MARS System Boosts LLM Competitive Programming Performance

Large Language Models often struggle with the nuanced challenges of competitive programming, particularly when existing multi-agent systems rely on generic roles like planner or coder. MARS introduces a novel approach by creating a relay system of highly specialized LLM agents. Each agent focuses on a specific algorithmic domain, such as dynamic programming or graph theory, and is enhanced by retrieval-augmented generation from an algorithm theory corpus. When presented with a problem, MARS selects a relevant team of specialists. A "starter" agent provides an initial solution, which is then iteratively refined by the active specialist. The solution is tested in a sandbox, and specialists can repair, keep, or hand off the draft to another specialist. This structured, collaborative approach significantly boosts the pass rate on competitive programming benchmarks, closing the gap with more expensive methods while being more cost-effective.