Comparison of hand-engineered inference strategies versus learned model-generated strategies

Determine how domain-specific, hand-engineered inference strategies for competitive programming (exemplified by the o1-ioi test-time pipeline that partitions IOI problems into subtasks, samples large candidate sets, clusters outputs on model-generated test inputs, and reranks submissions) compare to learned test-time reasoning strategies that are autonomously generated and executed by large reasoning models trained end-to-end via reinforcement learning (such as OpenAI o1 and o3).

Background

The paper examines performance of large reasoning models (LRMs) on competitive programming, contrasting specialized pipelines with general-purpose reinforcement learning approaches. OpenAI o1-ioi incorporates domain-specific, hand-engineered test-time strategies—such as subtask-based sampling, clustering using model-generated tests, and reranking—to compete in IOI 2024.

In contrast, the later o3 model relies on end-to-end RL to develop its own test-time reasoning behaviors (e.g., generating brute-force validators), avoiding manual inference heuristics. The authors explicitly raise the question of how these two paradigms compare, motivating their evaluations across CodeForces and IOI 2024.