Conquering Mathematical Olympiad with GPT-OSS-120B: A Competition-Level Solution Using Multi-Round Reasoning and Symbolic Verification

This article analyzes the winning solution for the Kaggle AI Mathematical Olympiad competition, showing how to solve high-difficulty Olympiad-level math problems using the GPT-OSS-120B large model combined with multi-round reasoning, symbolic verification, and an entropy scoring mechanism. This solution provides a reference for LLM reasoning research and mathematical AI system development.

Background: When Large Models Meet Mathematical Olympiad

Mathematical Olympiad problems have always been known for their strict logic and long reasoning chains, posing a great challenge even to human contestants. With the improvement of large language model capabilities, AI mathematical reasoning has become an important benchmark for measuring model intelligence. The AI Mathematical Olympiad – Progress Prize 3 competition hosted by Kaggle requires participating systems to output a non-negative integer between 0 and 99999 as the final answer.

This solution from Dimas Pasha Akrilian successfully addresses the competition challenges, demonstrating a technical approach that combines large model reasoning with symbolic computation and multiple verification.

Core Architecture: Design Philosophy of the Reasoning Pipeline

The core of the system is the AIMO3Solver custom reasoning engine, which adopts a structured multi-round reasoning framework with five stages: problem understanding, strategy exploration, path planning, reasoning execution, and verification. The GPT-OSS-120B is selected as the base model, and a local API interface is built via vLLM.

Multi-Round Reasoning and Voting Mechanism

The solution defaults to 8 independent attempts, and the final answer is determined through a voting mechanism. An entropy scoring mechanism is introduced to select the optimal answer by integrating frequency, reasoning consistency, and certainty; if invalid, it falls back to 0.

Python-Assisted Verification: Dual Guarantee of Symbolic and Numerical Computation

Pure neural networks are prone to arithmetic errors. The system integrates a persistent Jupyter kernel, supporting SymPy symbolic verification and NumPy numerical checks. Prompts guide priority on symbolic derivation, and tools cover multiple areas such as equation solving, modular arithmetic, and polynomial factorization, effectively reducing error rates.

Engineering Implementation Details

The system runs on an NVIDIA H100 GPU. The model weights are approximately 65.28GB, starting the inference server takes about 119 seconds, and preloading weights takes about 128 seconds. 16 persistent Jupyter kernels are initialized to support parallel tool calls.

Key configuration parameters: 8 attempts, 16 worker processes, maximum 128 rounds of dialogue, 65536 context tokens, early stop threshold of 4, batch size of 256—balancing reasoning quality and efficiency.

Implications for AI Reasoning Research

This solution provides several insights: multi-round reasoning is significantly better than single-round; symbolic verification compensates for the lack of precise computation in neural networks; the entropy scoring mechanism provides a quantitative basis for answer selection; structured prompts improve reasoning consistency. These techniques can be extended to fields such as code generation, scientific computing, and logical verification.

Conclusion: Technical Value of the Competition Solution

The AI-Mathematical-Olympiad project demonstrates a technical approach that combines large models with symbolic computation and multiple verification. It proves that in a resource-constrained competition environment, a competition-level mathematical reasoning system can be built through architectural design and engineering optimization. It is of great reference value to developers engaged in LLM reasoning research, mathematical AI development, or tool-enhanced reasoning pipeline construction.