Google and Penn State Introduce SIM, Enabling AI to Carry Optimization Experience Across Models and Tasks
Google Cloud AI, Google DeepMind, and Penn State jointly introduced Self-Improving Memory (SIM), which uses continuously updatable memory to retain the rationale, expected outcomes, and actual results of Harness code changes, enabling continuous recursive self-improvement across models and tasks. In 10 independent programming experiments, the SIM approach performed better in nine, with an average lead of 2.36 points. In cross-model office tasks, the SIM approach achieved a final score of 84.75, compared with 73.42 for the approach without SIM. The experiments modified the external runtime framework for large language models while leaving model weights unchanged, and the approach that inherited prior experience trailed by 0.8 points on Gemini 4 Argon.
Why read this
AI score 65
SIM incorporates the rationale, expected outcomes, and actual results of code changes into an updatable memory, providing a concrete way to reuse Harness optimization experience across models and tasks.
Topics
Body · AI translation
Zhidx reported on November 19 that a research team from Google Cloud AI, Google DeepMind, and Pennsylvania State University recently published a paper titled “Self-Improvement Memory Enables Continual Recursive Self-Improvement,” proposing a new method called Self-Improvement Memory (SIM).
▲“Self-Improvement Memory Enables Continual Recursive Self-Improvement”
Previously, although AI could autonomously optimize the operating framework, or Harness, external to a large model, the framework often had to be optimized all over again after switching models or tasks. For the first time, SIM organizes the reasons behind code changes, their expected effects, and their actual results into a continuously updated memory, enabling AI to inherit and adapt past improvement experience and achieve continual recursive self-improvement (CRSI) across models and tasks.
The research team then conducted tests in four fields: programming, text classification, office tasks, and engineering design. In 9 of 10 independent controlled programming experiments, the approach incorporating SIM performed better, with an average lead of 2.36 points. In another cross-model office-task test, the approach incorporating SIM ultimately scored 84.75 points, while the approach without SIM scored only 73.42, a difference of 11.33 points.
The study’s first author, Xin Yu, graduated from Nankai University in 2023 with a degree in statistics and is currently pursuing a PhD in statistics at Pennsylvania State University. The research was completed during an internship at Google.
▲Xin Yu
Xin Yu has interned at Microsoft Research Asia, ByteDance Seed, TikTok, and Meta, and received Microsoft Research Asia’s 2023 “Star of Tomorrow” award. Multiple research papers to which Xin Yu contributed have been published at leading international AI conferences including ICML, ICLR, and NeurIPS.
Another author of the paper, Adams Wei Yu, is a researcher at Google DeepMind who has worked on projects including Gemini, Bard, and the AI video generation model Veo.
▲Adams Wei Yu
After the paper was published, Xin Yu explained the motivation behind the work on X. Xin Yu noted that previous research on recursive self-improvement generally focused on a fixed large model and benchmark task, whereas real-world agents need to continually adapt to new tasks, datasets, and models. The research team therefore wanted AI’s optimization experience to continue accumulating as its environment changes.
▲Xin Yu’s post
“What carries forward should include the experience behind the code—not just the code itself. (What should be passed on is not merely the code itself, but also the improvement experience behind it.)” Xin Yu wrote.
Recursive Self-Improvement (RSI) refers to having AI participate in improving AI systems and enhancing their performance through repeated iteration. In this study, the object being improved by AI is not the large model itself, but a set of external programs that operate around the large model—the Harness.
The Harness manages prompts, context, tool calls, task execution workflows, and more. For example, when AI is asked to complete a programming task, the Harness can determine when to invoke the terminal, how to check the results of running the code, and what conditions must be met before the task can be declared complete.
Through RSI, AI can autonomously modify the Harness code, repeatedly test the effects of those changes, and retain the better-performing versions, thereby improving an agent’s task-completion rate without updating the large model’s weights.
However, this type of optimization is usually designed for a fixed model and task. A Harness optimized for Claude may no longer be effective after switching to Gemini, while rules previously designed for programming tasks may not apply to new tasks.
More importantly, although a modified Harness retains the code, it does not necessarily preserve the reasons behind each change or the experience gained from testing it. When the model or task changes, the AI responsible for optimization may repeat earlier attempts and may even remove features that were originally effective.
To address this, the research team proposed Continual Recursive Self-Improvement (CRSI), with the aim of allowing automated Harness optimization to persist across models and tasks while using SIM to preserve previous improvement experience. SIM is like a “lab notebook” that records each adopted modification, the reason for the change, the expected result, the actual effect, and the corresponding model and testing environment.
When a new task or model appears, the AI responsible for optimizing the Harness can consult these records to determine which rules should be retained, modified, or removed, rather than having to explore everything again from scratch.
