AIM

AI Mathehatician (Working In Process)

AIM is an agentic system targed at frontier mathematical research. This version of AIM is equipped with a multistep explore mechanism and a self-verification mechanism. It can further leverage the strong capability of large reasoning models, and provide valuable assistance for human mathematicians. Note that AIM is specifically designed for complex research-level problems, and is not suitable for exercises

Usage

Currently you can run AIM at your local machine with the following steps.

Obtain AIM Executable

This implementation of AIM is written in Rust language for better performance. You can directly download the executable in release page of this repo (if exists). Otherwise you'll need to build AIM by your self. Firstly make sure you have correctly setup the Rust toolchain on your local machine, after that you can compile AIM with the following commands:

git clone https://github.com/Carlos-Mero/AIM.git
cd AIM
cargo build --release

In this way you can find the standalone executable at path ./target/release/aim. You can also use the following command to compile and install aim at your system path, so that you can use it more conveniently.

cargo install --path .

Setup API Configs

AIM uses dotenv to configure API-keys for this agentic system. You need to create a file named .env at the project path or the parent path of that you want to run aim. Its contents should be as follows:

OPENAI_API_KEY=sk-xxxxxx
# Optional
OPENAI_API_BASEURL=https://api.openai.com
# Role & Invitation Codes (optional)
# Specify an admin account that bypasses project limits:
[email protected]
# Specify the invitation code for "invited" users (max 7 new projects per day):
AIM_INV_CODE=your-invite-code

After this you will be able to run aim in your project without environment errors.

Project Setup

Currently aim only supports locally running one session in each process. A project (or session) for AIM is a directory containing the target problem and relevant context of it. This directory should contain a file named problem.md which contains the target problem, and an optional file named context.md to provide background information of this problem if needed. More accurately, one project structure may look like follows:

 project
├──  .env
├──  context.md
└──  problem.md

Both the problem and context should be written in markdown format with LaTeX extension. It is better not to use headers in both files. After you have correctly setup and entered this project directory, you can start default aim pipeline with the following command:

aim -p .

You can also resume from an existing unsolved project by additionally passing --resume as a argument. You can run aim --help to obtain the usage of other commandline arguments.

Interactive Mode & Server Mode

AIM supports both a single-session CLI workflow and a long-running server mode for integration with a web frontend.

1. Server Mode

   aim --server

   Starts an HTTP API on port 4000 by default. Endpoints allow you to submit problems, stream progress, and fetch results.

2. Web Frontend A Next.js frontend is provided under ./frontend. To launch:

   cd frontend
   npm install
   echo "NEXT_PUBLIC_API_BASE_URL=http://localhost:4000" > .env.local
   npm run dev

   Open http://localhost:3000 to start new sessions, visualize theorem graphs, explore proofs, and view logs. You can also run npm run build under the frontend folder. This will build up static frontend files of this project. The aim executable already equipped with static file serving capability. Then you can directly visiting localhost:4000 when aim --server is running.

Deer-Flow Integration

We also include support using deer-flow to research for context of the given problem. With this feature we can simply start a research session with a given problem. deer-flow will first search for previous papers to obtain existing findings and methods around the given problem.

To set up this feature you should firstly make sure you have installed uv as the python virtual environment manager in your system. After that, please run these commands to initialize deer-flow under the root folder of AIM:

git submodule init
git submodule update

After that you can enter deer-flow folder and follow the steps of official README.md to set up its workflow, until you can successfully run uv run main.py and obtain the desired output. Then you can simply run aim without providing the context and it will automatically search for that for you via deer-flow.

Command-Line Options

Run aim --help for full details. Common flags include:

Session Persistence & Logs

• Intermediate state and memory are stored in aim.db in your project folder—use --resume to continue an interrupted run.
• Step-by-step logs are written under logs/ within the session directory.

Contributing

Contributions are welcome! Please open issues or pull requests on GitHub.