💬 NLP

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

This paper introduces Table-as-Search (TaS), a structured framework that reformulates long-horizon information seeking as a table completion task to maintain search coherence and state tracking, thereby unifying diverse search strategies and significantly outperforming state-of-the-art baselines in robustness and efficiency.

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

Published 2026-02-09

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Lost in the Middle" Agent

Imagine you hire a very smart but slightly forgetful research assistant (an AI agent) to find specific information on the internet. You ask them a complex question, like: "Find me 30 pet supply stores in North America that grew over 50% last year, and get their sales manager's phone number."

Current AI assistants try to do this by writing a long, continuous story in their head. They think, "I'll search for Store A, then Store B, then check their growth..." As they dig deeper, this mental "story" gets huge. Eventually, the assistant gets overwhelmed. They forget the beginning of the story, lose track of which stores they already checked, or get confused by the sheer volume of notes. In the paper, this is called being "lost in the middle." They might find the right stores but forget to get the phone numbers, or they might hallucinate (make up) facts because their "memory" is too cluttered.

The Solution: The "Table-as-Search" (TaS) Framework

The authors propose a new way to organize this work. Instead of writing a long, messy story, they turn the research task into filling out a spreadsheet (a table).

Think of it like this:

The Old Way: Trying to remember a grocery list while walking through a massive warehouse, hoping you don't forget the milk or the eggs.
The New Way (TaS): You have a clipboard with a grid. Every time you find a product, you write it in a specific box. If a box is empty, you know exactly what you still need to find.

How It Works: The Three Roles

The system uses a team of AI "agents" working together around this digital spreadsheet:

The Planner (The Manager): This AI looks at the user's question and sets up the spreadsheet. It decides what columns are needed (e.g., "Store Name," "Growth Rate," "Phone Number"). It looks at the empty boxes and says, "Okay, we need to find 30 stores to fill these rows."
The Sub-Agents (The Workers): These are the search engines. The Manager sends them out to find candidates.
- Row Expansion: If the table doesn't have enough stores yet, the workers go out and find new ones to add as new rows.
- Cell Population: If a store is found but the "Phone Number" box is empty, the workers go specifically to find that one piece of info and fill in that specific box.
The External Database (The Filing Cabinet): Instead of trying to hold all the search results in the AI's short-term memory (which is limited), the spreadsheet lives in an external database. The AI only looks at the parts of the table it needs right now. This keeps the AI's "brain" clear for thinking, not just storing data.

What They Tested

The paper tested this method on three types of difficult search tasks:

Deep Search: Finding one specific needle in a haystack (e.g., "Find the one singer who fits these 5 specific criteria").
Wide Search: Finding a huge pile of needles (e.g., "List every winner of a prize from 2005 to 2015").
DeepWide Search: The hardest one. Finding a huge pile of needles and verifying that each one meets strict rules (e.g., "Find 30 stores that meet these 5 rules AND get their phone numbers").

The Results: Why It's Better

The paper claims that TaS is significantly better than current state-of-the-art methods (like the standard "ReAct" method used by many big AI companies).

It doesn't get lost: Because the table tracks exactly what is done and what is missing, the AI never forgets its plan, even in very long searches.
It's more efficient: It found better answers using fewer search attempts than the competition. It didn't need to brute-force its way through; it was precise.
It works with smaller brains: Surprisingly, a smaller, cheaper AI model using this "Table" method beat a much larger, more expensive AI model using the old "Story" method. The structure helped the smaller model punch above its weight.
It's flexible: The "Manager" can swap out the "Workers." If you have a specialized tool for finding phone numbers, you can plug it in without breaking the whole system.

In a Nutshell

The paper argues that when AI agents go on long, complex internet scavenger hunts, they shouldn't rely on a messy stream of consciousness. Instead, they should use a structured table to keep their notes organized. This simple change prevents them from getting confused, helps them find more accurate answers, and allows them to do more with less computing power.

Technical Summary: Table-as-Search (TaS)

1. Problem Definition

Current Agentic Information Seeking (InfoSeeking) systems, particularly those relying on Large Language Models (LLMs), face significant challenges in long-horizon exploration. Traditional paradigms like ReAct manage search states (planning procedures and massive search results) within a single, unstructured plain-text context. This approach is inherently fragile for complex tasks, leading to:

"Lost in the Middle" Phenomenon: As the interaction horizon expands, relevant information density dilutes, causing agents to lose track of planning procedures and search history.
State Fidelity Loss: Tracking thousands of search results and corresponding planning steps in a single text trajectory leads to severe hallucinations and error propagation.
Inefficiency in Complex Scenarios: Agents struggle to simultaneously perform information extraction and state tracking within a single forward pass, particularly in tasks requiring both broad aggregation and deep verification.

The paper identifies a gap in handling three distinct InfoSeeking paradigms simultaneously: Deep Search (precise target filtering), Wide Search (broad information aggregation), and the challenging DeepWide Search (simultaneous broad exploration and deep verification).

2. Methodology: Table-as-Search (TaS)

To address these limitations, the authors introduce Table-as-Search (TaS), a structured planning framework that reformulates InfoSeeking as a Table Completion task.