▲SIM workflow
The system then writes the adopted modifications and their actual effects into SIM. Even if a modification fails to achieve the expected result, the unsuccessful outcome is retained as a reference for subsequent optimization.
In a Git server configuration task, the AI prematurely declared the task complete after local tests succeeded, but it had omitted the crucial step of creating the specified user account, causing external validation to fail. Through automated optimization, the researchers added a task-completion checking mechanism to the runtime framework, preventing the AI from ending a task prematurely when there was insufficient evidence of execution.
When testing shifted to an optical waveguide routing task, a new problem emerged: although the routing paths generated by the AI satisfied the constraints, this did not mean the routing cost was sufficiently low. Subsequent optimization therefore retained the previous checking mechanism while adding reminders to check solution quality, preventing the AI from ending a task merely because it had passed basic constraint checks.
This is also what SIM aims to achieve: enabling AI both to carry forward effective improvements from the past and to continue adapting to the requirements of new tasks.
To evaluate SIM, the research team conducted two main types of experiments: one kept the underlying model unchanged while changing the task data, and the other kept the test tasks unchanged while changing the underlying model.
In the first type of experiment, the researchers used Gemini 3.8 Flash as the underlying model and transferred an already optimized runtime framework to new tasks.
The results showed that across six test sets covering programming, text classification, office tasks, and engineering design, the approach incorporating the full SIM achieved the highest final score in every case, outperforming the approach that merely reused the existing runtime framework by 2.43 to 4.35 points.
The second type of experiment involved transfer between different models. The researchers first optimized a Harness for Claude Opus 5.5, then had Gemini 3.8 Flash use that Harness and continue the automated optimization process.
They found that after directly transferring the Harness designed for Claude to Gemini, its initial performance in all four test domains was worse than Gemini's performance with the original Harness.
After SIM was introduced, however, the AI responsible for optimization could draw on previous modification experience and adjust the parts unsuitable for Gemini. Ultimately, the SIM approach achieved the highest score in all four domains, outperforming the approach without SIM by 3.60 to 11.33 points.
The largest gap appeared in the office-task test. Both approaches started at 68.41 points. The approach with SIM ultimately reached 84.75 points, compared with 73.42 points for the approach without SIM.
▲Cross-model test results
To further validate the results, the researchers also conducted 10 independent controlled programming experiments.
With the same initial code, underlying model, tasks, and optimization budget, the approach with SIM scored higher in nine of the 10 experiments, achieving an average final score of 21.70 points, compared with an average of 19.33 points for the approach without SIM. The average difference was 2.36 points, with a statistical test result of p=0.003.
▲Results of 10 controlled programming experiments
The research team also attempted to have the runtime framework and its memory adapt sequentially to Gemini 2.5 Flash, Gemini 3.5 Flash, Gemini 3.8 Flash, and Gemini 4 Argon.
Compared with optimizing each generation of model from scratch, the approach that inherited past experience led by 3.5 and 1.6 points on Gemini 3.5 Flash and Gemini 3.8 Flash, respectively. On Gemini 4 Argon, however, the approach that inherited experience ultimately scored 0.8 points lower.
The study also sparked discussion on X. Tech blogger Superman noted that when AI systems previously attempted self-improvement, they tended to forget earlier successes, repeat mistakes, or become trapped in ineffective cycles of modification.
With SIM, AI can review past modification records and learn which approaches worked and which failed, allowing it to continue optimizing on the basis of earlier achievements.
Another developer, Brian Skellenger, cautioned that although the AI in this study could autonomously modify code and continuously optimize through testing feedback, it was modifying the runtime framework external to the large model, rather than the model's own architecture or parameters. The paper also explicitly states that the large model's weights remained unchanged throughout the experiments.
Overall, the study's advance lies in enabling AI to accumulate and reuse previous optimization experience. In addition to saving the modified code, the system records the reason for each modification, its expected effect, and its actual outcome, so that this experience can continue to be used after switching tasks or models.
As the paper's first author, Xin Yu, put it, AI systems need to accumulate not only optimization results, but also the processes through which those results were obtained.
In the past, recursive self-improvement (RSI) enabled AI to optimize its own Harness by repeatedly modifying code and testing the results, but previous optimization gains were often difficult to carry across models and tasks. Now, SIM allows AI not only to inherit modified code, but also to make use of past successes and failures, extending RSI to new models and tasks and providing a new method for continual recursive self-improvement (CRSI).
However, the continual optimization in these experiments still operated within preset test scopes and numbers of iterations, and cross-model transfer did not always yield better results. For AI to achieve truly long-term and stable self-evolution, the challenge of how to autonomously evaluate, select, and update historical experience still needs to be addressed.
This article is from the WeChat public account “Zhidx” (ID: zhidxcom), written by Jiang Yu and edited by Li Shuiqing. Published by 36Kr with authorization.
AI Trends