Competitive Programming with Large Reasoning Models

2025/02/03 by OpenAI, :, Ahmed El-Kishky +50 · 13 voices · 33 citations
Decision Sciences · Economics, Econometrics and Finance · #Economic theories and models #Multi-Criteria Decision Making #cs.AI #cs.CL #cs.LG

paper · pdf · doi:10.48550/arxiv.2502.06807

openalex publication_date 2025/02/03 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

We show that reinforcement learning applied to large language models (LLMs) significantly boosts performance on complex coding and reasoning tasks. Additionally, we compare two general-purpose reasoning models - OpenAI o1 and an early checkpoint of o3 - with a domain-specific system, o1-ioi, which uses hand-engineered inference strategies designed for competing in the 2024 International Olympiad in Informatics (IOI). We competed live at IOI 2024 with o1-ioi and, using hand-crafted test-time strategies, placed in the 49th percentile. Under relaxed competition constraints, o1-ioi achieved a gold medal. However, when evaluating later models such as o3, we find that o3 achieves gold without hand-crafted domain-specific strategies or relaxed constraints. Our findings show that although specialized pipelines such as o1-ioi yield solid improvements, the scaled-up, general-purpose o3 model surpasses those results without relying on hand-crafted inference heuristics. Notably, o3 achieves a gold medal at the 2024 IOI and obtains a Codeforces rating on par with elite human competitors. Overall, these results indicate that scaling general-purpose reinforcement learning, rather than relying on domain-specific techniques, offers a robust path toward state-of-the-art AI in reasoning domains, such as competitive programming.

Cited by

• Reasoning Before Translation: Enhancing Legal Machine Translation with Structured Reasoning
• Embarrassingly Simple Self-Distillation Improves Code Generation
• Humans Perceive Wrong Narratives from AI Reasoning Texts
• Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search
• Reinforcement Learning via Self-Distillation
• LA-RL: Label-Aware Self-Reflection for Reinforcement Learning in Information Extraction
• Offline-Online Curriculum RL for Multimodal Reasoning
• AlloBench: Measuring Online Tool Allocation Capability in LLM Agents
• Well Begun, Half Done: Reinforcement Learning with Prefix Optimization for LLM Reasoning
• Stabilizing Off-Policy Training for Long-Horizon LLM Agent via Turn-Level Importance Sampling and Clipping-Triggered Normalization
• Taming the Long-Tail: Efficient Reasoning RL Training with Adaptive Drafter
• Knowledge-Augmented Long-CoT Generation for Complex Biomolecular Reasoning
• Investigating CoT Monitorability in Large Reasoning Models
• HarnessLLM: Automatic Testing Harness Generation via Reinforcement Learning
• Limits of Generalization in RLVR: Two Case Studies in Mathematical Reasoning
• PRISM-Physics: Causal DAG-Based Process Evaluation for Physics Reasoning
• NeSyPr: Neurosymbolic Proceduralization For Efficient Embodied Reasoning
• LexChain: Modeling Legal Reasoning Chains for Chinese Tort Case Analysis
• Seesaw: Accelerating Training by Balancing Learning Rate and Batch Size Scheduling
• UniCode: A Framework for Generating High Quality Competitive Coding Problems
• Breaking the Self-Confirming Loop: Diagnosing and Mitigating Systemic Reward Bias in Self-Rewarding RL
• Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning
• Large Language Models Achieve Gold Medal Performance at the International Olympiad on Astronomy & Astrophysics (IOAA)
• Exploring Chain-of-Thought Reasoning for Steerable Pluralistic Alignment
• Graph-S3: Enhancing Agentic textual Graph Retrieval with Synthetic Stepwise Supervision
• Probing the Critical Point (CritPt) of AI Reasoning: a Frontier Physics Research Benchmark
• AutoLabs: Cognitive Multi-Agent Systems with Self-Correction for Autonomous Chemical Experimentation
• Scaling LLM Test-Time Compute with Mobile NPU on Smartphones
• AECBench: A Hierarchical Benchmark for Knowledge Evaluation of Large Language Models in the AEC Field
• Dream-Coder 7B: An Open Diffusion Language Model for Code
• GitTaskBench: A Benchmark for Code Agents Solving Real-World Tasks Through Code Repository Leveraging
• LegalΔ: Enhancing Legal Reasoning in LLMs via Reinforcement Learning with Chain-of-Thought Guided Information Gain
• SSRL: Self-Search Reinforcement Learning

Discussions

• AI can now compete with the most elite human programmers in competitive coding o3 won gold in the International Olympiad in Informatics beating a specialized version of o1 designed for the contest G [bsky, 80 points, 1 comments]
• Competitive Programming with Large Reasoning Models [hn, 16 points, 1 comments]
• Competitive Programming with Large Reasoning Models [hn, 6 points, 0 comments]
• "o3 achieves a gold medal at the 2024 IOI" arxiv.org/abs/2502.06807 [bsky, 6 points, 1 comments]
• Competitive programming with large language models [hn, 2 points, 0 comments]
• OpenAI: Competitive Programming with Large Reasoning Models [hn, 2 points, 1 comments]
• "o3 achieves a gold medal at the 2024 IOI and obtains Codeforces rating on par with elite human competitors... [S]caling general-purpose reinforcement learning, rather than relying on domain-specific [bsky, 1 points, 0 comments]
• Competitive Programming with Large Reasoning Models [hn, 1 points, 0 comments]
• OpenAI: Competitive Programming with Large Reasoning Models Research exploring how large language models can solve complex algorithmic and competitive programming challenges, potentially advancing AI [bsky, 0 points, 0 comments]
• Reading this paper, you feel the power of RL. RL is great, but not all we need. A blend of RL, inference time scaling, good pretraining, and domain-specific tuning is often the way to go! arxiv.or [bsky, 0 points, 0 comments]
• Paper: arxiv.org/abs/2502.06807 Our article about test-time compute for a deep dive in this topic: huggingface.co/blog/Ksenias... [bsky, 0 points, 0 comments]
• arxiv.org/abs/2502.06807 [bsky, 0 points, 0 comments]
• Codeforcesでレート2724はちょっと強いか。ちょっとだけ arxiv.org/abs/2502.06807 [bsky, 0 points, 0 comments]

Related

• Scaling Test-Time Compute to Achieve IOI Gold Medal with Open-Weight Models
• ProBench: Benchmarking Large Language Models in Competitive Programming
• LiveOIBench: Can Large Language Models Outperform Human Contestants in Informatics Olympiads?
• Scaling Up RL: Unlocking Diverse Reasoning in LLMs via Prolonged Training
• SRPO: A Cross-Domain Implementation of Large-Scale Reinforcement Learning on LLM