The Mechanisms, Recent Research, and Industrial Deployment Paths of Recursive Self-Improvement (RSI)

Recently I listened to an interview with Yuandong Tian about RSI, and then went on to read related research such as the Darwin Gödel Machine and AlphaEvolve, and I feel this direction is worth discussing. Over the past few years, what we have mainly focused on is what tasks large models can accomplish, from text generation and logical reasoning to code writing, and now to Agents that can autonomously plan and execute tasks. RSI, by contrast, discusses another question: once AI already possesses a certain level of research and development capability, can it be made to participate in its own improvement, and through such improvement continuously enhance its subsequent research capability? This question was in fact discussed a long time ago, but only recently have the technical conditions for carrying out engineering validation gradually become available. 1. What is RSI, and why is it worth attention now?

RSI stands for Recursive Self-Improvement. The key to understanding it is "recursive": /an AI system obtains improvement through its own research activities, and the improved system can then carry out the next round of research with stronger capabilities./

For example, today's Coding Agents can already automatically optimize a certain GPU Kernel, and by modifying code, running tests, and analyzing results, ultimately find a better-performing implementation. If we further let it analyze its own research process and improve candidate generation, experiment selection, or context management mechanisms so that the next round of algorithm optimization is more efficient, then the object of improvement expands from external programs to the research system itself. If this capability can be continuously passed on, it forms the recursive feedback that RSI focuses on.

This idea can be traced back to the intelligence explosion hypothesis proposed by I. J. Good in 1965. He envisioned that if a machine could design a machine stronger than itself, then the next generation of machines might in turn design even stronger successors, thereby forming a continuously strengthening feedback loop of intelligence. Later, Schmidhuber's Gödel Machine attempted to study self-modification from a formal perspective, requiring the system to prove that a modification can increase expected utility. However, the behavior of complex AI systems is difficult to fully cover through rigorous mathematical proof, so modern research has gradually shifted toward self-improvement based on experimental feedback.

In the past, AutoML, neural architecture search, and reinforcement learning had already achieved, to a certain extent, automatic optimization of models or policies by machines, but the search space and optimization objectives were usually defined in advance by researchers. The development of large language models has changed this condition; in particular, Coding Agents have begun to be able to understand complex code, modify training algorithms, organize experiments, and analyze results, and the objects and scope of automated search have therefore expanded.

Yuandong Tian also talked about this change in the interview. The gradual maturation of large models' programming capability and their understanding of AI research objects has given automated scientific research a more realistic foundation. After all, whether it is model architecture, training algorithms, or inference optimization, ultimately everything needs to be implemented through programs. When AI can understand and modify these programs, it also gains the ability to directly participate in the AI R&D process. 2. What is the core mechanism of RSI?

From a technical perspective, RSI can be understood as two interrelated feedback loops. The first loop is responsible for completing specific research tasks, and the second loop is responsible for improving the system that executes these research tasks itself.

The first loop is already relatively easy to understand. A research Agent proposes hypotheses according to the objective, generates candidate solutions, executes experiments through a Sandbox, and then an Evaluator verifies them according to metrics such as performance, correctness, and cost. Effective solutions are retained, and the system continues the next round of search. FunSearch and AlphaEvolve both embody this mechanism, generating candidate programs through language models and combining executable evaluation with evolutionary search to continuously discover better algorithm implementations.

The second loop is the problem that RSI truly needs to solve. The experience accumulated by the system during the research process needs to be able to in turn improve its own research strategies, tools, or model capabilities. For example, a certain research Agent discovers that a large number of experiments are consumed on repeated invalid candidates, and therefore modifies its own search strategy and historical experiment retrieval mechanism. If the new version can find more effective improvements under the same budget, then its research capability has been enhanced. Only by further letting the new version continue to modify itself can a sustained recursive process possibly form.

Several key technical issues are involved here. First is the modification boundary: what exactly can the system modify—Prompts, the Agent Harness, tool code, or further, model parameters and training algorithms? Modifications at different levels involve completely different costs and risks. Second is the evaluation mechanism: how to confirm that a modification truly enhances the system's capability, especially avoiding obtaining false gains by modifying test conditions. Finally is capability inheritance: the code, experimental results, failure experience, and strategy improvements formed during the research process need to be preserved as long-term assets and be effectively usable by subsequent versions.

This also means that RSI imposes new requirements on the Agent Runtime. In addition to task planning, tool invocation, and context management, the system also needs to support experiment orchestration, version branching, independent evaluation, and persistence of research state. In particular, the experience gained during research has genuine recursive improvement value only if it can change subsequent research behavior. III. What research today is truly approaching RSI?

In my view, the most noteworthy recent work is the Darwin Gödel Machine (DGM) and AIDE², both of which have begun to directly study the capability improvement of agents themselves.

Proposed in 2025, DGM allows a coding agent to directly modify its own code, judges the effect of modifications through coding benchmarks, and retains multiple historical versions to form evolutionary branches. One important design is open-ended exploration: the system can continue searching from different historical versions, thereby reducing the local-optimum problem caused by retaining only the current best solution. The paper reports that under specific experimental conditions, the system's score on SWE-bench rose from 20.0% to 50.0%. This shows that an agent can improve task performance by modifying its own program, without always relying on manual adjustment of the execution mechanism.

