SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research
The paper introduces SearchSwarm, a framework that synthesizes supervised fine-tuning data via a harness-guided process to train models with "delegation intelligence," enabling them to effectively decompose and manage long-horizon deep research tasks through agent subtask delegation, resulting in state-of-the-art performance on BrowseComp benchmarks.
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 Problem: The "Short-Term Memory" Bottleneck
Imagine you are a brilliant detective (the AI) trying to solve a massive, complex mystery that involves reading thousands of documents, visiting hundreds of websites, and connecting dots across years of history.
The problem? You have a very strict rule: You can only hold a few pages of notes in your hand at once.
In AI terms, this is the "context window." Even the smartest AI models have a limit on how much information they can remember in a single conversation. If a research task gets too long, the AI forgets the beginning of the investigation by the time it reaches the end.
Traditionally, when an AI hits this limit, it tries to "summarize" its notes. But this is like trying to remember a 10-hour movie by only remembering the plot summary written on a napkin. You lose the details, the evidence, and the nuance.
The Solution: The "SearchSwarm" Strategy
The authors of this paper realized that instead of trying to remember everything yourself, you should hire a team of assistants to do the heavy lifting.
They call this "Delegation Intelligence."
Think of the main AI as a Project Manager and the sub-AIs as Specialized Researchers.
- The Project Manager (Main Agent): This is the smart brain. It doesn't do the grunt work of reading 500 websites. Instead, it looks at the big question, breaks it down into smaller pieces, and says, "You, go find out about X. You, go find out about Y."
- The Researchers (Sub-Agents): These are copies of the same AI, but they work in their own separate "rooms" (independent contexts). They can read as many pages as they need without worrying about the Project Manager's memory limit.
- The Handoff: When a Researcher is done, they don't dump a 50-page report on the Manager. They write a one-page summary with specific citations (proof of where they found the info). The Manager reads the summary, checks the proof, and moves on.
The Secret Sauce: How They Taught the AI to Delegate
You might ask, "Can't the AI just figure this out on its own?"
The paper says no. Most AI models are trained on books and articles, not on scripts showing how to run a team. They don't naturally know how to ask for help, what to ask for, or how to check the work.
To fix this, the researchers built a "Training Harness" (a strict set of rules and instructions) to teach the AI how to be a good manager.
Here are the four golden rules they taught the AI:
- Rule 1: Don't do the grunt work yourself. If a task requires reading a lot of text, send it to a sub-agent. Save your own brainpower for connecting the dots.
- Rule 2: Give a great briefing. You can't just tell a sub-agent "Go find info." You have to say, "We already know X and Y, but we are stuck on Z. Go find Z, and here is why it matters." This prevents the sub-agent from wasting time re-discovering things the Manager already knows.
- Rule 3: Keep the final judgment. The Manager must never blindly trust the sub-agent. The Manager must verify the evidence. If a sub-agent says "The answer is 42," the Manager must check the citation to make sure it's not a hallucination (a lie).
- Rule 4: Always cite your sources. Every conclusion must come with a "receipt" (a link to the source). This ensures the final answer is trustworthy.
The Results: A Small Team Beating Giants
The researchers trained a model called SearchSwarm-30B (a model with 30 billion parameters, which is considered "medium-sized" in the AI world).
They tested it on difficult research benchmarks (like BrowseComp and GAIA), which are essentially "hard trivia" tests that require deep digging.
The Shocking Result:
SearchSwarm, a medium-sized model, beat almost every other model of its size. Even more impressively, it performed as well as, or better than, some of the giant, closed-source models that are 10 times larger and cost millions to run.
- Analogy: It's like a small, well-organized local library (SearchSwarm) beating a massive, chaotic national archive (the giant models) at finding specific, obscure facts, simply because the local librarian knows exactly how to delegate tasks to their assistants.
What This Means for the Future
The paper doesn't just show a cool trick; it provides a recipe. They released the code, the training data, and the model weights for everyone to use.
They proved that intelligence isn't just about having a bigger brain; it's about having better management skills. By teaching an AI how to delegate, break down tasks, and verify evidence, a smaller, cheaper model can solve complex, long-term problems that used to require massive, expensive super-computers.
In short: SearchSwarm teaches AI to stop trying to remember everything and start learning how to run a team.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.