Does an AI get better faster by critiquing itself alone, or by competing and learning alongside a group of peer AIs?

This explores whether a lone model improving itself is fundamentally different from improvement that comes from a group of models or a lineage of agent variants. The corpus has no head-to-head benchmark of the two. Read together, though, it suggests a clear pattern: single-improver loops tend to stall for structural reasons, and population-based setups work largely because the population supplies the outside signal a lone model can't give itself.

Start with why the solo version stalls. A model can only improve itself if it is better at checking answers than at producing them. This generation-verification gap grows with model size but disappears on factual tasks, where checking is no easier than recalling What limits how much models can improve themselves?. Left alone, self-improvement also runs into diversity collapse (outputs narrow toward one style) and reward hacking (the model learns to game its own scores). Self-improvement methods that do work turn out to quietly bring in an external anchor, such as an older model version, a third-party judge, user corrections, or tool feedback Can models reliably improve themselves without external feedback?. One framework makes this precise with two dials: whether the improver sits inside the agent, and whether the standard it is judged against comes from outside. Reliable improvement needs the second dial turned toward outside What separates self-improvement from policy improvement?.

Populations are one way to build that outside standard. In label-free reinforcement learning, where there are no answer keys to learn from, models rewarded by a diverse cohort of peers consistently beat models rewarding themselves, and they often match training on ground-truth answers Can peer models replace external judges for reward signals?. The key word is *diverse*. Different models have different blind spots, so peers catch errors a single model would keep approving. A related idea is to let the evaluator evolve alongside the agent instead of fixing one judge in advance. That makes self-improvement possible on tasks with no clean grader, like writing or proofs, and it uses fewer tokens than a fixed evaluator Can evaluators improve alongside the agents they score?.

The most surprising finding is about how you pick which agents to build on. In searches where coding agents modify their own code, the highest-scoring agent is often a poor parent: its descendants stop improving. Some lower-scoring agents start lineages that improve much more over time. The better predictor is how productive an agent's whole family of descendants turns out to be, not its own benchmark score Does benchmark score predict a coding agent's self-improvement capacity?. A single-improver loop can't see this, because it only has one line of descent to judge. Sakana AI's lab treats recursive self-improvement as a problem of learning more from each attempt (including failures) rather than of buying more compute Can sample efficiency replace compute scale in recursive self-improvement?.

Two caveats. First, most real self-improvement today is bounded self-refinement, not open-ended recursion Are self-refinement and recursive self-improvement actually the same thing?. Much of it happens in the fast, cheap loop of prompts, memory, and tools rather than in model weights Do self-improving agents really split into two distinct loops? Does recursive self-improvement start with harness engineering?. Second, claims of sustained gains are thin: one widely cited run reports seven successive improvements but no evidence about whether returns diminish Does recursive self-improvement sustain gains or hit diminishing returns?. So the takeaway is not that populations win. Diversity and lineage are ways of supplying the outside reference that single improvers are missing.