MARS: Multi-Specialist LLM Relay System for Competitive Programming
3.40T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.23918.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryMARS is a prompt-only multi-agent framework for competitive programming in which RAG over an algorithm-theory corpus selects a small team of topic-specialist LLM agents (dynamic programming, graphs, strings, geometry, etc.). A starter writes a C++17 solution; specialists iteratively run it against public examples in a sandbox and keep, repair, or hand off via structured packets. On CodeContests with Gemma 4, it reaches a 0.624 pass rate at 2.3 pipeline stages per task, gaining 14.4 points over direct prompting and approaching CodeSIM at 3.3x lower wall-clock cost.
Why it mattersA measured multi-agent relay pattern with topic-specialist selection and reported cost/accuracy tradeoffs against baselines; the architecture is concrete enough to adapt beyond competitive programming.
Cited by
No citations on record.
