When AI keeps improving its own answers, why do they all drift toward the same thing, and can separate groups help?

This explores how an island model, which runs several separate populations of candidate solutions and only occasionally lets them swap ideas, keeps an iterative refinement loop from converging on one answer. The corpus has no note on island models themselves, so I can't tell you how any specific paper implements one. What it does have is a lot of material on why iterative loops lose diversity and on the ingredients an island model would rely on.

First, the problem. Pure self-improvement loops tend to narrow. Can models reliably improve themselves without external feedback? lists diversity collapse as one of three structural failure modes, alongside the gap between generating and verifying answers and reward hacking. The same pattern shows up in RL training. Outcome-based RL sharpens a policy toward whatever already scores well, and Does outcome-based RL diversity loss spread across unsolved problems? finds this even reduces diversity on problems the model hasn't solved yet. Does reinforcement learning squeeze exploration diversity in search agents? sees the same squeeze in search agents. A single population with one selection pressure keeps rewarding its own best guesses, so it drifts toward one strategy.

An island model attacks this by structure. Each island is protected from the others' selection pressure, so different lineages can settle on different strategies. The corpus's closest match to this idea is Can reward vectors be the hidden source of solution diversity?. It keeps rewards separate per test case, criterion or persona instead of averaging them into one score, so different solutions can specialize on different trade-offs. That is a different mechanism from islands, but the goal is the same: give each subgroup its own reason to stay distinct, without an external diversity penalty.

The corpus also shows what other loops use to avoid collapse, and each of these could apply to islands. Step-level critique during training counteracts tail narrowing across self-training rounds (Do critique models improve diversity during training itself?). Progressive Draft Refinement compresses memory between iterations instead of piling up revisions, which avoids the noise buildup of plain sequential revision (Do iterative refinement methods suffer from overthinking?). The ACE framework uses incremental updates instead of full rewrites so detail isn't eroded (Can context playbooks prevent knowledge loss during iteration?). The self-improvement note also suggests that occasional migration between islands would work as an outside anchor, since reliable methods succeed by bringing in something external, such as past model versions or third-party judges.

Two cautions from the corpus apply here. Diversity effects depend on the domain: Does preference tuning always reduce diversity the same way? shows the same tuning shrinks diversity in code and increases it in creative writing. Islands that help in one setting may not help in another. And the corpus doesn't say how often islands should exchange candidates. If they mix too often they behave like one population, and if they never mix you get several separate loops that each still narrow. If you want an answer about island models specifically, the collection would need a paper on evolutionary or population-based search.