Reproduce “Amplifying human performance in combinatorial competitive programming” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Petar Veličković, Alex Vitvitskyi, Larisa Markeeva, Borja Ibarz, Lars Buesing, Matej Balog, Alexander Novikov
The paper applies FunSearch to improve scoring functions within human-designed greedy algorithms for combinatorial programming contests. Humans provide the backbone, while FunSearch proposes and evaluates scoring-rule changes. Across eight Hash Code qualification rounds, evolved scorers reached the top percentile in every round and beat the highest-scoring human team in five, showing gains on these backbones, not autonomous solution design.
Key insights
- Human-AI synergy, where humans provide the structural understanding and AI automates the discovery of complex scoring functions, is a highly effective paradigm for computationally intractable problems.
- Combinatorial competitive programming serves as a robust and distinct testbed for AI, rewarding good-enough solutions and allowing AI to excel where optimal solutions are elusive.
- The AI-evolved heuristics are often highly intricate and incorporate non-obvious strategies (e.g., hard thresholds, logarithmic scaling, preference adjustments) that can surpass human intuition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.