Following an AI improvement into the next round

An agent corrects a failed workflow. The correction becomes a reusable skill. Later, the agent uses that skill to design a better correction. Each step changes a different part of the system, and each deserves its own evaluation. Following that chain gives recursive self-improvement a concrete engineering meaning.

Recursive Self-Improvement in AI: A Survey, co-authored by Weiran Yao, organizes this growing body of work around the changes a system inherits and the effect those changes have on subsequent improvement. Its bibliography contains 404 distinct works.

Start with what survives

A useful account of an improvement loop begins with its persistent state. That state can include model weights, a memory entry, a tool, a workflow, or the program that proposes the next change. Recording the update target makes the experiment easier to inspect: a reader can trace an experience to a candidate modification and then to the artifact carried into the next round.

Persistence also gives teams a practical unit of analysis. A retained skill can be exercised on another task. A revised training mixture can produce another checkpoint. A changed search procedure can propose another experiment. These artifacts connect one round of work to the next.

Measure the improvement procedure

The survey separates task performance, retention of useful capabilities, and gains in the procedure that produces updates. Each level answers a useful question. Task evaluation measures the current system. Retention tests track capabilities across updates. Improver evaluation asks whether an inherited change helps produce better successors.

This framing connects research on generated training data, agent memory, program search, and automated experimentation. It gives researchers a common vocabulary for describing where learning occurs and how experience travels through the system.

Make successive rounds comparable

For an engineering team, the framework suggests an experiment log with explicit update targets, retained artifacts, evaluation sets, and resource budgets. Independent tests and transfer tasks can then follow a candidate through successive rounds. Recording compute and evaluation costs makes the benefit of each retained change easier to assess.

The survey is available as a preprint. Research credit: Hao Wu and collaborators, including Actava.ai co-founder Weiran Yao; Qingsong Wen is the corresponding author.

Source: Recursive Self-Improvement in AI: A Survey.