| | --- |
| | |
| | title: TinyThink |
| | sdk: gradio |
| | short_description: Interactive visualization of recursive reasoning concepts |
| | authors: |
| | - raayraay |
| | - samiksha-bc |
| | sdk_version: 6.3.0 |
| |
|
| | --- |
| | # 🧠 TinyThink: Glass-Box Reasoning Visualization |
| | An **interactive educational demo** that visualizes the concept of recursive reasoning from the paper ["Less is More: Recursive Reasoning with Tiny Networks"](https://github.com/SamsungSAILMontreal/TinyRecursiveModels) (Samsung SAIL Montreal, 2025). |
| |
|
| | ## What This Demo Shows |
| |
|
| | This Space provides a visual explanation of how **Tiny Recursive Models (TRM)** approach complex reasoning tasks: |
| |
|
| | - **Step-by-step visualization** of how a model iteratively refines its answer |
| | - - **ARC-AGI style puzzles** rendered with the official 10-color palette |
| | - - **Educational content** explaining the recursive reasoning mechanism |
| | |
| | - ## ⚠️ Important Note |
| | |
| | - **This is a visualization/simulation demo, not actual TRM inference.** The animation demonstrates the *concept* of recursive refinement—showing how a model progressively improves its answer over multiple iterations. |
| | - For the real model implementation, see: |
| | - - 📄 [Paper & Code](https://github.com/SamsungSAILMontreal/TinyRecursiveModels) |
| | - - 🤗 [Model Checkpoints](https://huggingface.co/arcprize/trm_arc_prize_verification) |
| | |
| | - ## The Key Insight |
| | |
| | - The TRM achieves **45% on ARC-AGI-1** with only **7M parameters** by using: |
| | - 1. **Recursive depth over model size** – Multiple "thinking" iterations |
| | 2. 2. **Latent state refinement** – The model updates internal representations z |
| | 3. 3. **Progressive answer improvement** – Each iteration corrects previous errors |
| | |
| | 4. This challenges the "bigger is better" paradigm in AI reasoning. |
| | |
| | 4. ## Try It Out |
| | 3. Select different ARC-AGI puzzles and watch the visualization of how recursive reasoning progressively transforms an initial guess into the correct solution. |
| | - --- |
| | |
| | *Created as an educational tool to make AI research accessible. Inspired by the innovative work of Samsung SAIL Montreal.* |
