Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
What this work does
Evolves agent source code through an archive of candidate agents, selecting self-modifications using downstream coding performance.
Connection to discovery
Makes harness/code improvement a distinct object from model-weight training and preserves stepping stones beyond the current best agent.
Evidence
SWE-bench and Polyglot comparisons include archive and self-improvement ablations as well as transfer across underlying models and languages.
Scope and limitations
Empirical selection replaces the original Gödel-machine proof requirement. Repeated benchmark selection and changing execution budgets need auditing; this is not a demonstrated unlimited self-improvement process.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.