🧠 Recursive Language Model (RLM)

A hands-on implementation of a Recursive Language Model (RLM) built from scratch using Python, OpenRouter, and a custom Python REPL. This project demonstrates how Recursive Language Models overcome long-context limitations by generating, executing, and refining Python code through an iterative reasoning loop.

---

📖 About the Project

Traditional Large Language Models (LLMs) struggle with long-context reasoning because all information must fit within the model's context window. Recursive Language Models (RLMs) address this limitation by allowing the model to generate Python code, execute it in a REPL, observe the results, and continue reasoning until the correct answer is found.

In this project, I implemented an end-to-end RLM that stores large contexts outside the prompt, executes model-generated code, supports recursive sub-LLM calls, and iteratively solves long-context reasoning tasks.

---

✨ Features

• 📄 Generate large synthetic datasets for long-context evaluation
• 🐍 Custom Python REPL for code execution
• 💾 Store large context outside the LLM prompt
• ⚡ Execute LLM-generated Python code
• 📤 Capture execution outputs automatically
• 🤖 Recursive Language Model (RLM) Agent Loop
• 🔄 Recursive sub-LLM calls using llm_query()
• 🎯 FINAL() mechanism for returning answers
• 📊 Evaluate long-context reasoning performance
• 🧩 Modular and beginner-friendly implementation

---

📚 What I Learned

By building this project, I gained hands-on experience with:

• 🧠 Understanding the architecture and workflow of Recursive Language Models (RLMs).
• 📖 Understanding why traditional Large Language Models (LLMs) struggle with long-context reasoning.
• 🐍 Building a custom Python REPL (Read-Eval-Print Loop) for executing AI-generated code.
• 💾 Managing large contexts outside the LLM prompt for efficient reasoning.
• ⚙️ Designing an iterative Reason → Execute → Observe agent loop.
• 🤖 Enabling LLMs to generate, execute, and refine Python code.
• 🔄 Implementing recursive sub-LLM calls using llm_query().
• 📤 Capturing execution outputs and using them as feedback for further reasoning.
• 🎯 Implementing the FINAL() mechanism to return validated answers.
• 📊 Evaluating Recursive Language Models on long-context reasoning tasks.
• 🛠️ Building an end-to-end Recursive Language Model from scratch.
• 🚀 Understanding how code execution improves the reasoning capabilities of modern AI agents.

---

🏗️ Project Workflow

                 User Query
                      │
                      ▼
             Recursive Language Model
                      │
          Generates Python Code
                      │
                      ▼
              Execute Inside REPL
                      │
                      ▼
             Capture Execution Output
                      │
                      ▼
          More Reasoning Required?
               │              │
             Yes              No
               │              │
               ▼              ▼
      Recursive Sub-LLM     FINAL(answer)
               │
               └──────────────► Repeat

---

🛠️ Tech Stack

• Python
• Google Colab
• OpenRouter API
• Gemini / OpenAI Compatible Models
• Python REPL
• Regular Expressions (re)
• JSON
• Context Managers (contextlib)

---

📂 Project Structure

Recursive-Language-Model/
│
├── RLM_MODEL.ipynb          # Complete notebook implementation
├── README.md
└── Documatation

---

🎯 Key Concepts Covered

• Recursive Language Models (RLM)
• Long-Context Reasoning
• Read-Eval-Print Loop (REPL)
• Iterative Agent Loops
• Program-Aided Reasoning
• Recursive AI Agents
• Code Execution by LLMs
• Recursive Sub-Agent Calls
• Execution Feedback Loops
• Context Management

---

🚀 Future Improvements

• Secure sandboxed code execution
• Multi-agent collaboration
• Parallel recursive reasoning
• Memory optimization
• Support for external tools and APIs
• Visualization of the reasoning process

---

📚 References

• Recursive Language Models (RLM) Research Paper
• OpenRouter API Documentation
• Python Documentation

---

⭐ Support

If you found this project helpful, consider giving it a ⭐ Star. It motivates me to build and share more AI, LLM, and Agentic AI projects.