AI systems that write code, score it, and breed the winners can improve fast, but do they ever break past their built-in limits?
Do evolutionary discovery systems like FunSearch count as bounded or open-ended improvement?
This explores whether an evolutionary loop like FunSearch (an LLM proposes candidate programs, an automated scorer keeps the best, and the winners seed the next round) can keep improving without a ceiling, or stalls at whatever its fixed setup allows.
This explores whether an evolutionary loop like FunSearch can keep improving without a ceiling, or stalls at whatever its fixed setup allows. The collection has no note on FunSearch itself, so this answer comes from its closest relatives. They suggest FunSearch-style systems are bounded in what they can change, even though the gains inside that boundary can be large.
The useful question is what the loop is allowed to rewrite. In the closest relatives here, humans fix the operators and the scorer, and the loop only evolves the candidates. Frontis-MA1 evolves programs with four fixed moves (Draft, Improve, Debug, Crossover). Training the model on those same moves stacked with search instead of replacing it, and lifted Medal Average on MLE-Bench Lite from 39% to 71% Can training and search gains add together in program evolution?. Mind Evolution uses the same recipe at inference time, with crossover and mutation across separate islands of candidates, and solves 98%+ of planning tasks Can evolutionary search beat sampling and revision at inference time?. Those are big numbers, but they measure progress toward a target someone else defined, using a search procedure someone else designed. That is bounded improvement.
The scorer is what makes these systems work, and it is also their ceiling. One note argues that pure self-improvement is circular: checking your own work is too close to the same skill as producing it, diversity collapses, and rewards get gamed. It finds that the methods that reliably work all bring in an outside anchor, such as tool feedback or a third-party judge Can models reliably improve themselves without external feedback?. A FunSearch-style automated evaluator is exactly that kind of anchor. The loop can only get as good as the evaluator can tell the difference.
Open-endedness starts when the loop edits the things FunSearch holds fixed. The Darwin Gödel Machine rewrites its own code and keeps an archive of agent variants, gated by empirical benchmarks rather than proofs, and gets 2.5× on SWE-bench Can AI systems improve themselves through trial and error?. In bilevel autoresearch, an outer loop reads the inner loop's code and writes new search mechanisms at runtime, for a 5× gain on GPT pretraining Can an AI system improve its own search methods automatically?. SAGA goes further and evolves the objective itself, compiling new goals into scoring functions the inner loop can use immediately Can agents evolve their own objectives during search?. A survey lays this out as a ladder: dynamic peers first, then adaptive environments and feedback, then the evolution mechanism itself Can agents evolve beyond the constraints humans engineer?. FunSearch-style systems sit on the first rung.
The ladder is not a clean story, because higher rungs bring their own bounds. Gains only count as open-ended if they transfer. An evolved agent matched its human-built version on four held-out benchmarks Does automated evolution match human-built agent performance?, and ModularRSI evolved harness modules on benchmark-disjoint data and still improved unseen tasks Can harness modules improve separately from benchmark data?. The model doing the editing is also a limit. The ability to write useful harness edits is flat across model tiers, but the ability to benefit from them peaks in mid-tier models Do stronger models always evolve harnesses better?. Dream-RSI points at one way past this: it replays accumulated discovery history to score exploration policies cheaply, so the system learns how to explore rather than only exploring Can past discoveries train better exploration policies?.
So FunSearch-style evolution is bounded by its evaluator and its fixed operators. It is open-ended only in the size of the candidate space it can search. It becomes open-ended in the stronger sense once it can rewrite its own search mechanism or scoring function, and that is where the newer systems in this collection are heading.
Sources 11 notes
Frontis-MA1 trained a single model on four program-evolution operators (Draft, Improve, Debug, Crossover), then reused those operators in long-horizon evolutionary search. The result was complementary gains: learning and search improved performance together rather than substituting for each other, lifting Medal Average from 39.39% to 71.21% on MLE-Bench Lite.
Mind Evolution, an evolutionary search strategy using LLM-generated crossover and mutation with island model diversity, solves 98%+ of planning tasks and significantly outperforms best-of-N and sequential revision strategies while working directly in natural language without task formalization.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.
DGM replaces formal proofs with empirical benchmarking and maintains an evolutionary archive of agent variants, achieving 2.5× improvement on SWE-bench and 2.2× on Polyglot by discovering capabilities like better code editing and context management.
An outer loop successfully read inner loop code, identified bottlenecks, and generated new Python mechanisms at runtime, discovering combinatorial optimization and bandit methods that broke the inner loop's deterministic patterns and improved performance on GPT pretraining by 5x.
Show all 11 sources
SAGA's bi-level architecture closes a feedback loop from optimization results back to goal design by having an outer LLM loop propose new objectives and compile them into code the inner loop can immediately use, enabling objective formulation as part of discovery rather than a fixed input.
A survey framework organizes co-evolving systems into three stages that progressively remove human engineering: dynamic peers first, then adaptive environments and feedback, finally the evolution mechanism itself. Single-entity self-improvement stalls in static contexts; co-evolution supplies adaptive pressure across multiple components.
AIDE85, evolved through seven accepted rewrites in 8 days, equals or surpasses AIDEhuman on four held-out benchmarks spanning in- and out-of-distribution tasks including weather forecasting. The result shows automated design iteration can match human-driven R&D on generalization.
ModularRSI evolves harness modules independently using contrastive trajectories on benchmark-disjoint data, showing consistent gains across unseen tasks and domains. The approach isolates mechanism-level improvements from task-specific adaptation by aggregating evidence across tasks before updating components.
Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.
Dream-RSI demonstrates that accumulated discovery trees can be replayed off-policy to score exploration policies without repeated online evaluation. The framework loops between policy evaluation on historical data, online redeployment, and simulator expansion, reportedly achieving competitive discovery quality at lower cost.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Self-Improvements in Modern Agentic Systems: A Survey
- DarwinX: Evolving Agent Harnesses Through Natural Selection
- Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
- The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
- Dream-RSI: Recursive Self-Improvement through Evolving Worlds
- Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
- Evolving Deeper LLM Thinking