🤖 AI

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

This paper introduces WideSeek-R1, a multi-agent reinforcement learning framework that employs a lead-agent-subagent architecture to achieve broad information-seeking performance comparable to massive single-agent models by leveraging width scaling and parallel execution.

Original authors: Zelai Xu, Zhexuan Xu, Ruize Zhang, Chunyang Zhu, Shi Yu, Weilin Liu, Quanlu Zhang, Wenbo Ding, Chao Yu, Yu Wang

Published 2026-03-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zelai Xu, Zhexuan Xu, Ruize Zhang, Chunyang Zhu, Shi Yu, Weilin Liu, Quanlu Zhang, Wenbo Ding, Chao Yu, Yu Wang

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 Big Idea: From "Super-Genius" to "Super-Team"

Imagine you need to organize a massive international conference. You have two ways to do it:

The "Depth" Approach (The Old Way): You hire one Super-Genius. This person is incredibly smart, knows everything, and works alone. They spend days reading thousands of documents, making lists, and calling people one by one. Eventually, they finish the job. This is how current AI models (like DeepSeek-R1) work. They get smarter by getting bigger and thinking longer in a single line of thought.
The "Width" Approach (The New Way): You hire a Project Manager and a team of 20 Specialists. The Manager breaks the big job into small chunks (e.g., "Find venues in Tokyo," "Find venues in Paris"). The 20 Specialists all go to work at the same time, gathering info in parallel. The Manager then stitches the pieces together.

WIDESEEK-R1 is the paper that says: "Stop trying to make the Super-Genius smarter. Instead, build a better Team Manager and train the whole team to work together perfectly."

The Problem: The "Context Pollution" Mess

Why not just use the Super-Genius for everything?

Imagine the Super-Genius is trying to research 50 different universities.

Turn 1: They read about Harvard.
Turn 2: They read about Yale.
Turn 3: They read about Princeton.

By Turn 50, the Super-Genius is drowning in a sea of information. They forget what they read in Turn 2 because the "memory" (context window) is too full. This is called Context Pollution. Also, they have to do it one by one, which is slow.

The Solution: WIDESEEK-R1

The authors built a system called WIDESEEK-R1 to solve this. Here is how it works, using our analogy:

1. The Structure: The Conductor and the Orchestra

The Lead Agent (The Conductor): This is the "Manager." Its only job is to listen to the big question, break it down into small, independent tasks, and hand them out. It doesn't do the research itself; it just directs traffic.
The Subagents (The Musicians): These are the workers. They receive a small task (e.g., "Find the founding year of Yale"). They work in parallel (all at once), so they don't get confused by each other's work. They have their own "clean" workspace.

2. The Secret Sauce: Multi-Agent Reinforcement Learning (MARL)

You might think, "Okay, just give them a checklist." But the paper says that doesn't work well. If you just tell them what to do, they might argue or get lazy.

Instead, the authors used Reinforcement Learning (like training a dog or a video game character).

They created a massive training ground with 20,000 complex research tasks.
They let the "Conductor" and the "Musicians" try to solve these tasks.
The Reward: If the final table of information was perfect, the whole team got a treat (a reward). If the Conductor gave bad instructions or the Musicians got lazy, they got no treat.
The Result: Over time, the Conductor learned exactly how to split the work, and the Musicians learned exactly how to do their part efficiently. They learned to synergize (work together better than the sum of their parts).

The Amazing Results

The paper tested this system on a benchmark called WideSearch (finding broad information about many things at once).

The Underdog Wins: They used a tiny model (4 Billion parameters) for their team.
The Giant Loses: They compared it to a massive, single-agent model (DeepSeek-R1 with 671 Billion parameters).
The Outcome: The tiny team (WIDESEEK-R1-4B) performed just as well as the massive giant!

Why is this a big deal?
It means you don't need a super-expensive, massive computer to do complex research. You can use a small, cheap model if you organize it into a smart, parallel team.

The "Scaling" Lesson: Depth vs. Width

The paper draws a cool line between two ways to get better:

Depth Scaling (The Ladder): You keep climbing the ladder (adding more turns of thinking). But eventually, you hit the ceiling. The Super-Genius gets tired, confused, or stuck.
Width Scaling (The Net): You throw a wider net. You add more people to the team. As long as the Conductor is smart enough to manage them, the more people you add, the better the results get.

Summary in One Sentence

WIDESEEK-R1 proves that instead of building one giant, overworked AI brain, we should build a small, well-trained AI team that works in parallel, allowing small models to solve massive problems as well as giant ones.

1. Problem Statement

The paper addresses the limitations of current Large Language Model (LLM) scaling strategies, which predominantly focus on depth scaling (increasing model parameters or extending reasoning chains via multi-turn interactions for a single agent). While effective for deep, complex queries, depth scaling hits a bottleneck when tasks require broad information seeking—gathering and synthesizing attributes of multiple entities into structured formats (e.g., tables).

