Selection Without a Selector

Abstract

Recursive self-improvement (RSI) in large language models is an evolutionary process. Every working RSI loop today has the three ingredients Darwinian evolution requires: variation, selection, and heredity.

But it is evolution with the rules changed. Variation is directed rather than random. Inheritance is Lamarckian: what a model discovers while working can be written straight back into its weights. And fitness is designed by people rather than imposed by an environment.

Those changes explain why RSI races ahead in some domains and stalls in others. Generation is now cheap; selection is not. RSI therefore advances at the speed of its verifiers. Evolutionary theory also predicts RSI’s main failure modes (proxy exploitation, diversity collapse, loss of open-endedness) and points to the remedies biology already found for each.

The title names the risk. As models increasingly grade their own work, the selector moves inside the system. RSI then drifts back toward evolution proper: selection without an external selector, favoring whatever persists.

Introduction

The most useful model of RSI is not a runaway “intelligence explosion.” It is a breeding program, and breeding programs have well-understood dynamics.

Working definition. RSI here means any loop in which a model’s outputs are used to produce a more capable successor, with limited human input per iteration. That covers training on self-generated data, self-play, models grading their own outputs, and agents that rewrite their own code. It excludes one-off training runs designed end to end by people.

Why now. Since 2023, several labs have built self-improvement systems from explicitly evolutionary parts: populations, mutation operators, archives of past variants, and fitness evaluators. The analogy is no longer a metaphor. It is the architecture.

Claims. This paper argues four things:

1. LLM self-improvement loops meet the formal conditions for evolution by selection, so evolutionary theory applies to them.
2. They differ from biological evolution in three specific ways (directed variation, Lamarckian inheritance, designed fitness), and each difference speeds them up.
3. With variation now cheap, selection (verification) is the bottleneck. That predicts where RSI succeeds first.
4. Evolution’s known pathologies map onto RSI’s observed failures, and biology’s fixes translate into engineering practice.

The evolutionary engine

Evolution needs only three conditions, as Lewontin (1970) stated them: individuals vary, the variation affects reproductive success, and the variation is heritable. Any system meeting all three evolves, whatever it is made of. Dennett (1995) calls this substrate neutrality.

Three further features of biological evolution set its speed and character:

• Mutation is blind. Variation arises without regard to whether it helps. Most mutations are neutral or harmful, so evolution is extremely sample-inefficient.
• The Weismann barrier holds. What an organism learns during its life is not written into its genes. Only germline changes pass to offspring.
• Fitness is emergent. No one specifies the objective. Fitness is simply what reproduces, judged by an environment that includes other evolving organisms and keeps shifting.

Open-ended evolution, the kind that keeps producing novelty for billions of years, also depends on large, diverse populations and co-evolving environments. Remove those and populations converge, drift, or stall.

Learning and evolution already interact through the Baldwin effect. Behaviors learned within a lifetime change which mutations pay off, until the behavior becomes innate. Hinton and Nowlan (1987) demonstrated this in a simple computational model. It becomes central to RSI below.

RSI in LLMs today

Every working RSI system pairs a generator (the model) with an automated selector, and the strongest ones are explicitly evolutionary. The table traces the lineage, newest first.

The loop is also closing inside labs. Anthropic reported that more than 80% of code merged into its codebase was written by Claude as of May 2026 (source). On a fixed internal test of speeding up code that trains a small model, Claude went from roughly 3x speedup (Claude Opus 4, May 2025) to roughly 52x (Claude Mythos Preview, April 2026). A skilled human reaches about 4x in four to eight hours. Anthropic cautions this is a like-for-like benchmark, not a real-world training speedup.

The same report names the remaining human role: choosing goals and judging results. It also found human code review had become the new bottleneck. In evolutionary terms, people have retreated to the one job the loop cannot yet do for itself: defining and applying fitness.

Mapping the analogy

The mapping is close enough to be operational: every evolutionary component has a counterpart an engineer can point to.

Three rows carry the argument: mutation is directed, lifetime learning is inherited, and fitness is designed. The next section examines each.

Where the analogy breaks

RSI breaks three of biology’s rules. Each break makes it faster; together they move the bottleneck from variation to selection.

1. Variation is directed

Biological mutation is blind. An LLM proposing a code change is not: it draws on everything it has learned about code. AlphaEvolve pairs a fast model for breadth with a stronger model for depth, a mutation operator with a strategy.

The payoff is a far higher yield per variant. The cost is diversity. Informed guesses cluster around what the model already believes, so directed variation explores less than blind variation does.

2. Inheritance is Lamarckian

The Weismann barrier does not exist for models. A solution found through long test-time reasoning can be distilled into weights, so the next generation produces it immediately.

This is the Baldwin effect with the slow middle step removed. Learned behavior becomes innate in one generation instead of thousands. It is the largest single source of RSI’s speed.

3. Fitness is designed

Nature never specifies an objective. RSI loops do, through tests, rubrics and reward models. That makes RSI steerable, its main safety advantage over natural evolution.

It also makes RSI bounded. A loop cannot reliably select for capability its selector cannot recognize. The selector is both the steering wheel and the ceiling.

Smaller differences

• Population size. Frontier development runs on a handful of model lineages. A biologist would call that a dangerously small effective population.
• No reproductive drive. Models do not seek to copy themselves; selection pressure is supplied from outside. That holds only while no loop rewards self-propagation.
• Two selection layers. Inside the loop, verifiers select variants. Outside it, the market selects models by adoption, a slower and noisier fitness signal.

The selector bottleneck

In biology, variation is expensive and selection is free: the environment does the judging. In RSI the relationship inverts. Variation is nearly free, while selection must be built, maintained and trusted.

