MARS: Multi-Specialist LLM Relay System for Competitive Programming
MARS routes contest problems through algorithm specialists instead of generic coding agents.
The system uses retrieval to pick specialists for areas like dynamic programming, graphs, strings, or geometry, then passes a C++17 draft through them in sequence. Each stage tests against public examples in a sandbox and can keep, repair, or hand off the solution. On CodeContests with Gemma 4, the paper reports a 0.624 ± 0.006 pass rate, up 14.4 points over direct prompting, while using fewer stages and lower wall-clock cost than CodeSIM. HF Daily Papers' note
The system uses retrieval to pick specialists for areas like dynamic programming, graphs, strings, or geometry, then passes a C++17 draft through them in sequence. Each stage tests against public examples in a sandbox and can keep, repair, or hand off the solution. On CodeContests with Gemma 4, the paper reports a 0.624 ± 0.006 pass rate, up 14.4 points over direct prompting, while using fewer stages and lower wall-clock cost than CodeSIM. HF Daily Papers' note
score 4