The current frontier of AI research is heavily focused on "depth scaling"—the pursuit of making single, massive models increasingly intelligent to handle complex, multi-step reasoning. However, when tasks require searching through vast, broad pools of information, these singular giants often struggle with organizational efficiency. WideSeek-R1 introduces a compelling alternative called "width scaling," which distributes the workload across a collaborative team of smaller AI agents rather than relying on a single, monolithic brain.

At the heart of this system is a lead-agent-subagent architecture trained through a specialized process called multi-agent reinforcement learning. This method allows a central "lead" agent to act as a manager, delegating specific parts of a query to multiple "subagents" that work in parallel. By isolating contexts and utilizing specialized tools, these agents can process information simultaneously. Remarkably, a WideSeek-R1 configuration using only 4 billion parameters achieved an F1 score of 40% on the WideSearch benchmark, rivaling the performance of the gargantuan 671 billion parameter DeepSeek-R1.

Perhaps most exciting is the system's inherent scalability; performance improves consistently as more subagents are added to the mix, all without requiring additional training. This shift from individual model size to collective "organizational capability" suggests a future where swarms of smaller, efficient models could outperform today’s most expensive AI titans. It marks a significant step toward more agile and resource-efficient information-seeking systems.

The current frontier of AI research is heavily focused on "depth scaling"—the pursuit of making single, massive models increasingly intelligent to handle complex, multi-step reasoning. However, when tasks require searching through vast, broad pools of information, these singular giants often struggle with organizational efficiency. WideSeek-R1 introduces a compelling alternative called "width scaling," which distributes the workload across a collaborative team of smaller AI agents rather than relying on a single, monolithic brain.

At the heart of this system is a lead-agent-subagent architecture trained through a specialized process called multi-agent reinforcement learning. This method allows a central "lead" agent to act as a manager, delegating specific parts of a query to multiple "subagents" that work in parallel. By isolating contexts and utilizing specialized tools, these agents can process information simultaneously. Remarkably, a WideSeek-R1 configuration using only 4 billion parameters achieved an F1 score of 40% on the WideSearch benchmark, rivaling the performance of the gargantuan 671 billion parameter DeepSeek-R1.

Perhaps most exciting is the system's inherent scalability; performance improves consistently as more subagents are added to the mix, all without requiring additional training. This shift from individual model size to collective "organizational capability" suggests a future where swarms of smaller, efficient models could outperform today’s most expensive AI titans. It marks a significant step toward more agile and resource-efficient information-seeking systems.