Self-Harnessing Recursive Language Models
Submitted to the NeurIPS 2026 Workshop on Meta Agents
Can a frozen recursive language model mine its own failures, rewrite its own harness, and keep working on inputs 8–32× longer than it was optimized on?
• By William Stanford, Mohammed Akram Khan Lodi, Jaloliddin Boymakhammadov, Eliaz Calvar and Simon Coumes