Source-linked AI summary
EvoX: Meta-Evolution for Automated Discovery
Shu Liu, Shubham Agarwal, Monishwaran Maheswaran, Mert Cemri, Zhifei Li, Qiuyang Mang, Ashwin Naren, Ethan Boneh, Audrey Cheng, Melissa Z. Pan, Alexander Du, Kurt Keutzer, Alvin Cheung, Alexandros G. Dimakis, Koushik Sen, Matei Zaharia, Ion Stoica
TL;DR
Existing LLM-driven evolutionary methods often use fixed search strategies that cannot adapt reliably across tasks or changing optimization stages. EvoX co-evolves candidate solutions and the strategies that select and vary them, updating strategies from progress and population state. Across nearly 200 real-world tasks, it outperforms prior AI-driven evolutionary methods on the majority of tasks.
Problem
Existing evolutionary systems often fix search-strategy choices such as candidate selection and exploration-exploitation settings, limiting adaptation across tasks and optimization stages.
Method
EvoX uses two-level evolution to generate candidate solutions while dynamically evolving search strategies based on progress, prior strategy performance, and population state.
Results
Across nearly 200 real-world optimization tasks, EvoX outperforms OpenEvolve, ShinkaEvolve, and GEPA on the majority, including 96% of math and system optimization benchmarks.
Takeaways & Limitations
EvoX dynamically shifts between search strategies as optimization progresses, rather than relying on a fixed strategy throughout execution.
Takeaways & Limitations
The evaluation uses a fixed budget of 100 iterations per task for open frameworks, while AlphaEvolve's publicly unavailable iteration budget prevents matched-budget comparison.
Abstract
from arXiv · showhide
Recent work such as AlphaEvolve has shown that combining LLM-driven optimization with evolutionary search can effectively improve programs, prompts, and algorithms across domains. In this paradigm, previously evaluated solutions are reused to guide the model toward new candidate solutions. Crucially, the effectiveness of this evolution process depends on the search strategy: how prior solutions are selected and varied to generate new candidates. However, most existing methods rely on fixed search strategies with predefined knobs (e.g., explore-exploit ratios) that remain static throughout execution. While effective in some settings, these approaches often fail to adapt across tasks, or even within the same task as the search space changes over time. We introduce EvoX, an adaptive evolution method that optimizes its own evolution process. EvoX jointly evolves candidate solutions and the search strategies used to generate them, continuously updating how prior solutions are selected and varied based on progress. This enables the system to dynamically shift between different search strategies during the optimization process. Across nearly 200 real-world optimization tasks, EvoX outperforms existing AI-driven evolutionary methods including AlphaEvolve, OpenEvolve, GEPA, and ShinkaEvolve on the majority of tasks.
1. Introduction
LLM-driven evolutionary optimization depends on how prior candidates are selected and varied, yet existing strategies are often fixed and can stagnate across tasks or optimization stages. EvoX addresses this by co-evolving solutions and search strategies, with broad improvements across nearly 200 tasks.
- Existing paradigm: LLM-driven evolutionary systems select and vary previously evaluated candidates to generate, evaluate, and reintegrate new solutions.The search strategy determines which candidates the generator sees and how they are modified.
- Limitations of fixed strategies: Fixed strategies and manually configured knobs can fail to generalize across problems, require retuning, or cause stagnation as optimization progresses.Some tasks favor repeated local refinement, while others require qualitatively different solution structures.
- Within-run adaptation: A MAP-Elites-style strategy can make rapid early progress yet later stagnate, whereas sampling candidates along different trade-off objectives can enable continued improvement.This within-run shift is illustrated in a multi-objective signal processing task.
- EvoX: EvoX treats search strategy as an evolvable object in a two-level process combining solution evolution with meta-evolution.The solution loop generates candidates, while the meta-evolution loop updates the strategy.
- EvoX: EvoX dynamically updates strategies using optimization progress, prior strategies, their performance, and the current solution-population state.When progress stagnates, an LLM generates a new strategy conditioned on these signals.
- Evaluation: Across nearly 200 real-world optimization tasks, EvoX outperforms OpenEvolve, ShinkaEvolve, and GEPA on the majority, including 96% of math and system optimization benchmarks.It also often matches or surpasses the best human-designed solutions.
2. Related Work
Prior LLM-guided evolutionary systems differ mainly in candidate selection and variation, while recent adaptive methods often modify the generator rather than the search strategy. EvoX extends meta-learning for optimization by making the search strategy itself evolvable.
