Survey Maps AI Agent Self-Evolution From One-Shot Learning to Recursive Improvement
A comprehensive survey covering 2023 to 2026 examines how AI agents are evolving beyond static, one-shot execution into systems capable of continuous self-improvement. Researchers identify three core dimensions of agent evolution: what to evolve, when to evolve, and how to evolve, spanning model weights, memory, tools, and architecture. Notable systems reviewed include Voyager, SWE-RL, MUSE, and DGM, each demonstrating distinct approaches to turning execution outcomes into learning signals. Meta's SWE-RL, released in late 2025, achieved a 10.4 percentage point gain on the SWE-bench Verified benchmark using only self-play — with no external training data. The survey marks 2025–2026 as a turning point where recursive self-improvement shifted from theoretical concept to empirical validation.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in