity.Row expansion formulates n diverse search strategies when candidates are insufficient or fail constraints, while Sub-Agents populate cells for each candidate.
  • Dynamic Orchestration: Both Main-Agent and Sub-Agent update the shared table through database interfaces, using AppendRow for candidates and UpdateRow for results.The framework can integrate specialized off-the-shelf search agents as Sub-Agents.
  • Answer Synthesis: After table saturation or timeout, the Planner retrieves structured evidence to synthesize the final response and export verified candidates with SQL for Wide and DeepWide Search.For Deep Search, filled table entries are used to cross-verify constraints for a precise conclusion.

5 Experimental Setup

The evaluation covers Deep, Wide, and DeepWide Search through specialized benchmarks and metrics, comparing TaS with agentic framework baselines under a cost-conscious long-context setup.

  • Benchmarks: Three benchmark categories evaluate distinct long-horizon InfoSeeking capabilities: Deep Search, Wide Search, and DeepWide Search.Deep Search uses GAIA and BrowseComp-ZH for multi-step reasoning and precise filtering, while WideSearch measures broad information aggregation.
  • Evaluation Metrics: Deep Search reports Accuracy using standard LLM-as-a-Judge protocols, whereas Wide Search reports stable Avg@4 values for four table-completion metrics.The Wide Search metrics are Column-F1 (Candidate Acc.), Row-F1 (Row-level Acc.), Item-F1 (Cell-level Acc.), and Success Rate (SR, Table-level Acc.).
  • Benchmarks: DeepWide Search introduces 20 challenging long-horizon E-commerce queries requiring extensive candidate discovery, constraint verification, and information collection.Experts annotate the benchmark and report Column-F1 and Item-F1.
  • Experimental Scale and Cost: The setup balances representative evaluation with computational feasibility because long-horizon execution costs over $5,000.The authors note that broader benchmark coverage is constrained by this prohibitive execution cost.
  • Baselines: TaS is compared with Single-Agent ReAct, Multi-Agent ReAct, and compute-scaled variants instantiated with diverse foundation models.Multi-Agent ReAct serves as the state-of-the-art baseline for Wide Search and Deep Search.
  • Implementation: Experiments use SmolAgent and WideSearch with Google Search and Webpage Visit, an 64k-token maximum context window, and summarization strategies to reduce cost.Training-based search sub-agents run on 8 NVIDIA A100 GPUs, with further implementation details provided in Appendix A.

6 Main Results

TaS consistently outperforms unstructured and proprietary baselines across Deep, Wide, and DeepWide Search benchmarks. Its structured planning improves long-horizon search quality while preserving efficiency and enabling flexible execution architectures.

  • Deep Search: TaS surpasses Multi-Agent ReAct by 14.0% on GAIA with Gemini-2.5-Flash, scoring 52.4% versus 38.4%.It also outperforms the stronger Qwen3-Max counterpart in this comparison.
  • Deep Search: TaS maintains a +2.5% advantage on GAIA’s search-dependent subset, despite a -18.2% regression on non-search tasks.The regression is attributed to structured-table overhead being unnecessary for simple internal agentic tasks.
  • Wide Search: On WideSearch, TaS matches ReAct-MA on Success Rate at 3.5% versus 3.6%, then leads at Max@4 with 9.1% versus 6.5%.The benchmark averages 274.8 table cells per query, stress-testing agents over a massive search space.
  • Wide Search: TaS improves WideSearch coverage over ReAct-MA by +8.4% in Column-Recall and +6.9% in Item-Recall, while raising Item-Precision by +4.4%.The results indicate that increased recall does not require sacrificing precision.
  • DeepWide Search: On DeepWide Search, TaS exceeds ReAct-MA and Gemini DeepResearch by +4.7% in Column-F1 and +5.1% in Item-Precision.The result supports structured planning as an advantage in complex long-horizon InfoSeeking.
  • DeepWide Search: Replacing the sub-agent with a fine-tuned 32B deep search model yields Column-F1 of 55.9% versus 52.7% and Item-Precision of 67.7% versus 63.5%.Candidate discovery shows a marginal trade-off, while information retrieval precision improves substantially.

