If an AI grades its own homework, why doesn't it just get smarter and smarter on its own?

This explores why a model that grades and retrains on its own work tends to stall, and what outside reference actually makes the loop produce progress. The core problem is simple: a model can only teach itself something if it is better at checking answers than at producing them. That difference, the generation-verification gap, sets a hard limit on how far self-improvement can go What limits how much models can improve themselves?. The gap gets larger as models get bigger, but it disappears for factual tasks. A model that doesn't know a fact can't recognize the right answer any better than it can produce it, so self-checking gives it nothing to learn from. That is why the corpus's broader verdict is that pure self-improvement goes in circles. It stalls through the verification gap, through outputs getting less varied over time, and through the model learning to game its own reward Can models reliably improve themselves without external feedback?.

You can see the circularity in how these loops fail. Small mistakes in self-generated training data don't stay small. They compound within two or three rounds, leaving the model at a level of performance set by how well it checks its work, not by what it can actually do How quickly do errors compound during model self-training?. Self-rewarding training breaks down in a quieter way. As the model improves, its 'better' and 'worse' answers become nearly identical, so the gap between them, which is the only thing the training learns from, shrinks about ninefold. The fix is telling: compare current answers against an *older* version of the model Why does self-rewarding training collapse when responses improve?. Once the loop stops looking only at itself, the signal comes back.

Here's the twist worth knowing: many methods that advertise 'no external signal' still have one, just disguised. Meta-Rewarding adds a judge that grades the model's own judging, and scores jump Why do self-improvement loops plateau without updating the judge?. SERL has the model take turns answering and judging, and rewards it when its own rankings stay consistent Can models learn to judge themselves without external rewards?. Asymmetric self-play has one copy of the model write problems for another copy to solve, checking answers by majority vote Can language models improve themselves without any external training data?. Another approach starts self-improvement from just 1,000 human-made examples of how to deepen shallow reasoning Can models improve themselves on tasks without verifiable answers?. In each case, something outside the loop being trained keeps it from drifting: a second layer of judging, a consistency requirement, a split between roles, or a small seed of human examples. The clearest version uses a *different* model as the outside view. Rewards from a varied group of peer models beat a single model rewarding itself, and often come close to training on real answer keys Can peer models replace external judges for reward signals?. Diversity matters because a group of different models is less likely to share the same blind spots.

The same logic shows up at a much larger scale. Whether AI-assisted research truly compounds depends on a 'reproduction number': is the feedback strong enough to outrun how fast research problems get harder What determines whether AI self-improvement actually compounds?? One estimate finds today's feedback loops are getting stronger but can't yet sustain themselves Are AI feedback loops strong enough to sustain recursive self-improvement?. Headline claims of recursive self-improvement, like seven successive self-rewrites, often don't report how big each gain was or whether the gains were shrinking Does recursive self-improvement sustain gains or hit diminishing returns?. The takeaway: when you meet a 'self-improving' system, the useful question isn't whether it improves itself. Ask where its outside anchor is hidden, and whether that anchor knows anything the model doesn't.