- Foundations: Evolutionary search maintains a population of prior solutions and explicitly selects and varies candidates over time.This paradigm builds on iterative systems that organize and reuse past information.
- LLM-guided evolutionary search: LLM-guided evolutionary systems span prompt optimization, program discovery, and algorithm discovery, with methods differing primarily in candidate selection and variation.Examples include MAP-Elites, Pareto-frontier selection, and diversity-driven selection.
- Learning-based adaptation: Some methods adapt the generator through search alternation, fine-tuning, or reinforcement learning, but do not adapt the search strategy governing selection and variation.The distinction is between changing the model and changing the evolutionary process around it.
- Meta-learning: Meta-learning and learned optimization treat the optimization procedure itself as the object of adaptation rather than a fixed component.Prior work includes learned optimizers, symbolic update-rule discovery, gradient-based meta-learning, and reinforcement-learning-based optimizers.
- EvoX: EvoX treats the search strategy as an evolvable object and adapts candidate selection and variation across optimization stages and heterogeneous solution landscapes.It uses evolutionary feedback rather than fixed selection and variation mechanisms.
3. Problem Formulation
The formulation represents LLM-driven optimization as iterative candidate evaluation under a fixed budget, with a search strategy constructing each generation context. Strategies choose parents, variation operators, and optional inspirations to maximize the final best score.
- Candidate evaluation: Candidate solutions x ∈ X are generated by a language model, evaluated by a task-specific evaluator, and assigned a scalar score with auxiliary artifacts.The artifacts may include logs, traces, or other feedback.
- Solution population database: Optimization proceeds for T sequential evaluation steps, appending each evaluated candidate to the database D_t to form D_t+1.The database contains all candidates evaluated up to the current step.
- Search strategy: A search strategy S constructs the next-generation LLM input from the current database D_t.It determines the generation context used for proposing the next candidate.
- Search strategy: The strategy selects one or more parent candidates, specifies a variation operator π, and may provide an inspiration set of additional exemplars.These components define what is modified and which other candidates inform generation.
- Variation operators: EvoX uses local refinement for fine-grained exploitation, structural variation for coarse-grained exploration, and free-form variation without edit-scope constraints.The operators are instantiated with task-specific prompts, and their semantics depend on the task.
- Optimization goal: The optimization goal is to adaptively select and improve the search strategy to maximize the final best score under a fixed evaluation budget of T steps.The fixed budget defines the scope of the optimization objective.
4. Co-evolving Solution and Search Strategy
EvoX co-evolves the solution population and the search strategy through alternating solution evolution, progress monitoring, and strategy updates. Strategy changes are state-conditional, use historical strategy performance and population descriptors, and preserve the existing population.
- Algorithm setup: EvoX co-evolves the solution population D_t and the search strategy under an initial budget, evaluator, solution generator, strategy generator, and validity test.Algorithm 1 initializes the population and strategy before alternating the evolution phases.
- Solution evolution: During solution evolution, the active strategy selects parents, variation, and inspiration context; the solution generator produces a candidate, and the evaluator appends it to the population.This repeats for a window of W evaluation steps or until the total budget is exhausted.
- Progress monitoring: EvoX evaluates a strategy over multiple steps using progress across a sliding window because a strategy shapes a sequence of candidates rather than one candidate in isolation.The window-based assessment captures downstream progress induced by the strategy.
- Strategy updates: If progress falls below the stagnation threshold τ, EvoX generates and validates a replacement strategy; otherwise, it continues with the current strategy.This makes switching demand-driven rather than periodic at a fixed interval.
- State-conditional adaptation: Conditioning on strategy history and current population signals enables state-conditional adaptation instead of fixed or periodic switching.The approach targets non-stationarity caused by changing candidate quality, diversity, and variation outcomes.
- Strategy memory: The strategy database records each deployed strategy, its population-state descriptor, and observed performance, enabling strategy selection based on prior outcomes and similar conditions.Population descriptors include score statistics, frontier structure, progress indicators, and recent window statistics.
- Meta-evolution: Meta-evolution selects high-performing parent strategies and inspirational strategies, then uses an LLM to mutate strategy components conditioned on the current population descriptor.Mutations can change parent selection, inspiration construction, or variation-operator preferences.
- Strategy deployment: Validated strategy switches preserve the solution database, so evolution continues from the current search state rather than resetting the population.Invalid candidates are retried within a fixed budget before falling back to the previous strategy.
