Team Infinity | Leaderboard Rank: 276 / 1393
Contributors: Arunesh Kumar Lal (aklal@bu.edu) and Sri Harsha Kotamraju (shk8@bu.edu)
This project was our submission for the Artificial Intelligence Mathematical Olympiad (AIMO) hosted on Kaggle.
The goal was to solve national-level math problems using AI models by interpreting LaTeX-formatted questions across algebra, number theory, geometry, combinatorics, and more.
We developed advanced AI workflows combining:
- Fine-tuned LLMs (Large Language Models)
- Self-Consistency Chain of Thought (SC-CoT)
- Symbolic Code Reasoning and Validation
Our best model could solve 21 problems successfully, achieving rank 276/1393.
- Model: MMOS-DeepSeekMath-7B (Transformer-based, 7B parameters)
- Self-Consistency Sampling: Multiple solutions generated and majority-voted
- Python Code Reasoning: LLM-generated Python code executed and verified using SymPy
- Dynamic Prompt Engineering: Automatic LaTeX parsing, contextualized prompt generation
- Zero-shot and Few-shot Approaches: No fine-tuning, dynamic temperature and top-p control
- Memory Optimizations: 4-bit quantization, multi-GPU inference
- Containerized Deployment: Docker-based setup for reproducibility
| Folder/File | Description |
|---|---|
notebooks/ |
Jupyter Notebooks for model runs, prompt engineering, code execution |
configs/ |
Model and environment setup configurations |
scripts/ |
Utilities for LaTeX parsing, dynamic prompt creation, answer validation |
assets/AIMO_Submissions.jpg |
Screenshot of final Kaggle submissions |
assets/AIMO.png |
Project logo/banner |
AIMIO Project.pdf |
Full technical report explaining methods, challenges, and results |
| Experiment | Description | Problems Solved | Notes |
|---|---|---|---|
| Final Submission | Self-consistency + code generation with DeepSeekMath-7B | 21 | Main final submission |
| Dual-VLLM Inference | Two model instances on separate GPUs | 20 | Parallel inference speedup |
| Zero-Shot Self-Consistency | Basic self-consistency on zero-shot | 19 | Baseline |
| Challenge | Solution |
|---|---|
| High GPU memory requirements | 4-bit quantization + dynamic memory clearing |
| Code execution errors | Built custom subprocess execution wrapper with error handling |
| Managing notebook execution time | Timeout constraints and dynamic control of self-consistency sampling |
| Arithmetic hallucinations | Validated solutions using symbolic computation libraries like SymPy |
- DeepSeekMath GitHub
- SC-CoT Paper
- Parameter Efficient Fine-Tuning (LoRA, QLoRA)
- Kaggle AIMO Competition
"Pushing the limits of AI in mathematical reasoning — one problem at a time."