Key challenges in broad information seeking for single agents include:

Context Pollution: As a single agent processes multiple subtasks sequentially, irrelevant information accumulates in the context window, degrading performance.
Sequential Inefficiency: Processing independent subtasks one-by-one is inefficient and fails to leverage parallel computational resources.
Organizational Bottleneck: As tasks broaden, the bottleneck shifts from individual agent competence to the system's ability to orchestrate and coordinate multiple agents.

Existing multi-agent systems often rely on hand-crafted workflows or turn-taking interactions that serialize execution, failing to achieve true parallelism or learn scalable orchestration end-to-end.

2. Methodology: WIDESEEK-R1

The authors propose WIDESEEK-R1, a hierarchical multi-agent framework trained via Multi-Agent Reinforcement Learning (MARL) to synergize scalable orchestration and parallel execution.

A. System Architecture

The system consists of two roles instantiated using a shared LLM (Qwen3-4B) but with isolated contexts and specialized tools:

Lead Agent:
- Role: Task decomposition and orchestration.
- Tools: Restricted to a single tool, call_subagent, to prevent context pollution.
- Function: Decomposes broad objectives into independent subtasks, generates clear prompts for subagents, and waits for all parallel subagents to complete before proceeding to the next turn.
Subagents:
- Role: Parallel information seeking.
- Tools: Equipped with search (retrieving snippets/URLs) and access (summarizing specific URLs).
- Function: Execute assigned subtasks in parallel within isolated contexts, gathering and synthesizing information to return findings to the lead agent.

B. Training via MARL

The framework is trained end-to-end using a modified Group Relative Policy Optimization (GRPO) algorithm. Key innovations include:

Multi-Agent Advantage Assignment: Instead of complex credit assignment, the system assigns a group-normalized advantage ( $\hat{A}_i$ ) to all agents and tokens within a single rollout. The reward is based on the final answer's consistency with the ground truth. This ensures joint optimization without reward hacking.
Dual-Level Advantage Reweighting:
- Token-level: Averages advantages over all tokens produced by an agent across turns to prevent short turns from being diluted.
- Agent-level: Averages advantages over all agents in a rollout. This prevents the system from spawning excessive subagents that do not improve the final reward, ensuring that adding agents only happens if it benefits the outcome.

C. Data Construction

To train the system, the authors constructed a 20k broad information-seeking dataset using an automated pipeline:

Query Generation: Extracts intents from HybridQA and refines them into complex, schema-constrained queries requiring tabular outputs (10–50 rows).
Answer Generation: Uses a high-capacity model (Gemini-3-pro-preview) to generate two independent answers and identifies "unique columns" for robust row alignment.
Filtering: Applies strict consistency checks (cell-wise match rate > 0.9) and complexity filters to ensure high-quality, challenging training data.

3. Key Contributions

WIDESEEK-R1 Framework: A novel lead-agent–subagent system trained via MARL that effectively synergizes scalable orchestration with parallel execution, moving beyond hand-crafted workflows.
New Dataset: Release of a large-scale, 20k-instance dataset of broad information-seeking tasks, filling a gap in existing resources that are mostly tailored for depth scaling (multi-hop QA).
Width Scaling Demonstration: Empirical evidence that width scaling (adding parallel agents) yields consistent performance gains, whereas depth scaling (adding turns) quickly plateaus.

4. Experimental Results

The model was evaluated on the WideSearch benchmark and standard QA datasets.

Performance on WideSearch:
- WIDESEEK-R1-4B achieved an Item F1 score of 40.0%.
- This performance is comparable to the single-agent DeepSeek-R1-671B (41.3%), despite using nearly 170x fewer parameters.
- It significantly outperformed other 4B and 8B baselines (e.g., Qwen3-4B, Search-R1-7B, AgentFlow-7B).
Width Scaling Behavior:
- As the number of parallel subagents increased (up to 10), WIDESEEK-R1-4B showed continuous performance gains.
- In contrast, the base model (Qwen3-4B) saw performance decline after a certain number of subagents due to noise accumulation and lack of orchestration training.
Generalization:
- On standard single-hop and multi-hop QA benchmarks (NQ, HotpotQA, etc.), WIDESEEK-R1-4B surpassed larger multi-agent systems (OWL-8B, MiroFlow-8B) and its single-agent variant, proving that width scaling does not compromise general reasoning capabilities.

5. Significance and Impact

Paradigm Shift: The paper challenges the prevailing "bigger is better" (depth scaling) narrative, demonstrating that organizational capability (width scaling) via multi-agent collaboration can rival massive models with significantly smaller parameter counts.
Efficiency & Democratization: By achieving state-of-the-art results with a 4B model, the approach lowers the computational barrier for advanced reasoning tasks, making high-performance agentic systems accessible to organizations with limited compute resources.
Future Directions: The work highlights the potential of MARL in training scalable agent organizations and suggests that future research should focus on efficient parallel architectures rather than solely increasing model size.

In summary, WIDESEEK-R1 proves that for broad, complex information tasks, a well-orchestrated swarm of small, parallel agents trained via reinforcement learning can outperform or match massive single agents, offering a new, efficient path for scaling LLM capabilities.