FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based Agents
FS-Researcher is a file-system-based dual-agent framework that overcomes context window limitations in long-horizon research tasks by using a persistent external memory to enable effective test-time scaling, achieving state-of-the-art report quality across different backbone models.
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
Imagine you are a brilliant detective hired to solve a massive, complex mystery. You have a super-smart assistant (the AI) who can read books and talk to people, but there's a catch: your assistant has a very short-term memory.
If the mystery is small, your assistant can remember everything. But if the case involves reading 500 websites, interviewing 50 experts, and writing a 100-page report, your assistant's brain gets full. They start forgetting what they read five minutes ago, or they get overwhelmed and give up. This is the problem with current AI "researchers." They run out of "mental space" (context window) before they can finish the job.
Enter "FS-Researcher": The Detective with an Infinite Filing Cabinet.
This paper introduces a new way to help AI do deep research. Instead of trying to force the AI to remember everything in its head, the researchers gave it a digital filing cabinet (a file system) that never runs out of space.
Here is how it works, broken down into a simple story:
The Two-Act Play
The system uses two different AI "characters" who work together, passing a physical folder back and forth.
Act 1: The Librarian (The "Context Builder")
Imagine a super-organized Librarian. Their only job is to go out into the chaotic internet, find clues, read articles, and take notes.
- The Old Way: The Librarian would try to memorize every fact and shout them all out at once. Eventually, they'd get confused and forget the beginning of the story.
- The FS-Researcher Way: The Librarian doesn't memorize anything. Instead, they open a filing cabinet.
- They find a document about "Insurance Company A." They read it, summarize the key points, and file it in a folder labeled "Company A."
- They find a document about "Global Trends." They summarize it and file it in a "Trends" folder.
- They keep doing this for hours, days, or weeks. The cabinet grows huge, but the Librarian's brain stays fresh because they only need to remember what's in the current folder they are looking at.
- They also keep a "To-Do List" on the desk to track what they've done and what's missing.
Act 2: The Writer (The "Report Writer")
Once the Librarian says, "The cabinet is full and organized," the Writer steps in.
- The Writer never looks at the internet. They only look at the filing cabinet.
- They need to write a chapter on "Financial Strength." They open the "Financial Strength" folder, read the specific notes the Librarian wrote, and draft that section.
- Then they close that folder, open the "Growth" folder, and write the next section.
- Because they are only looking at one small folder at a time, they never get overwhelmed. They can write a perfect, long report by piecing together the notes from the cabinet, one section at a time.
Why This is a Game-Changer
1. The "Infinite" Memory
Current AI is like a person trying to hold a whole library in their hands. If the library is too big, they drop books. FS-Researcher is like a person with a giant warehouse. They can store as much information as they want, and they can walk back to the warehouse to grab a specific book whenever they need it.
2. No More "Forgetting"
In old systems, if the AI had to read 100 pages, it might forget the first 50 by the time it got to page 100. With the filing cabinet, the AI can go back, re-read the notes from page 1, and make sure the final report is perfectly consistent.
3. The "Test-Time Scaling" Secret
The paper discovered something cool: The more time you let the Librarian work, the better the final report.
- If you let the Librarian spend 3 hours filling the cabinet, the report is good.
- If you let them spend 10 hours, the cabinet is fuller, the notes are more detailed, and the Writer produces a masterpiece.
- This proves that you can make AI smarter not just by building a "smarter brain," but by giving it more time and a better workspace to do the work.
The Result
The researchers tested this on hard questions (like "Compare the top 10 insurance companies and predict who will win in China").
- Old AI: Often gave up, missed key facts, or wrote shallow reports.
- FS-Researcher: Produced reports that were deeper, more accurate, and better cited than almost any other system, even those using the most expensive AI models.
In a Nutshell
Think of FS-Researcher as giving the AI a permanent notebook and a library card. Instead of trying to hold the whole world in its head, it builds a personal library, organizes it perfectly, and then writes its report using that library. It turns a "short-term memory" problem into a "long-term organization" solution.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.