WideSeek-R1: Exploring Width Scaling for Broad Information Seeking via Multi-Agent Reinforcement Learning

Zelai Xu^1*, Zhexuan Xu^1*, Ruize Zhang^2*, Chunyang Zhu³, Shi Yu⁴,
Weilin Liu³, Quanlu Zhang³, Wenbo Ding², Chao Yu^2†, Yu Wang^1†

¹EE, Tsinghua University    ²SIGS, Tsinghua University    ³Infinigence AI    ⁴IIIS, Tsinghua University

^*Equal Contribution.   ^†Corresponding Authors.

Paper: arXiv:2602.04634

Code: WideSeek-R1 code

Data: Training Data | Corpus

Model: WideSeek-R1-4B

Quick Start: Quick Start Guide

Abstract

Recent advancements in Large Language Models (LLMs) have largely focused on depth scaling, where a single agent solves long-horizon problems with multi-turn reasoning and tool use. However, as tasks grow broader, the key bottleneck shifts from individual competence to organizational capability. In this work, we explore a complementary dimension of width scaling with multi-agent systems to address broad information seeking. Existing multi-agent systems often rely on hand-crafted workflows and turn-taking interactions that fail to parallelize work effectively. To bridge this gap, we propose WideSeek-R1, a lead-agent–subagent framework trained via multi-agent reinforcement learning (MARL) to synergize scalable orchestration and parallel execution. By utilizing a shared LLM with isolated contexts and specialized tools, WideSeek-R1 jointly optimizes the lead agent and parallel subagents on a curated dataset of 20k broad information-seeking tasks. Extensive experiments show that WideSeek-R1-4B achieves an item F1 score of 40.0% on the WideSearch benchmark, which is comparable to the performance of single-agent DeepSeek-R1-671B. Furthermore, WideSeek-R1-4B exhibits consistent performance gains as the number of parallel subagents increases, highlighting the effectiveness of width scaling.

Figure 1: Comparison of depth and width scaling. While depth scaling enhances performance through sequential multi-turn interactions, width scaling orchestrates multi-agent systems for parallel execution. WideSeek-R1 pushes the frontier of width scaling via MARL for synergized orchestration and execution.

Contributions

• We introduce WideSeek-R1, a multi-agent system trained via MARL to synergize scalable orchestration and parallel execution for broad information seeking.

• We open-source a large-scale dataset of 20,000 broad information-seeking tasks, offering a complementary training resource to existing multi-hop datasets.

• We demonstrate the effectiveness of width scaling with WideSeek-R1-4B, which achieves comparable performance to DeepSeek-R1-671B and exhibits consistent gains as the number of parallel agents increases.

Motivation

As tasks grow broader, width scaling via multi-agent systems becomes essential, yet both single-agent methods and existing multi-agent systems fall short in different ways. Broad information seeking, which requires gathering and synthesizing attributes of multiple entities into a structured table, serves as an ideal testbed for this challenge.

Limitations of Single-Agent Methods

Single-agent methods face two fundamental limitations when tasks grow in breadth.

• Context pollution. As the agent’s context accumulates information from previous subtasks, irrelevant content increasingly interferes with reasoning, degrading performance on later subtasks.

• Sequential execution. A single agent must process independent subtasks one by one, leaving parallelizable work serialized and making the overall process inefficient.

These limitations underscore the necessity of multi-agent systems, which naturally enable context isolation and parallel execution for effective width scaling.

Limitations of Existing Multi-Agent Systems

Despite their promise, existing multi-agent systems have not fully realized the potential of width scaling, primarily because few systems are trained end-to-end to learn scalable orchestration and parallel execution.

• Hand-crafted orchestration. Most prior work relies on manually designed workflows rather than learned agents, hindering flexible and scalable coordination as the number of agents grows.

• Turn-taking execution. Current systems typically process subtasks one at a time through turn-taking interactions, serializing progress and failing to parallelize independent work.

As a result, the performance of existing multi-agent systems is bottlenecked by limited scalability and insufficient parallelization. WideSeek-R1 is designed to address both levels through end-to-end multi-agent reinforcement learning.

