TL;DR: Current AI agents can iterate and improve results, but rarely rethink the strategy behind that improvement. New research explores how to fix this limitation: Metaⁿ recursively adds layers that improve the agent’s strategy, while Recuris evolves how agents store, retrieve, verify, and apply experience through memory.
Housekeeping, folks: AI 101 is becoming ∇ Guide and moving to Friday. We’re going beyond the basics of AI concepts, models, and techniques, connecting these explanations to our Almanac. The symbol is called nabla, used in mathematics to express a gradient. A small nod to how models learn, as we keep learning ourselves.
RSI is an abbreviation that has become all too familiar over the past few months. That’s how we describe it: Recursive self-improvement (RSI) is the idea that AI can help improve future AI systems, which can then help build even better ones.
For years, developers have worked to improve model capabilities. Now, as models become more capable, the temptation is to hand more of that work over to them. The promise is just too compelling: models that could
learn from experience, much as humans get better at their jobs
help design and train stronger successors
adapt to unfamiliar problems without waiting for human updates.
These capabilities could accelerate scientific and engineering progress beyond anything we have seen before. And it’s hard not to get excited about AI agents becoming first-class contributors to science.
Today, we already see that AI can improve AI (we are going to publish an interesting interview about that next week). But why hasn’t that turned into sustained recursive self-improvement? A lot depends on the answer, including how quickly we can expect AI to advance.
In this article, we’ll look at what’s holding RSI back. Even if RSI isn’t your thing, these bottlenecks are worth understanding because many of them slow down AI development more broadly. So let’s figure out where progress gets a little bit stuck.
In today’s episode:
What limits Recursive Self-Improvement? Iteration Is Not the Same as Rethinking
Why Iteration Matters for RSI but Is Not Enough
Metan: A Layered Approach to Improving RSI
Recuris: The Power of Memory in RSI
Conclusion: Everything is a puzzle piece
Sources and further reading
Not interested in this topic? In this video we examine how NVIDIA went from nearly dying with NV1 to finding CUDA’s market in deep learning, and how that journey led to its reported $12.9 billion acquisition of Hugging Face. Has NVIDIA figured out how to make open-source AI extraordinarily profitable?
Follow us on
What limits Recursive Self-Improvement? Iteration Is Not the Same as Rethinking
We’ll start our exploration with a paper that, at first glance, doesn’t seem to be about RSI: “What is Missing from AI Post-Training AI: An Empirical Analysis” by researchers from Tsinghua University and other institutions. It looks at AI agents post-training other AI models. This may sound like a narrow training problem, but it gives us a concrete way to see how far AI can already go in improving AI, and what is still missing from that process.
Researchers separate an agent’s ability into two levels:
Execution – the agent follows an existing training plan. It prepares data, tunes settings, fixes bugs, adjusts rewards, and chooses the best checkpoint.
Strategy – the agent changes the plan itself. It might switch to another training method, add or remove a training stage, use a different type of data, or redirect its remaining compute.
The researchers counted something as a training experiment only when the agent actually updated the model’s parameters. Then they looked at what changed between experiments:
Small changes within the same plan like tuning hyperparameters were counted as execution changes.
Changes to the training method were counted as strategy changes.
The results of their experiment unveiled an important insight. Agents already succeed at training models. Across 1,338 trajectories, they completed 5,111 training runs and improved average benchmark performance from 10.4% to 23.0%. On HumanEval, for example, average performance increased from 22.0% to 41.4%. So agents can do the work: prepare data, launch training, diagnose technical problems, repair failed runs and iterate for hours.
The main limitation appeared at the strategy level: agents usually pick a training strategy at the very beginning and tend to stay there, spending the rest of their time on making small adjustments. Agents tested a different strategy in only about 2% of cases. They don’t change the method of improvement even when the results suggest they should try something different.
It is interesting that the preferred approach often depended more on the agent than on the task. While Claude Code mostly chose full-parameter supervised fine-tuning, Codex CLI “prefered” parameter-efficient fine-tuning.
The researchers suggested that there has to be a way to solve this. They tried three fixes but non of them work toward the goal:
More experience like giving agents memory, useful skills, and feedback improves only their execution. Agents achieved better scores: for example, HumanEval performance increased from 22.0% to 62.8% (compared to 41.4% without the fix). However, more experience doesn’t make agents change their strategy when needed.