7 Analysis

TaS becomes more advantageous as long-horizon search complexity and test-time compute increase, with gains attributed to structured planning and state management rather than search volume alone. Ablations identify the Planner Main-Agent as the critical component while showing that specialized Sub-Agents can be plugged in to amplify performance.

  • RQ1: Robustness: TaS’s Deep Search performance gap over baselines expands from +14.3% on Med-Hard to +17.9% on Hard instances as complexity increases.Difficulty is measured by constraint count |C| for Deep Search.
  • RQ2: Search Efficiency: TaS (Max@2) significantly outperforms ReAct-MA (Max@4) across all Wide Search tool-usage segments with comparable or lower tool usage.This supports structured planning and state management as the source of TaS’s advantage, rather than increased search volume.
  • RQ2: Search Efficiency: TaS Num@1 surpasses ReAct-MA Num@4 (199.7 > 199.4), while TaS Num@4 approaches the ground-truth upper bound (251.1 vs. 274.8).Num@k measures the maximum valid cells achieved across k trials in WideSearch.
  • RQ3: Test-Time Scaling: On BrowseComp-ZH, TaS’s performance gap over ReAct-MA widens from +2.4% at N=1 to +7.2% at N=2 during test-time scaling.On WideSearch, the advantage also increases from +4.0% at N=3 to +4.4% at N=4.
  • RQ4: Ablation Studies: Downgrading the Planner from Qwen3-Max to Qwen3-30B-A3B causes a significant performance drop, whereas downgrading the Sub-Agent has a much milder impact.The ablation identifies the Planner Main-Agent as the critical bottleneck in TaS.
  • RQ4: Ablation Studies: Replacing Gemini-2.5-Flash with MiroThinker-8B substantially improves most metrics, and TaS with Gemini significantly outperforms standalone MiroThinker-8B on all metrics.These results characterize the Sub-Agent as plug-and-play and show that TaS unlocks and amplifies specialized deep-search models.

8 Conclusion

The Table-as-Search (TaS) framework reformulates long-horizon agentic information seeking as table completion with structured search-state tracking. Experiments show that TaS outperforms state-of-the-art baselines across Deep, Wide, and DeepWide Search benchmarks while demonstrating robustness, efficiency, scalability, and flexibility.

  • 8 Conclusion: TaS reformulates long-horizon agentic InfoSeeking as a Table Completion task.The framework maps each user query to a structured table schema for precise search-state tracking.
  • 8 Conclusion: TaS significantly outperforms state-of-the-art baselines across Deep, Wide, and DeepWide Search benchmarks.The reported benchmarks cover the three search settings unified by the framework.
  • 8 Conclusion: TaS exhibits superior robustness, efficiency, scalability, and flexibility for building more robust InfoSeeking agents.These properties are reported as framework-level advantages in the conclusion.

Limitation … C More Experimental Results

The paper identifies limitations in TaS’s generalization and planner dependence while detailing benchmark construction, fine-tuning, database tools, complexity analysis, and the framework’s documented process. Its experimental appendices define specialized evaluation protocols and implementation settings for Deep, Wide, and DeepWide Search.

