AIエージェントが最適化プログラミングコンテストで初優勝

2025-12-23 · Source: Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Intermediate, medium

Summary

Sakana AI's "ALE-Agent" won the AtCoder Heuristic Contest 058 on December 14, 2025, outperforming 804 participants in a 4-hour optimization programming challenge. This marks the first known instance of an AI winning such a contest in real-time, demonstrating AI's capability to rival top human experts in complex, multi-hour tasks. The contest problem involved optimizing hierarchical production planning, a scenario common in supply chains. ALE-Agent's solution, while structurally similar to human approaches (greedy algorithm + simulated annealing), leveraged extensive trial-and-error, unique greedy heuristics like "virtual power," and diverse "giant neighborhood" operations for simulated annealing. It also featured highly optimized execution efficiency, including pre-computation tables. The agent utilized 2654 GPT-5.2 and 2119 Gemini 3 Pro Preview calls, costing approximately $1300, to generate and refine its solution.

Key takeaway

For AI scientists developing optimization agents, this achievement highlights the power of combining scaled LLM inference with self-learning mechanisms. You should focus on designing agents that can perform extensive, diverse trial-and-error and extract actionable insights from execution results. Consider integrating novel heuristics and varied neighborhood operations to push beyond conventional problem-solving boundaries, even if it incurs higher computational costs.

Key insights

AI agents can now win complex, real-time optimization contests by scaling LLM inference and self-learning.

Principles

• Extensive trial-and-error improves AI optimization.
• Diverse neighborhood operations enhance local search.
• Self-learning mechanisms refine AI problem-solving.

Method

ALE-Agent uses multiple LLMs to generate programs in parallel, synthesizes insights from trial-and-error, and applies these insights for iterative program generation and refinement.

In practice

• Implement "virtual power" heuristics for non-operational assets.
• Employ "giant neighborhood" operations in simulated annealing.
• Optimize execution with pre-computation and constant-factor tweaks.

Topics

• AI Agents
• Heuristic Optimization
• Competitive Programming
• Large Language Models
• Simulated Annealing

Best for: AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• Step-by-Step Optimization-like Reasoning in LLMs over Expanding Search Spaces — OPT* tasks enable scalable training and evaluation of LLM optimization reasoning by expanding search spaces without new labels.
• Scaling up multi-agent systems: an interview with Minghong Geng — Hierarchical and hybrid designs are key to scaling multi-agent systems for complex, long-horizon problems.
• Large Language Model-Driven Full-Component Evolution of Adaptive Large Neighborhood Search — LLM-driven evolution can automate and enhance complex metaheuristic algorithm design by decoupling and optimizing core components.
• Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning — LLM agents can perform interactive discovery but struggle with scalability and efficiency compared to classic automata learning.
• LLM-Evolved Domain-Independent Heuristics for Symbolic AI Planning — LLMs can evolve domain-independent heuristics for symbolic AI planning that outperform human-designed ones.
• Sakana AI Agent Wins AtCoder Heuristic Contest (First AI to Place 1st) — AI agents can now win complex optimization contests by autonomously discovering novel algorithms and scaling inference.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Blog.