Image Credit: “What is Missing from AI Post-Training AI: An Empirical Analysis” paper
Human guidance helps agents choose a different initial approach at the beginning. But once training begins, they again become stuck in small local adjustments.
Image Credit: “What is Missing from AI Post-Training AI: An Empirical Analysis” paper
And finally, giving agents 2–8x more inference compute helps on easier tasks, but makes almost no difference on the hardest ones.
The problem, then, is not a lack of experience, guidance, or compute. AI can already iterate to improve AI. But being able to iterate is not the same as being able to rethink. Current agents rarely can stop and “think”: maybe my whole approach is wrong? They just seem to lack a reliable way to recognize a mistake during the iteration process, so they risk to keep iterating in the wrong direction and get stuck in the cycle of local unnecessary corrections.
So it is bigger than a post-training problem. It is an obstacle for RSI →
Why Iteration Matters for RSI but Is Not Enough
In recursive self-improvement, AI becomes part of the process that improves future AI systems. The key word here is recursive: the results of one research cycle are fed back into the next cycle, helping the system choose better options for improvement. To iterate means to repeat automated research loop while learning from each attempt:
Propose an idea → implement it → run an experiment → evaluate the results → revise the plan → try again.
This loop becomes recursive only when AI uses what it learns to rethink the strategy behind the next round including the training method, data setup, evaluation process, or model-building pipeline.
Now most agents complete this loop at the execution level. But the broader process of “AI improving AI” remains mostly linear.
Image Credit: “What is Missing from AI Post-Training AI: An Empirical Analysis” paper
The “recursive” part is still limited: agents keep working within the original plan. They can try a different approach when a human points them toward it. But they don’t make that change on their own.
A bigger model may not solve this. Agents may need to be trained to recognize when an approach isn’t working, with checkpoints where they stop and ask whether to keep going or try something else.
That’s why the researchers pay so much attention to strategy. It determines what the agent tries next. If the approach is wrong, another round of work may just take it further in the wrong direction.
Fortunately, there is another fascinating study that proposes a practical solution to improve the process of RSI. →
Metan: A Layered Approach to Improving RSI
Don’t settle for shallow articles. Learn the basics and go deeper with us. Truly understanding things is deeply satisfying.
Join Premium members from top companies like Microsoft, NVIDIA, Google, HF, OpenAI, a16z, plus AI labs such as Ai2, MIT, Berkeley, .gov, and thousands of others to really understand what’s going on in AI.
How did you like it?
FAQ
What Is Recursive Self-Improvement in AI?
Recursive self-improvement (RSI) is the idea of AI becoming part of the process that improves future AI systems. The recursive part appears when results from one improvement cycle feed back into the next, allowing the system to improve not only its outputs but also the strategy, tools, memory, or processes used to produce future improvements.
What Is Limiting Recursive Self-Improvement Today?
One major limitation is strategy. AI agents can already run experiments, train models, fix errors, and iterate on solutions, but research shows they rarely abandon an initial strategy and choose a fundamentally different approach on their own. More experience, human guidance, or inference compute does not fully solve this problem.
How Can AI Improve Its Own Improvement Strategy?
Metaⁿ explores one approach by recursively adding new layers above an existing solver. Each layer can examine the code, results, and failures produced below it, then add strategic guidance and reusable tools. Higher layers therefore get a broader view of the improvement process without modifying the underlying solver or the fixed meta-operation.
Can Memory Improve Recursive Self-Improvement?
Yes, but simply giving an agent more memory or skills may not be enough. Recuris combines experiential memory with working memory, a skill-invocation policy, and verification. It also learns from failures across tasks to update the relevant parts of this memory system, creating a controlled recursive improvement loop outside the model weights.
What Are the New Approaches to Recursive Self-Improvement?
Recent research is exploring several ways to improve different parts of RSI. Metaⁿ focuses on recursively improving strategy through increasingly deep meta-layers, while Recuris evolves an external memory system that controls what experience an agent uses, when it uses it, and how task progress is verified. Neither represents complete RSI, but each addresses a specific limitation.