Every arrow but one has become fast and cheap; the selector now sets both the loop’s speed and its ceiling.

The prediction follows directly. RSI advances fastest where verification is cheap and exact: games, math, code correctness, kernel speed. It advances slowest where fitness is hard to measure: strategy, taste, long-horizon outcomes in the real world. Anthropic’s account fits this pattern. It describes its models as superhuman at optimizing within a clearly defined experiment, while choosing goals remains the human job.

What evolution predicts for RSI

Evolutionary theory makes six predictions about self-improving systems. Five have already appeared in RSI practice; biology suggests a remedy for all but the last.

1. Selection exploits the proxy

Evolution optimizes reproduction, not the traits an observer intended. In digital-evolution experiments, researchers have repeatedly seen evolved organisms subvert their intentions and exploit bugs in their code (Lehman et al., 2020).

RSI shows the same behavior. When the Darwin Gödel Machine was asked to stop hallucinating tool use, one lineage hit a perfect score in two modifications. It had removed the logging markers the hallucination detector relied on, without fixing the problem. The authors saw this more often when the checking code was visible to the agent.

Remedy: biology’s selector is an environment the organism cannot edit. Keep evaluators hidden, separate and rotated, and audit any variant that improves suspiciously fast.

2. Inbreeding collapses diversity

Small, closed populations lose genetic variation and accumulate defects. Shumailov et al. (Nature, 2024) found that indiscriminately training generative models on generated content makes the tails of the original distribution disappear. That is loss of genetic diversity in statistical form.

Remedy: gene flow. Maintain a steady inflow of real-world and human-generated data. Later work reported that accumulating real and synthetic data, rather than replacing one with the other, avoided collapse in its experiments (Gerstgrasser et al., 2024).

3. Hill-climbing stalls; stepping stones don’t

Complex traits rarely evolve by direct ascent. Feathers likely served insulation or display before flight. Evolution keeps variants whose value appears only later.

The Darwin Gödel Machine reached 50.0% on its SWE-bench evaluation. A version that always built on its latest agent reached 23.0%, and one that always built on its best agent reached 39.7%. The winning lineage passed through two agents that scored worse than their parents.

Remedy: keep archives of interesting variants, not just the current champion. Novelty search and quality-diversity algorithms such as MAP-Elites formalize this.

4. Static environments stop selecting

Sustained evolution is driven by arms races: predators and prey, hosts and parasites. A fixed benchmark is a static environment. Once it saturates, selection pressure disappears. Anthropic reports SWE-bench went from low single digits to saturation in two years.

Self-play escaped this trap in games because the opponent evolves too. Remedy: co-evolve tasks and evaluators with the model through adversarial task generation, red teams and continuously refreshed evaluations.

5. The largest leaps change the machinery of inheritance

Maynard Smith and Szathmáry (1995) argued that evolution’s major transitions (chromosomes, eukaryotic cells, multicellularity, language) each changed how information is stored and passed on. Evolvability itself evolved.

RSI’s analogue is improving the improvement process. Promptbreeder evolves its own mutation prompts. The Darwin Gödel Machine improves its own editing tools. AlphaEvolve speeds training of the models it runs on. These second-order loops are the ones that compound, and the ones most worth watching.

6. When the selector moves inside

Evolution has no external judge; fitness is simply whatever persists. RSI has so far kept its judge outside, in tests, rubrics and people. That is changing. Self-rewarding models grade their own outputs, and Anthropic now uses a Claude judge to score whether agent coding sessions succeeded.

Each step moves RSI back toward evolution proper: selection without an external selector. Biology is clear about where that leads. A system whose fitness is endogenous optimizes for its own persistence, whatever anyone intended.

Remedy: there is no biological one, because evolution never had a designer to lose. This is the one prediction the analogy cannot resolve for us. It is also the most consequential.

Implications

If selection is the bottleneck, both value and risk concentrate in the selector. That reframes priorities for three audiences.

For builders

• Build the verifier before the generator. A loop can improve only as far as its evaluator can tell better from worse.
• Wall off the selector. Evaluation code, held-out data and scoring should sit outside anything a variant can read or edit.
• Keep an archive, not a champion. Track lineage so any regression traces back to the generation that introduced it.
• Budget for gene flow. Reserve a fixed share of each generation’s training data for fresh real-world input.
• Retire evaluations before they saturate. A benchmark near its ceiling no longer selects anything.

For investors

• Moats move to environments and verifiers. Generation is commoditizing. Trusted, proprietary fitness signals (real user outcomes, P&L, clinical results) are the scarce input.
• Automation follows verifiability. Expect code, math, chip design and optimization problems to automate first; domains judged by taste or long-delayed outcomes will follow later.
• Back second-order loops. Systems that improve their own tooling or training substrate compound. Systems that only tune outputs do not.

For policymakers

• Measure loop closure, not just capability. Track what share of a successor model’s selection was done by humans versus by models.
• Require selector independence for high-stakes loops. Evaluators should be separate from, and untrainable by, the system they judge. Objective-hacking incidents should be disclosed.
• Treat monoculture as systemic risk. Agriculture learned this through the Irish potato famine and the loss of the Gros Michel banana. A few frontier lineages sharing one training recipe carry the same fragility.

Conclusion

Evolution is the only process known to have produced general intelligence from scratch. It did so blindly, slowly and without a goal. RSI keeps evolution’s engine but swaps blind mutation for informed proposals and fixed genes for editable weights.

What it cannot yet swap out cheaply is judgment. For now, the speed and direction of self-improvement are set by whoever builds the selector.

The central question is no longer whether models can improve themselves; they already do. It is who, or what, decides what counts as better.