A 113-page survey of more than 500 studies organizes a fast-growing field into a closed-loop framework — and lays the groundwork for a working system.
For years, the dominant formula for building more capable artificial intelligence was straightforward: give models more data, more computing power and more guidance from people.
But the human part of that formula may eventually become a limiting factor. High-quality annotation and expert feedback are expensive and difficult to scale. And as large language models approach human-level performance in some specialized areas, people may not always be able to provide the knowledge needed to push them further.
A team of Stony Brook University researchers is exploring another possibility: Could increasingly capable AI models help drive their own continued improvement?
In the paper “Self-Improvement of Large Language Models: A Technical Overview and Future Outlook,” the researchers reviewed more than 500 studies and organized a rapidly expanding but fragmented area of AI research into a unified framework. The 113-page paper is currently under review at Transactions on Machine Learning Research and is also available as an arXiv preprint.
The work was led by first author Haoyan Yang, a PhD student in the Department of Computer Science, under the guidance of Jiawei “Joe” Zhou, an assistant professor in the departments of Applied Mathematics and Statistics and Computer Science.
“Much like writing a program that recursively improves itself, we wanted to explore how LLMs could drive their own advancement,” Zhou said.
Here, “self-improvement” does not mean that an AI system has already become capable of independently redesigning itself without limits or human oversight. Instead, the paper examines how a model’s existing abilities could be used to automate specific parts of its own development.
The distinction is important. Much of the current research on self-evolving AI agents focuses on improving the system surrounding a fixed model — such as its prompts, tools and memory — primarily while the model is completing a task. The Stony Brook team asks a broader question: not just what a model can adjust while completing a task, but what it can contribute across the entire process of its own development.
The researchers envision that process as a continuous, closed loop composed of four interconnected stages with an evaluation process:
- Data acquisition: The model gathers material from existing sources, interacts with an environment or generates new training data.
- Data selection: The system evaluates which information is valuable enough to learn from while filtering out low-quality, repetitive or misleading data.
- Model optimization: Through a cycle of generation, reward and optimization, the model produces potential answers or reasoning, evaluates their quality and uses the results to update its capabilities.
- Inference refinement: The model improves its performance when responding to a request through methods such as reflection, tool use, more effective reasoning strategies and test-time adaptation.
- Autonomous Evaluation: Throughout the cycle, an automated evaluation layer keeps checking whether a claimed improvement is real, and uses what it finds to shape the next step.
Lav R. Varshney, Della Pietra Infinity Professor and director of the AI Innovation Institute at Stony Brook University, who was not involved in the paper, said the framework gives researchers a common structure for approaching the problem.
“The paper introduces a coherent, four-stage closed-loop lifecycle — Data Acquisition, Data Selection, Model Optimization and Inference Refinement — wrapped in an Autonomous Evaluation Layer, which gives the field a clean engineering blueprint to build within,” Varshney said.
He added that advances in LLM capabilities have made the prospect of automating portions of model development increasingly plausible.
“Since modern LLMs are now robust enough to write code, evaluate complex logic and run various computing environments, it makes sense to delegate the engineering tasks of model development to the model itself,” Varshney said. “This can accelerate progress, as long as there is some external information coming in.”
Building a Framework for a Moving Field
The project grew out of Yang’s earlier research on self-improvement and conversations he began having with Zhou in 2025. Rather than simply compiling existing methods, they wanted to determine how those methods could function together as parts of a larger system.
“We did not want to simply catalog and list existing methods; we wanted to organize them within the survey into a unified, model-driven framework,” Yang said.
As the paper’s student lead, Yang designed the overall taxonomy, coordinated weekly meetings and individual workflows, and worked closely with Zhou to refine the framework. He also led the research and writing for the sections on model optimization, applications and the field’s future direction.
Constructing a coherent system from a fast-moving body of research proved to be the project’s greatest challenge. The team revised its framework repeatedly as new studies, industry investments and technical approaches continued to emerge.
“The most interesting part was watching the field take off as we worked,” Yang said. “We kept finding more and more papers and industrial investments on self-improvement, especially this time around 2026, along with a growing number of talks, which made us realize our framing was well-founded and timely.”
In addition to Yang and Zhou, the paper was co-authored by Stony Brook graduate students Mario Xerri, Solha Park, Huajian Zhang, Yiyang Feng and Sai Akhil Kogilathota.
Balancing Improvement with Oversight
The same closed loop that could allow a model to improve also creates substantial risks. A model trained on its own output could gradually reinforce its mistakes or lose performance as low-quality information accumulates. It could learn to exploit the metric used to evaluate it rather than develop the intended capability. As more of the process becomes automated, it may also become harder for people to understand, control and redirect the system.
For that reason, the researchers emphasize the need to balance automation with human oversight and greater capability with safety. Autonomous evaluation would need to rely on dynamic benchmarks and interactive environments rather than a single fixed test, and outside information would still be necessary to prevent the improvement cycle from becoming closed off from reliable evidence.
The team has already taken the next step, turning the framework from the survey into a functioning self-improvement system. A comprehensive evaluation of that system is now underway, aimed at carrying the design from blueprint to dependable technical infrastructure — and at establishing how much of AI’s future development AI can safely help perform itself.
Add comment