CP-Agent: A Calibrated Risk-Controlled Agent for Feedback-Driven Competitive Programming
Researchers have developed CP-Agent, a new system designed to improve the performance of large language models in competitive programming tasks. The agent utilizes a calibrated stopped process model to effectively incorporate execution feedback, focusing on reducing false admissions and increasing evidence against incorrect programs. By implementing mechanisms like Dual-Granularity Verification and Test Augmentation, CP-Agent significantly boosts success rates on benchmarks like LiveCodeBench Pro and ICPC-Eval without requiring model parameter updates. AI
IMPACT Enhances LLM capabilities in complex problem-solving, potentially improving agent performance in specialized domains.