The 2026 paper Recursive Self-Improvement of AI Research Agents further proposes AIDE², which directly takes the AI research agent itself as the object of improvement. The system modifies the code of the research Agent and then uses independent AI R&D tasks for evaluation, letting the validated new version carry out the next round of modification. The paper reports seven consecutive improvements over eight consecutive days of operation, involving mechanisms such as search strategy and context management, and observed capability transfer on some tasks that were not involved in optimization selection.

These studies have moved RSI from a theoretical concept into a stage where empirical analysis is possible. However, what is still relatively well demonstrated is self-improvement under specific tasks and specific system structures. Whether the foundation model itself can continuously design better training algorithms and model architectures, or even form cross-domain autonomous scientific research capability, still needs further verification.

This also corresponds to a question emphasized by Tian Yuandong in an interview: AI can already undertake a considerable amount of scientific research execution work, but scientific research also includes discovering important problems, proposing new concepts, building theoretical explanations, and judging the value of research directions. The progress of Coding Agents has solved a large number of difficulties at the execution level, while higher-level scientific research judgment remains a capability boundary that needs to be broken through. IV. How can RSI enter industrial scenarios?

Judging from the current level of technological maturity, I believe RSI is best suited to first enter AI Infra, especially model training, inference optimization, and algorithm engineering. These fields share favorable common conditions: /the research object can be expressed through code, experiments can be executed automatically, and results can be compared through clear performance metrics./

Taking inference systems as an example, we now need to continuously adjust parallelism strategies, batching parameters, caching mechanisms, and low-level computational operators according to model, hardware, and workload characteristics. If a research Agent is introduced, the system can analyze performance bottlenecks based on runtime metrics, automatically generate candidate configurations and implementations, and then select better solutions through isolated experiments. Further, the experience accumulated from historical experiments can be used to improve search strategies, so that the system reduces repeated exploration when facing new models and hardware environments.

Training systems offer similar opportunities; for example, learning rate strategies, data mixing ratios, optimizers, and training scheduling can all serve as objects of automated research. However, training experiments are usually more costly, and conclusions obtained from small-scale experiments may not necessarily transfer to large-scale training, so staged validation and strict computational budget management are needed.

If these capabilities are placed into existing Agent architectures, one finds that the required underlying components are not unfamiliar. The Agent Runtime is responsible for research planning and tool scheduling, the Sandbox provides an isolated execution environment, the experiment platform is responsible for computing resource management, the Evaluator is responsible for result validation, and code repositories, object storage, and experiment databases preserve cross-version research assets.

What truly needs further development is the cross-generational capability evaluation mechanism. How much performance gain a single optimization brings is relatively easy to measure; whether the optimized research system is better at discovering new optimization schemes requires designing independent tasks and comparing the performance of different versions under the same computational budget.

From the perspective of enterprise input and output, research cost also needs to be considered. The fact that an automated system can execute more experiments does not mean that R&D efficiency will necessarily improve. If finding an optimization scheme with limited returns consumes a large amount of compute, the actual value may not be high. Therefore, the number of effective improvements per unit of computational budget, the transferability of research results, and the research efficiency of subsequent versions should become more important metrics than the number of experiments. V. How will RSI affect the future development of AI?

The most noteworthy potential impact of RSI is that AI R&D efficiency may form a new accumulation mechanism. In the past, model R&D mainly depended on researchers, computing resources, and experiment cycles. If AI can continuously participate in research, and research results can further enhance the next round of research capability, then R&D efficiency may generate a stronger positive feedback.

This change will first affect scientific research organizations and AI infrastructure. Researchers can devote more energy to problem definition, research judgment, and key experiment design, while automated research systems undertake a large amount of candidate exploration, code modification, and experimental validation. AI Infra may also gradually develop, from a resource management and task execution platform, the capability to continuously optimize models, algorithms, and computing systems.

But whether recursive improvement can ultimately lead to an intelligence explosion may be too early to conclude. Improvements in research capability may be jointly constrained by compute, data, validation cost, and algorithmic bottlenecks. Even if the system can complete multiple effective improvements in succession, one cannot directly infer long-term exponential growth, let alone guarantee that this capability can naturally transfer to open-ended scientific research.

Judging from the papers publicly available so far, RSI already has an early engineering foundation, but there is still a considerable research path before general, continuous, and reliable recursive self-improvement.

What I pay more attention to is its impact on Agent technical architecture. At present, the lifecycle of most Agents is still centered on a single task; even if memory and historical records are preserved, the execution system itself usually does not undergo substantive capability changes as a result. RSI further requires that research results can enter subsequent versions, so that the system's execution strategies, tools, and research methods continue to improve as tasks accumulate.

Along this direction, the future Agent Runtime may need to expand from task execution management to capability version management, experimental evaluation, and long-term evolution management.

With regard to RSI, what is truly worth continuous observation is whether AI can increasingly effectively improve its own ability to improve. This is also the core of what distinguishes recursive self-improvement from ordinary automated optimization.