Method

WideSeek-R1 is a hierarchical lead-agent-subagent system trained via end-to-end MARL to synergize scalable orchestration and parallel execution for width scaling. The lead agent and subagents share a single LLM but operate with isolated contexts and specialized tools: the lead agent focuses on task decomposition and orchestration, while each subagent executes its assigned subtask in parallel using external tools to gather information and return findings.

Figure 2. Overview of WideSeek-R1 rollout and training pipeline. Rollout: The lead agent coordinates task decomposition while subagents execute parallel subtasks. Training: A shared model is trained via GRPO with multi-agent advantage assignment and dual-level advantage reweighting.

Lead Agent for Scalable Orchestration

The lead agent decomposes a broad task into parallelizable subtasks and delegates them to subagents. Unlike systems with fixed hand-crafted workflows, the lead agent is trained for scalable and learnable orchestration as the number of subagents increases. The lead agent is restricted to a call_subagent-style tool to reduce context pollution.

Subagents for Parallel Execution

Subagents are responsible for parallel information seeking, enabling width scaling by executing multiple subtasks simultaneously. This design addresses both context pollution and sequential execution bottlenecks in single-agent pipelines. Subagents use two tools:

• search: retrieves relevant snippets and URLs for a query.

• access: summarizes information from a specific URL.

Multi-Agent Reinforcement Learning

WideSeek-R1 jointly optimizes the lead agent and subagents with a shared model, enabling simultaneous learning of orchestration and information-seeking behavior. The method extends GRPO for multi-agent systems with two key designs:

• Multi-Agent Advantage Assignment: use a verifiable outcome reward per multi-agent rollout and assign the same advantage to all agents and all tokens for stable training.

• Dual-Level Advantage Reweighting: use both agent-level and token-level reweighting in the policy gradient objective for multi-agent multi-turn optimization.

\[\mathbb{E}\left[ \frac{1}{G}\sum_{i=1}^{G} \frac{1}{N_i}\sum_{a=1}^{N_i} \frac{1}{\sum_{t=1}^{T_{i,a}} |o^t_{i,a}|} \sum_{t=1}^{T_{i,a}}\sum_{j=1}^{|o^t_{i,a}|} \min\left( r^{t,j}_{i,a}\hat{A}_i, \operatorname{clip}\left(r^{t,j}_{i,a},1-\epsilon_{\mathrm{low}},1+\epsilon_{\mathrm{high}}\right)\hat{A}_i \right) \right]\]

Training Data Construction

To unlock width scaling, WideSeek-R1 requires large quantities of broad information-seeking tasks. We build a fully automated data construction pipeline that synthesizes high-quality training instances with schema-constrained queries and standardized tabular outputs, yielding 20,000 instances.

Figure 3. Automated data construction pipeline with three stages: Query Generation, Answer Generation, and QA Pair Filtering.

The pipeline includes three stages:

1. Query Generation: extract user intents from HybridQA and refine them into complex schema-constrained queries with explicit table requirements.

2. Answer Generation: generate two independent responses and identify unique columns for self-consistency verification.

3. QA Pair Filtering: remove low-consistency or low-difficulty samples to keep robust and challenging instances.

Experiment Results

Main Results on WideSearch

WideSeek-R1-4B achieves the best results on five out of six metrics among 4B and 8B baselines. The multi-agent system consistently outperforms the single-agent variant, with an absolute improvement of 11.9 points in item F1 score, and attains an 8.8-point gain over the base Qwen3-4B in the same multi-agent setting. Notably, WideSeek-R1-4B is comparable to single-agent DeepSeek-R1-671B with nearly 170x fewer parameters.

Table 1a. Single-agent results on the WideSearch benchmark

Model	Item F1 Avg@4	Item F1 Max@4	Row F1 Avg@4	Row F1 Max@4	Success Avg@4	Success Pass@4
SingleSeek-R1-4B	28.1	39.2	6.5	12.5	0.3	1.0
Qwen3-4B	20.1	30.2	3.0	4.8	0.0	0.0
Search-R1-7B	15.5	24.4	2.0	4.4	0.0	0.0
ASearcher-7B	16.5	26.0	2.8	5.8	0.0	0.0
DeepSeek-R1-671B	41.3	55.1	20.7	31.7	0.4	1.5