Core Architecture

Structured Schema: Instead of free-form text generation, TaS maps a user query $q$ into a structured schema $S = \langle K, C, I \rangle$ , where $K$ represents key candidates, $C$ denotes constraints, and $I$ denotes required information.
External Database: The framework maintains a structured table in an external database (e.g., MongoDB). Rows represent candidate entities, and columns represent attributes or constraints.
- Filled Cells: Strictly record search history and results.
- Empty Cells: Serve as an explicit, pending search plan.
Offloading: Massive search results are offloaded to the external database, preserving the agent's limited context window for complex reasoning rather than passive information storage.

Multi-Agent Orchestration

TaS is implemented via a multi-agent system centered around a shared database, executing a three-phase process:

Table Initialization: A central Planner (Main-Agent) parses the query and constructs the table schema.
Dynamic Orchestration: The Planner iteratively selects actions:
- Row Expansion: If candidates are insufficient, the Planner formulates diverse search strategies to discover new candidates (Sub-Agents perform parallel broad searches).
- Cell Population: If candidates exist but information is incomplete, Sub-Agents are dispatched in parallel to populate specific cells (constraints verification or information collection).
Answer Synthesis: Upon reaching a saturated state, the Planner retrieves structured evidence from the database to synthesize the final response (e.g., cross-verifying constraints for Deep Search or executing SQL queries for Wide Search).

Unified Representation

TaS provides a unified representation for:

Deep Search: Identifying a unique candidate row satisfying all constraints ( $|C| > 0$ ).
Wide Search: Gathering information for massive candidates under minimal constraints ( $|I| > 0$ ).
DeepWide Search: Maximizing candidate discovery under strict constraints followed by dense information collection ( $|C| > 0, |I| > 0$ ).

3. Key Contributions

Framework Innovation: Proposes TaS, the first framework to reformulate long-horizon InfoSeeking as a structured Table Completion task, unifying Deep, Wide, and DeepWide search paradigms.
State Management: Introduces a mechanism to precisely manage search states by separating "pending actions" (empty cells) from "verified history" (filled cells), mitigating the "lost in the middle" problem.
Benchmark Creation: Curated a challenging DeepWide Search benchmark based on real-world E-commerce Business Development scenarios, addressing the lack of public high-quality evaluations for hybrid search tasks.
Modular Design: Demonstrates a "plug-and-play" architecture where specialized deep search agents (e.g., fine-tuned models) can serve as Sub-Agents, decoupling planning from execution.

4. Experimental Results

Extensive experiments were conducted across three benchmark categories: GAIA/BrowseComp-ZH (Deep Search), WideSearch (Wide Search), and the curated DeepWide Search benchmark.

Performance Superiority: TaS significantly outperforms state-of-the-art baselines (including Multi-Agent ReAct and commercial systems like Gemini DeepResearch) across all three paradigms.
- Deep Search: On GAIA, TaS with the cost-efficient Gemini-2.5-Flash surpassed the Multi-Agent ReAct baseline by +14.0%, demonstrating that structured state management can enable smaller models to rival larger counterparts.
- Wide Search: TaS with Claude-Sonnet-4 (No-Think) achieved comparable Success Rates to the computationally heavier Claude-Sonnet-4 (Thinking) baseline, while significantly outperforming it in Max@4 metrics (9.1% vs. 6.5%).
- DeepWide Search: TaS outperformed the proprietary Gemini DeepResearch system by +4.7% in Column-F1 and +5.1% in Item-Precision.
Efficiency and Robustness:
- Search Efficiency: TaS achieves higher performance with comparable or lower tool usage volumes compared to baselines, proving gains stem from planning quality rather than brute-force search scaling.
- Robustness: As task complexity (difficulty level) increases, the performance gap between TaS and baselines widens. TaS maintains stability in "Hard" tiers where baselines suffer significant degradation (>30% drop).
- Test-Time Scaling: TaS benefits more effectively from increased inference compute (scaling $N$ ) than unstructured ReAct, with performance gaps widening as $N$ increases.
Ablation Studies: The Planner (Main-Agent) is identified as the critical bottleneck; downgrading the planner causes significant performance drops, whereas downgrading Sub-Agents has a milder impact. This confirms the planner's role in orchestrating effective exploration.

5. Significance and Claims

The paper claims that TaS represents a paradigm shift from unstructured text generation to data-centric structured planning for agentic systems.

Architectural Orthogonality: TaS is orthogonal to recent advancements in context management (e.g., summarization) and model training (e.g., RL). It does not rely on compressing text but rather imposes a structured schema to offload state tracking.
Scalability: The framework demonstrates that effective long-horizon InfoSeeking is less dependent on the raw reasoning capability of the foundation model and more dependent on structured state management. This allows cost-effective, smaller models to perform complex tasks when paired with the TaS architecture.
Industrial Applicability: By decoupling planning from execution, TaS allows for the integration of specialized, cost-efficient models for high-frequency search actions, making it a viable solution for industrial-scale applications like E-commerce Business Development.

Limitations Acknowledged:

Non-Search Tasks: The structured schema may introduce rigidity for tasks relying solely on internal knowledge or simple instruction following, leading to performance fluctuations in non-search GAIA instances.
Planner Dependency: Performance is currently bounded by the reasoning capability of the central Planner.
Evaluation Scalability: The DeepWide benchmark relies on human verification due to the open-ended nature of the tasks, which limits large-scale reproducibility compared to closed-domain benchmarks.