Skip to content

Colombo: a frontier-to-frontier automated research system for oracle-based optimization

Colombo: A Frontier-to-Frontier Autoresearch System for Oracle-Based Optimization

arXiv· arXiv:2611.23080· Haihan Zhang, Wendao Wu, Chenheng Zhang, Chunyuan Zheng, Yanyi Li, Cong Fang, Haoxuan Li, Zhouchen Lin·Indexed by this site · 2 days ago· Originally publishedAI updates· Its place in the scan record →
AI summary

The research team proposes Colombo, a self-organizing AI researcher system, first instantiated in oracle-based optimization, with the core questions being how efficiently a problem class can be solved given an information interface, and what cost any admissible algorithm cannot avoid. The system maintains the Complexity Atlas, a dynamic map recording frontier settings and the current state of knowledge, in which the gap between achievable guarantees and unavoidable limits defines the frontier; the self-organizing mechanism turns unsolved gaps into adaptive research plans, coordinating conjecture formation, mathematical research and replanning as evidence accumulates. Adversarial Co-Ascent simultaneously develops algorithms and hard problem families by having them challenge each other, revealing information use and hidden mechanisms; mathematical claims are checked by independent review and Lean-based formal certification, and long-term memory preserves supported findings and research experience across settings. The authors state that Colombo has established optimal or near-optimal rates on more than fifty settings in the Atlas, and call this a concrete step toward Autopoietic Science.

Read originalarxiv.org