Table 1b. Multi-agent results on the WideSearch benchmark

Model	Item F1 Avg@4	Item F1 Max@4	Row F1 Avg@4	Row F1 Max@4	Success Avg@4	Success Pass@4
WideSeek-R1-4B	40.0	51.8	15.3	24.4	0.4	1.0
Qwen3-4B	31.2	42.3	8.4	15.5	0.0	0.0
AgentFlow-7B	28.7	45.4	9.0	20.2	0.4	1.5
OWL-8B	20.2	29.3	3.1	5.8	0.0	0.0
MiroFlow-8B	23.7	37.7	5.8	12.7	0.4	1.0

Exploring Width Scaling

To compare depth scaling and width scaling under test-time compute, we plot performance as a function of the number of turns (depth) and the number of agents (width).

• Depth Scaling: performance initially improves with more turns, then quickly plateaus due to fixed context limits.

• Width Scaling: performance initially improves with more subagents but can decline at larger width when an untrained lead agent receives conflicting signals.

• Width Scaling + MARL: performance improves consistently as subagents increase, reaching 40.0% item F1 with 10 subagents.

Figure 4. Comparison of depth and width scaling with respect to test-time compute. The blue curve is depth scaling vs turns; the red curves are width scaling vs number of subagents.

Ablation Studies

We conduct ablations to answer two questions: (1) whether jointly optimizing both the lead agent and subagents is necessary, and (2) how data composition affects final capability.

Figure 5. Ablation on lead agent and subagents.

Figure 6. Ablation on training data composition.

• Lead Agent and Subagents: best performance is achieved when both roles use WideSeek-R1-4B. Upgrading either role alone gives comparable gains, and combining both yields additional improvements, confirming role synergy under end-to-end MARL.

• Training Data: a hybrid dataset (wide + deep) consistently outperforms wide-only or deep-only training at equal data size, indicating complementary benefits between orchestration-focused and execution-focused data.

Standard QA Benchmarks

To evaluate generalization beyond broad information seeking, we test on seven open-domain QA benchmarks (three single-hop and four multi-hop). WideSeek-R1-4B reaches an average score of 59.0%, outperforming its backbone multi-agent Qwen3-4B by 7.7 points and surpassing larger multi-agent systems such as OWL-8B and MiroFlow-8B.

Table 2a. Single-agent results on standard QA benchmarks (3 single-hop + 4 multi-hop)

Model	Avg.	NQ	TriviaQA	PopQA	2Wiki	HotpotQA	Bamboogle	MuSiQue
SingleSeek-R1-4B	57.0	58.8	78.3	48.0	70.9	62.1	54.6	26.5
Qwen3-4B	48.3	48.5	68.7	43.0	58.9	51.4	48.2	19.2
Search-R1-7B	55.4	49.9	78.0	55.7	58.1	60.8	58.4	27.1
ASearcher-7B	61.0	54.5	79.3	55.9	77.6	67.6	60.0	32.6

Table 2b. Multi-agent results on standard QA benchmarks (3 single-hop + 4 multi-hop)

Model	Avg.	NQ	TriviaQA	PopQA	2Wiki	HotpotQA	Bamboogle	MuSiQue
WideSeek-R1-4B	59.0	56.1	78.5	48.5	75.0	64.2	61.8	28.9
Qwen3-4B	51.3	49.6	70.7	44.9	65.0	54.3	52.6	21.7
AgentFlow-7B	61.0	58.5	87.0	52.5	77.2	57.0	69.6	25.3
OWL-8B	57.2	64.0	74.2	52.2	62.6	61.0	55.8	30.4
MiroFlow-8B	50.0	50.9	73.1	42.8	58.6	52.4	50.8	21.3

Citation

@article{xu2026wideseek,
  title={WideSeek-R1: Exploring Width Scaling for Broad Information Seeking via Multi-Agent Reinforcement Learning},
  author={Xu, Zelai and Xu, Zhexuan and Zhang, Ruize and Zhu, Chunyang and Yu, Shi and Liu, Weilin and Zhang, Quanlu and Ding, Wenbo and Yu, Chao and Wang, Yu},
  journal={arXiv preprint arXiv:2602.04634},
  year={2026},
}