Recoleta
GitHub
Source note

ARIADNE: Agentic Reward-Informed Adaptive Decision Exploration via Blackboard-Driven MCTS for Competitive Program Generation

Authors
Minnan Wei, Xiang Chen, Xiaoshuai Niu, Siyu Chen
Source
arXiv
Published
Collection
Software Intelligence
Code GenerationCompetitive ProgrammingMctsLLM AgentsCode Repair
Open arXivMarkdown source

Summary

ARIADNE is an LLM-based competitive programming system that uses MCTS and a shared blackboard to plan, test, and repair code under a fixed search budget. It reports higher Pass@1 than agentic coding baselines on APPS, CodeContests, CodeContests+, LiveCodeBench, and two recent contest sets.

Problem

  • It addresses unreliable one-shot program generation for contest problems, where a model must choose the right algorithm, satisfy time and memory limits, and handle hidden edge cases.
  • The problem matters because small strategy errors or missed edge cases can make generated code fail all hidden tests, even when the code looks plausible.
  • Prior agent pipelines, MCTS code search, and blackboard coordination each miss part of the needed loop: adaptive planning, persistent failure evidence, and budget allocation.

Approach

  • ARIADNE treats program generation as a sequential decision process over states (current code, blackboard evidence).
  • MCTS chooses among five action families: strategy selection, code generation, test generation, quality evaluation, and code repair.
  • The blackboard stores parsed constraints, candidate strategies, generated tests, counterexamples, diagnostics, and repair notes so later branches can reuse earlier evidence.
  • Evaluation returns both a scalar reward and structured feedback. The reward uses weights of 0.6 for correctness, 0.2 for performance, and 0.2 for code structure.
  • UCB-guided selection and reward backpropagation move the search budget toward branches with better test results and useful repairs.

Results

  • With GPT-4o, ARIADNE reports Pass@1 of 41.30 on APPS, 46.67 on CodeContests, 27.27 on CodeContests+, and 20.91 on LiveCodeBench.
  • The paper says ARIADNE beats the strongest listed baseline, CodeSim, by up to 26.06 Pass@1 points.
  • On the 2025 ICPC Asia Shenyang Regional Contest set, ARIADNE reports pass@1/3/5 of 3/13, 6/13, and 7/13, compared with 1/13, 4/13, and 5/13 for the best benchmark baseline.
  • On the 2025 CCPC Fujian Invitational set, ARIADNE reports pass@1/3/5 of 4/13, 5/13, and 7/13, compared with 2/13, 2/13, and 5/13.
  • The excerpt also claims further gains with DeepSeek-V3.2, but it does not provide exact DeepSeek-V3.2 scores.

Link