The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
Abstract
The abstract outlines a roadmap for recursive self-improvement in AI, from autonomy stages to meta-improvement, across domains like scientific discovery and software engineering, while identifying practical challenges.
Recursive self-improvement (RSI) enables AI systems to turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement. We first use the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, then introduce the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. Next we examine RSI across scenarios (e.g., scientific discovery, embodied intelligence, software engineering), highlighting their distinct requirements and development speeds. Drawing on diverse industry practices and preliminary empirical evidence, we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.
Community
RSI is the capability of an intelligent system to transform acquired experience and feedback into persistent changes to itself across interaction rounds, such that those changes can affect how later improvements are generated, evaluated, selected, and consolidated. The updated object may be model weights, prompts, code, memory, skills, task distributions, or the improvement mechanism itself. Levels capture autonomy over what is changed, how it is changed, and where later learning experience comes from; they are not paper-quality rankings.
The survey introduces the Headroom-Closed Index (HCI), develops the RSI roadmap represented by the L1-L5 taxonomy, and examines RSI in scientific discovery, embodied intelligence, and software engineering. This repository provides the paper-level, auditable companion to that roadmap.
The 491 baseline papers and 28 table-derived extensions in this collection are retained as RSI-related under the L1-L5 taxonomy. The taxonomy intentionally includes bounded forms and precursors:
L1 - Autonomy over Improvement Execution: the system executes a human-defined improvement procedure, and its accepted results persist into later tasks or rounds.
L2 - Autonomy over Improvement Strategies: the system chooses how to improve a specified target, while the objective, evaluation criterion, or acceptance rule remains external.
L3 - Autonomy over Future Learning Experience: the learner's evolving state influences the experience, task, or curriculum acquired next.
L4 - Autonomy in Deployment and Environmental Adaptation: reusable memory, skills, or deployed agent components are retained and alter later behavior within a fixed improvement process.
L5 - From Environmental Adaptation to Meta-Improvement: the system improves the mechanism that produces future improvements, such as search, evaluation, or research-control policies.
Thus, inclusion does not claim that every entry is a fully autonomous or open-ended RSI system. The L1 and L2 labels make their bounded autonomy explicit.
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper