Does an AI that upgrades itself from the inside actually differ from a model that just gets retrained the usual way?

This explores what separates recursive self-improvement (RSI) from the familiar process of updating a model's policy, meaning the way it chooses actions or outputs. Surprisingly, the corpus suggests the two are closer than the dramatic framing implies. One framework treats them as the same improvement cycle and tells them apart by two dials. The first is whether the thing doing the improving sits inside the agent or outside it. The second is whether the standard for 'better' comes from the agent itself or from the world What separates self-improvement from policy improvement?. Ordinary policy improvement keeps both dials on the outside: an external trainer applies an external reward. RSI turns both dials inward, and that is where the trouble starts.

Turning the second dial inward runs into a wall. When a model judges its own progress, it hits a circularity problem: it can't reliably check work it couldn't have produced better in the first place. It also tends to narrow its range of outputs and to game its own reward Can models reliably improve themselves without external feedback?. Methods that work in practice quietly bring an outside anchor back in, such as an earlier model version, a third-party judge, user corrections, or feedback from tools. So most of what gets called 'self-improvement' today is closer to policy improvement with clever plumbing. A 1,250-paper survey draws the same line between bounded self-refinement, which is checkable and common in industry, and open-ended RSI, which is still limited by grounding, collapse dynamics and compute Are self-refinement and recursive self-improvement actually the same thing?.

A second, less obvious distinction is *what* gets updated. Policy updates usually mean changing weights, and even there reinforcement learning turns out to touch only 5–30% of parameters, in patterns that repeat across random seeds Does reinforcement learning update only a small fraction of parameters?. Self-improving agents, by contrast, increasingly work through a fast loop that rewrites prompts, memory and tools rather than weights. This is cheaper and easy to undo, and it's where most recent progress has come from Do self-improving agents really split into two distinct loops?. A stronger version of RSI targets the improvement process itself. An agent automating R&D makes better products, but it only escapes diminishing returns if it also rewrites the code that does the researching Can recursive self-improvement speed up the research process itself?. Dream-RSI shows a middle path. It replays its own past discoveries as a cheap simulator for scoring and improving how it explores, so the agent's history becomes its training ground Can past discoveries train better exploration policies?.

The real dividing line may be objectives, not mechanics. One debate participant argues that fast RSI depends on AIs proposing their own research goals without drifting from what was intended. That is exactly the step policy improvement never takes, because a human always sets the goal Can AIs learn to specify their own research objectives?. Whether the loop then compounds is an open question. Modeling suggests acceleration depends on the *product* of several feedback strengths, so one weak link stalls the whole loop, and current loops are strengthening but not yet self-sustaining Are AI feedback loops strong enough to sustain recursive self-improvement?. Direct evidence is thin: one 8-day run reported seven accepted self-rewrites but gave no sizes or timings for the gains Does recursive self-improvement sustain gains or hit diminishing returns?. Sakana AI is betting that the path forward is sample efficiency rather than more compute Can sample efficiency replace compute scale in recursive self-improvement?. Anthropic has warned that if the loop does close, the societal risks justify considering slowing down Does recursive self-improvement pose serious risks to society?.