AI BriefingCombining Weak Agents to Build a State-of-the-Art Deep Researcher
Key point
Instead of developing a new agent, merging the outputs of existing agents achieved the #1 spot on the Deep Research benchmark.
Details
It took the #1 spot on the Deep Research Bench II (DRB II) benchmark, breaking the existing SOTA (State-of-the-Art) record. This achievement was accomplished not by building a stronger single agent, but through an Ensembling technique that merges the outputs of existing Weak Agents.
The core of Deep Research tasks lies not in writing, but in Information Recall — the ability to find necessary facts without missing them. The research found that reports written by different agents contained different facts, and combining them produced synergy.
The research team demonstrated this performance through the following process.
- Data Collection: Used publicly available report data from agents ranked 7th to 13th on the leaderboard.
- Oracle Experiment: Proved that the 'Merger' approach, which combines multiple reports, achieved a much higher score than the 'Selector' approach, which selects a single optimal report.
- Result: Recorded a TotalScore of 64.38, unreachable by a single agent, surpassing the existing record by 3.2 points.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Qwen Intelligence Launches Three Mobile AI Agents and Open Benchmark Suite