AI is starting to research how to self-evolve, Tencent draws a 5-level evolution roadmap

According to Dynasty Beating monitoring, how to let AI improve itself without relying on manual tuning by humans repeatedly, but rather by summarizing its own experience, modifying itself, and getting stronger with more usage, is becoming an important focus of Agent research. Tencent MIIT, together with Zhejiang University and other teams, reviewed 549 related studies, categorizing various aspects of "self-reflection, self-training, memory evolution, skill evolution, and recursive improvement" into five levels from L0 to L4.

L0: Modify Answers. Reflect, retry, and change approaches when making mistakes, but without leaving long-term changes after the task ends.
L1: Modify Model. Incorporate learned experiences into model parameters for future tasks.
L2: Modify Agent. Begin to adjust prompts, memory, skills, tools, workflows, and harness outside the model.
L3: Modify "Evolution Method." Even how to learn in the future, propose updates, choose updates, and rollback updates can be modified. The paper sees this as the starting point of recursive self-improvement.
L4: Even "what is considered progress" can be modified. Evaluation criteria, rewards, constraints, and testing methods also fall under the scope of self-modification.

As we progress to higher levels, we get closer to true "AI upgrading AI," but the problems become more challenging: it may not actually be getting stronger but simply getting better at taking its own exam. If AI can both modify itself and change the grading rules, it could ultimately prove its "progress" solely by altering the exam content.

Therefore, the paper presents a key principle: AI can evolve on its own but cannot have ultimate approval authority. Tests, evidence, and release rules used to prove its improvement must be kept beyond its modification reach. Rejections or rollbacks for inadequate updates should be carried out, and if necessary, handled by humans.