Scaling Test-Time Compute to Achieve IOI Gold Medal with Open-Weight Models

2025-10-15 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

NVIDIA researchers introduce GenCluster, a scalable test-time compute framework that enables open-weight large language models (LLMs) to achieve gold medal-level performance at the International Olympiad in Informatics (IOI) 2025. This framework addresses the challenge of matching proprietary models' performance with transparent, reproducible methods. GenCluster integrates large-scale solution generation, behavioral clustering, LLM-based ranking via a tournament, and a round-robin submission strategy to navigate IOI's strict validation budgets and submission limits. Experiments demonstrate that the gpt-oss-120b model, when combined with GenCluster and 5000 generations per subtask, achieved a gold medal score of 446.75, marking the first time an open-weight model has reached this level. The approach shows consistent performance scaling with increased compute, narrowing the gap between open and closed AI systems in competitive programming.

Key takeaway

For research scientists developing competitive programming LLMs, GenCluster offers a transparent and reproducible method to achieve top-tier performance with open-weight models. You should consider implementing its four-stage pipeline—parallel generation, behavioral clustering, tournament-based ranking, and round-robin submission—to maximize scores under strict competition constraints. This approach demonstrates that scaling test-time compute is crucial for bridging the performance gap between open and proprietary systems, providing a clear path to gold-level achievements.

Key insights

GenCluster enables open-weight LLMs to achieve IOI gold medal performance through scalable test-time compute and strategic solution selection.

Principles

• Test-time compute scales LLM performance.
• Behavioral clustering improves solution selection.
• LLM-as-a-judge can rank code solutions.

Method

GenCluster generates many candidate solutions, clusters them by behavioral similarity, ranks clusters using an LLM-based tournament, and employs a round-robin submission strategy under IOI constraints.

In practice

• Generate 5000+ solutions per subtask for optimal results.
• Use C++ for competitive programming solutions.
• Employ longest reasoning trace as a correctness proxy.

Topics

• Competitive Programming
• Large Language Models
• Test-Time Compute
• GenCluster
• IOI Gold Medal

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Winning a Kaggle Competition with Generative AI–Assisted Coding — LLM agents, paired with GPU acceleration, dramatically speed up ML experimentation by automating code generation and execution.
• ThinkBooster: A Unified Framework for Seamless Test-Time Scaling of LLM Reasoning — ThinkBooster unifies LLM test-time compute scaling with a modular framework, benchmark, and deployable proxy for improved reasoning.
• EGG: An Expert-Guided Agent Framework for Kernel Generation — EGG guides LLM-based kernel generation with expert principles, decomposing the process for superior performance and correctness.
• Adaptive Test-Time Compute Allocation for Reasoning LLMs via Constrained Policy Optimization — Optimizing LLM test-time compute involves balancing accuracy and cost via a constrained policy.
• 3x Faster Search: Parallel Test-Time Scaling with Instructed-Retriever-1 — Parallel test-time scaling with a specialized model significantly boosts retrieval speed and quality by executing search stages concurrently.
• Inference Scaling (Test-Time Compute): Why Reasoning Models Raise Your Compute Bill — Inference scaling dynamically allocates compute for reasoning, but requires strategic management to avoid excessive costs and performance issues.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.