When an AI teaches itself to improve, does it rewrite its own brain, or just upgrade its notes and tools?

This explores what changes, and what doesn't, depending on whether a self-improving AI updates its own weights or updates the things wrapped around a frozen model, such as prompts, memory and tools. One survey framework splits self-improving agents into a slow parametric loop that updates foundation-model weights and a fast non-parametric loop that updates prompts, memory and tools Do self-improving agents really split into two distinct loops?. The practical difference is cost and reversibility. A scaffold edit is cheap and can be undone, while a weight update is neither. That is why recent progress clusters in the fast loop.

Parametric loops make the model itself better at the task. Transformers taught 10-digit addition can reach 100 digits by generating solutions, keeping only the correct ones, and retraining. Their out-of-distribution gains kept compounding across rounds instead of flattening Can transformers improve exponentially by learning from their own correct solutions?. Asymmetric self-play does something similar with no human labels. One model proposes problems, another solves them, and both improve through RL Can language models improve themselves without any external training data?. A small set of 1,000 reasoning-enrichment demonstrations can also seed repeated self-improvement on tasks that have no checkable answer Can models improve themselves on tasks without verifiable answers?. Even the slow loop is less sweeping than it sounds. RL changes only 5 to 30 percent of a model's parameters, and it picks nearly the same ones across random seeds Does reinforcement learning update only a small fraction of parameters?.

In a non-parametric loop the model stays fixed and the machinery around it evolves. In bilevel autoresearch, an outer loop reads the inner loop's code, spots its bottlenecks, and writes new Python mechanisms at runtime. It found combinatorial-optimization and bandit methods that broke the inner loop's rigid patterns, for roughly a 5x improvement on a GPT pretraining task Can an AI system improve its own search methods automatically?. The skill that matters here looks more like stubborn iteration than raw talent. Across 17 frontier models on long optimization tasks, repeated benchmark-edit-incorporate cycles predicted success better than initial quality. Most models quit early or wasted their budget What predicts success in ultra-long-horizon agent tasks?. That study doesn't test parametric against non-parametric, but it shows what a fast loop rewards.

The two loops share the same weak spots. Pure self-improvement is circular in either form. Reliable methods succeed by bringing in an outside anchor such as an earlier model version, a third-party judge, user corrections or tool feedback Can models reliably improve themselves without external feedback?. Both are also capped by the generation-verification gap. A model can only improve where it checks answers better than it produces them, and that gap vanishes on factual tasks What limits how much models can improve themselves?. A frozen judge stalls either loop, and letting the judge co-evolve with the actor pushed AlpacaEval 2 from 23% to 39% Why do self-improvement loops plateau without updating the judge?. Today's loops of both kinds are also human-designed and fixed, and they break when the domain shifts Can AI systems improve their own learning strategies?.

The corpus doesn't compare the two loops head to head on their ceilings, so whether weight updates or scaffold updates eventually plateau sooner is open. Long-run evidence is thin for both. One recursive-improvement run reports seven accepted rewrites but gives no per-step gains, so it can't show that returns don't diminish Does recursive self-improvement sustain gains or hit diminishing returns?.