Recoleta
GitHub
Source note

MAS-Algorithm: A Workflow for Solving Algorithmic Programming Problems with a Multi-Agent System

Authors
Yuliang Xu, Xiang Xu, Yao Wan, Hu Wei, Tong Jia
Source
arXiv
Published
Collection
Software Intelligence
Multi Agent SystemsCode GenerationAlgorithmic ReasoningCompetitive ProgrammingSoftware Agents
Open arXivMarkdown source

Summary

MAS-Algorithm is a five-agent workflow for competitive-programming problems. It improves acceptance rates for Qwen coding models by adding algorithm selection, retrieval, planning, coding, judging, and error feedback around the same base model.

Problem

  • Algorithmic programming tasks test whether coding models can choose algorithms, reason about constraints, and produce correct efficient code.
  • Direct prompting gives weak control over intermediate reasoning, so errors in algorithm choice, complexity, edge cases, or implementation are hard to isolate.
  • Fine-tuning can be expensive and gave small gains in this study, which makes inference-time coordination useful for code intelligence systems.

Approach

  • Agent1 selects likely algorithms and data structures from candidate labels, with support for multiple solution branches.
  • Agent2 retrieves algorithm knowledge from a local OI-WIKI store using bge-zh embeddings and summarizes it for the solver.
  • Agent3 writes a structured solution plan, Agent4 turns the plan into C++ code, and Agent5 diagnoses failures.
  • A judging tool compiles and runs code in a Docker gcc environment, compares outputs, and sends failed cases back to Agent5.
  • Agent5 returns PASS, FIX, or RETHINK, routing the workflow back to code revision or solution replanning until success or an iteration limit.

Results

  • On the self-built dataset, MAS-Algorithm improved average AC rate across five Qwen models by 6.48 percentage points over direct prompting, with gains from +4.39 to +9.00 points.
  • Qwen3-14B rose from 28.98% AC to 37.98% AC, a +9.00 point gain; its case pass rate rose from 29.34% to 40.14%, a +10.80 point gain.
  • Qwen3-Coder-30B-A3B-Instruct rose from 32.25% AC to 38.87% AC, a +6.62 point gain. LoRA fine-tuning on accepted solutions improved direct prompting only to 33.14% AC, a +0.89 point gain.
  • Qwen3-235B-A22B-Instruct-2507 rose from 62.87% AC to 67.26% AC, a +4.39 point gain.
  • On LiveCodeBench-Pro, Qwen3-Coder-30B-A3B-Instruct improved from 5.28% AC to 10.00% AC, a +4.72 point gain; case pass rate rose from 9.07% to 14.05%.
  • Replacement studies reported large upper-bound gains: improving Agent3 reached 66.57% AC, +27.70 points over the base MAS-Algorithm setting in that experiment; improving Agent2 reached 61.82% AC, +22.95 points.

Link