Sakana AI's ALE-Agent Achieves Top 2% in AtCoder Heuristic Contest
Sakana AI's ALE-Agent, an AI designed for hard optimization problems, ranked 21st out of 1,000 human participants in a live AtCoder Heuristic Contest, showcasing AI's potential in algorithm engineering.
Sakana AI has developed ALE-Agent, an AI system specialized in solving hard optimization problems, which recently achieved 21st place out of over 1,000 human participants in a live AtCoder Heuristic Contest (AHC). This milestone marks a significant advancement in AI's ability to tackle NP-hard problems with real-world applications, such as logistics and factory planning.
The Challenge of Hard Optimization Problems
Combinatorial optimization problems, like the Traveling Salesperson Problem, are computationally intensive and often lack exact solutions. Human experts spend weeks refining algorithms to approximate optimal solutions. Sakana AI's ALE-Bench benchmark, developed in collaboration with AtCoder, evaluates AI systems on 40 such problems from past AHCs. Unlike traditional coding benchmarks, ALE-Bench focuses on long-horizon reasoning and iterative refinement, mirroring real-world problem-solving.
ALE-Agent's Breakthrough Performance
ALE-Agent, built on Gemini 2.5 Pro, combines domain knowledge with inference-time scaling to generate diverse solutions. In the May 2025 AHC047 contest, it:
- Ranked 21st, outperforming 98% of human competitors.
- Improved solutions 100x faster than humans, leveraging rapid code revisions.
- Applied advanced techniques like Poisson approximation and simulated annealing to boost scores.
Key Insights and Limitations
- Strengths: ALE-Agent excels in short-duration contests and problems suited for simulated annealing.
- Weaknesses: Struggles with longer contests and experimental algorithm design.
- Future Work: Enhancing feedback mechanisms and integrating human expert tools could further improve performance.
Conclusion
Sakana AI's work with ALE-Bench and ALE-Agent demonstrates AI's growing capability in algorithm engineering. The team plans to refine the agent for broader applications, potentially revolutionizing industries reliant on optimization. For details, read the technical paper or explore the ALE-Bench dataset.
Interested in joining Sakana AI? Check their careers page.
Related News
Zscaler CAIO on securing AI agents and blending rule-based with generative models
Claudionor Coelho Jr, Chief AI Officer at Zscaler, discusses AI's rapid evolution, cybersecurity challenges, and combining rule-based reasoning with generative models for enterprise transformation.
Human-AI collaboration boosts customer support satisfaction
AI enhances customer support when used as a tool for human agents, acting as a sixth sense or angel on the shoulder, according to Verizon Business study.
About the Author

David Chen
AI Startup Analyst
Senior analyst focusing on AI startup ecosystem with 11 years of venture capital and startup analysis experience. Former member of Sequoia Capital AI investment team, now independent analyst writing AI startup and investment analysis articles for Forbes, Harvard Business Review and other publications.