  • Limitation: TaS remains unstable on non-search GAIA instances because its retrieval-oriented tabular schema can be unnecessarily rigid for internal-knowledge or instruction-following tasks.The paper identifies general-purpose agentic-task applicability as an open limitation.
  • Limitation: TaS’s execution layer can use smaller, cost-efficient models without performance loss, but weaker Planner models cause significant performance degradation.Training-based search agents can integrate as plug-and-play Sub-Agents, whereas future work prioritizes specialized Planner optimization, potentially through Agentic RL.
  • A.1 Benchmarks and Metrics: DeepWide evaluation fixes each query’s retrieval target, such as 30 candidates, and uses expert human annotation because exhaustive ground truth is computationally infeasible.Four BD experts annotate results, and the benchmark reports Column-F1 and Item-F1.
  • A.1 Benchmarks and Metrics: DeepWide Search uses 20 high-quality, complex Business Development samples requiring complex filtering followed by deep information extraction.The benchmark addresses the lack of open datasets combining horizontal breadth with vertical depth.
  • A.2 Fine-tuning Deep Search Sub-Agent: The Deep Search Sub-Agent uses Qwen3-32B, trained from approximately 12K samples through trajectory distillation and reverse synthesis.The training setup applies SFT with 64K context windows, a learning rate of 5 × 10−5, and 64 NVIDIA A100 GPUs for five hours.
  • A.3 Tools for Table Operation: The tabular memory system uses MongoDB with PyMongo and six atomic primitives for creating tables, adding and updating records, inspecting snapshots, counting rows, and filtering pending records.Operations follow standard PyMongo syntax, including operators such as $set and $exists.
  • A.4 Experimental Setup for Analysis: The analysis classifies Deep Search and Wide Search samples into Easy, Med-Easy, Medium, Med-Hard, and Hard categories using benchmark-specific complexity metrics.Subset experiments use 100 BrowseComp-ZH samples and 40 WideSearch samples because of limited API quotas.
  • B Detailed Process of TaS: Figure 8 presents the detailed TaS process in alignment with Algorithm 1.This appendix documents the framework’s procedural correspondence between the figure and algorithm.

C.1 Full Results on GAIA … Case Study on WideSearch #EN-059

TaS consistently improves search-oriented performance, including upper-bound aggregation, efficiency, and robustness under extreme context loads. Case studies illustrate how schema filling supports global verification and complete structured outputs across Deep, Wide, and DeepWide Search settings.

  • C.1 Full Results on GAIA: TaS consistently outperforms state-of-the-art baselines on search-dependent GAIA samples, but its performance is unstable on tasks that do not require searching.Table 8 separates 80 search-dependent samples from 23 no-search samples.
  • C.2 Max@4 Performance on WideSearch: 9.1% Success Rate is achieved by TaS with Claude-Sonnet-4 (NoThink), surpassing Multi-Agent ReAct with Claude-Sonnet-4 (Thinking) at 6.5%.The comparison evaluates Max@4 performance as an upper bound for massive information aggregation.
  • C.2 Max@4 Performance on WideSearch: TaS achieves its WideSearch performance gains with comparable or lower tool usage than the Multi-Agent ReAct baseline.The analysis attributes these gains to structured planning precision rather than brute-force search scaling.
  • C.4 Robustness Analysis on WideSearch: 32.3% Item-F1 is maintained by TaS in the Hard setting, exceeding Multi-Agent ReAct’s 21.4% by +10.9% when the state space exceeds 1,500 cells.TaS outperforms the baseline across all difficulty tiers, indicating robustness against extreme context overload.
  • D.1 Qualitative Analysis: TaS mitigates premature convergence by enforcing global verification through schema filling across multiple candidates and constraints.The case study contrasts this with ReAct stopping at partial matches, such as identifying “Hu Xia” while missing the album age.
  • D.2 Search and No-Search Cases in GAIA: GAIA analysis distinguishes 80 search-dependent samples requiring web interaction from 23 no-search samples solvable through reasoning, calculation, or coding.Figure 11 contrasts the behavioral requirements of the two categories.
  • Case Study on BrowseComp-ZH #141: The BrowseComp-ZH case asks for a singer satisfying university, birthplace, television-theme-song, and debut-album constraints.The query requires identifying one candidate meeting all stated biographical and career conditions.
  • Case Study on WideSearch #EN-059: The WideSearch case requires verifying every TED Prize winner from 2005 to 2015 in a Markdown table with Year, Winner, TED Talk Title, and Host City columns.A related DeepWide case requests 20 US lighting manufacturers or merchants with platform, store, contact, and product-count fields.
Loading 2602.06724v1…