RSI: why recursive self-improvement in AI is still stalled

By Turing Post.

This issue launches the newsletter's format change: the section formerly called "AI 101" is renamed "∇ Guide" (nabla, the mathematical symbol for the gradient) and will be published on Fridays, connecting explanations of AI concepts with the publication's "Almanac."

The central topic is Recursive Self-Improvement (RSI): the idea that AI can help improve future AI systems, which would in turn help build still better ones. According to the email, it has already been shown that AI can improve AI, but that has not translated into sustained RSI, and the article explores why.

The analysis starts from an explicitly cited paper: "What is Missing from AI Post-Training AI: An Empirical Analysis," produced by researchers at Tsinghua University and other institutions. The study examines AI agents that carry out post-training of other models, distinguishing two levels of capability: 'execution' (following an existing training plan: preparing data, tuning hyperparameters, fixing errors, picking the best checkpoint) and 'strategy' (changing the plan itself: switching training method, adding or removing stages, using a different kind of data, redirecting the remaining compute).

The researchers counted as a 'training experiment' only those cases in which the agent actually updated the model's parameters. By that criterion, the results show that 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 rose from 22.0% to 41.4%. In other words, agents are already capable of preparing data, launching training runs, diagnosing technical problems, repairing failed runs and iterating for hours.

The main limitation, according to the paper, appears at the strategic level: agents tend to settle on a training strategy at the outset and stick with it, spending the rest of the time on minor adjustments. In only about 2% of cases did they try a different strategy, even when results suggested they should change approach. Curiously, the preferred approach depended more on the specific agent than on the task: Claude Code tended to choose full-parameter supervised fine-tuning (full-parameter SFT), while Codex CLI preferred parameter-efficient fine-tuning.

The researchers tested three possible fixes, and none solved the underlying problem:

1. More experience (giving agents memory, useful skills and feedback) improves only execution: on HumanEval performance rose from 22.0% to 62.8% (versus 41.4% without this enhancement), but it does not make agents change strategy when they should.

2. Human guidance helps in choosing a different initial approach, but once training begins the agents again get stuck on minor local adjustments.

3. More inference compute (between 2 and 8 times more) helps on easy tasks, but makes little difference on the hardest ones.

The conclusion of this first part is that the problem is not a lack of experience, guidance or compute: AI can already iterate to improve AI, but iterating is not the same as rethinking the approach. Today's agents rarely stop to think "maybe my whole approach is wrong"; they lack a reliable way to recognize an error mid-process, so they risk continuing to iterate in the wrong direction and getting trapped in unnecessary local fixes. This, the article notes, is not just a post-training problem but an obstacle to RSI in general.

The second section, "Why Iteration Matters for RSI but Is Not Enough," digs into this distinction: in recursive self-improvement, the key word is 'recursive,' because the results of one research cycle feed back into the next, helping the system choose better improvement options. Iterating means repeating the 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 only becomes recursive when the AI uses what it has learned to rethink the next round's strategy (the training method, the data configuration, the evaluation process or the model-building pipeline). According to the email, most agents today complete this loop only at the execution level, while the broader process of 'AI improving AI' remains largely linear.

The issue's table of contents also announces other sections —"Meta^n: A Layered Approach to Improving RSI," "Recuris: The Power of Memory in RSI," a conclusion titled "Everything is a puzzle piece" and a list of sources— whose content was not included in the body of the email received, so it is not possible to summarize them accurately.

As an additional note outside the main topic, the email links to a YouTube video from Turing Post itself about how NVIDIA went from nearly disappearing with its NV1 chip to finding in deep learning the market for CUDA, framing it within its reported $12.9 billion acquisition of Hugging Face.