5. Case Study: Signal Processing
On a noisy, changing time-series filtering task, EvoX adapts its search strategy as progress stalls, moving from simple search and refinement toward diverse sampling, structural exploration, and final polishing. This adaptive process discovers more capable filtering programs and achieves a 34.1% higher final score than a static baseline under the same 100-iteration budget.
- Task and evaluation: The task evaluates filtering programs on signal fidelity, smoothness, low lag, and minimal false trend changes using a combined score.These objectives capture competing requirements for filtering a noisy, changing time series.
- Overall result: 34.1% higher final score is achieved by EvoX than by the static baseline through strategy changes driven by observed progress.Figure 2 reports major gains at approximately iterations 48 (+0.119), 70 (+0.056), and 96 (+0.022).
- Static baseline: The static baseline improves from 0.499 to 0.530 before stagnating with simple moving-average and exponential-moving-average filters.Its random selection and generic free-form variation produce only tiny later adjustments, limiting performance on complex noise patterns.
- Adaptive search phases: At iteration 40, EvoX switches from greedy refinement to stratified multi-objective sampling, combining programs strong on different objectives to discover hybrid designs such as SSA with Whittaker smoothing.The strategy samples parents and inspirations from diverse score tiers and objective-specific rankings.
- Adaptive search phases: When progress slows and candidates become similar, EvoX increases UCB-guided structural variation to explore complex program changes and underused parent programs.The resulting solutions incorporate higher-order filters, smoothing kernels, and forward–backward filtering, yielding a further +0.056 improvement.
- Adaptive search phases: By iteration 90, EvoX shifts to UCB-guided local refinement, making precise changes to top solutions while avoiding premature narrowing and adding the final +0.022 gain.Large structural changes had begun destabilizing performance, so the search moved toward fine-tuning.
6. Evaluation
EvoX is evaluated across 196 real-world optimization tasks and consistently performs strongly against LLM-driven evolutionary baselines. Its adaptive strategy evolution also improves robustness, solution quality, and cost efficiency across task types and optimization stages.
- Evaluation setup: EvoX is evaluated on 196 tasks spanning mathematics, systems, ALE-Bench-Lite, and Frontier-CS, with comparisons against major LLM-driven evolutionary frameworks.The evaluation also includes ARC-AGI-2 and ablations of cost and scaling behavior.
- Mathematical optimization: Under GPT-5, EvoX achieves the best or tied-best result on 7 of 8 mathematical tasks, while Gemini-3.0-Pro yields the best result on all 8.Mean performance is best on 6 of 8 GPT-5 tasks and 7 of 8 Gemini-3.0-Pro tasks.
- Mathematical optimization: EvoX matches or exceeds AlphaEvolve on 5 of 7 mathematical tasks within 100 iterations.Direct cost comparison is unavailable because AlphaEvolve’s iteration budget is not publicly specified.
- System optimization: EvoX exceeds human-best results on all six system benchmarks and achieves the best GPT-5 mean score on every task.Under Gemini-3.0-Pro, it has the best mean on 5 of 6 tasks; telemetry repair is the sole exception.
- Algorithmic and research benchmarks: On algorithmic and research benchmarks, EvoX records a highest average private score of 1958.2 on ALE-Bench-Lite and a Frontier-CS mean of 62.6 with median 75.5.These results improve over OpenEvolve, GEPA, and ShinkaEvolve on Frontier-CS, while some ALE-Bench-Lite gains remain smaller when strong initial solutions bias search toward local optima.
- Adaptation and efficiency: Across Heilbronn triangle ablations, EvoX improves beyond fixed strategies regardless of initialization and reaches 0.0339 while breaking through GEPA’s 0.0323 plateau.EvoX and GEPA exceed 0.031 for less than $1, compared with $7.6 for ShinkaEvolve and $15.4 for OpenEvolve.
A. Benchmark Details
EvoX is evaluated on optimization benchmarks spanning mathematical, system-performance, algorithmic, and research problems.
- 196 problems cover 24 optimization tasks across mathematical optimization, system performance, and algorithmic challenges, plus 172 FrontierCS problems.
- The mathematical benchmarks include geometric, combinatorial, and signal processing optimization problems.
- The system benchmarks are drawn from ADRS-Bench and represent realistic production-system optimization problems.
- The algorithmic and research suite contains 182 open-ended problems from ALE-Bench-Lite and FrontierCS.
B. Additional Results
Additional evaluations examine EvoX beyond the primary optimization benchmarks.
- Additional evaluations assess EvoX behavior beyond the primary optimization benchmarks discussed in the main paper.
B.1. ARC-AGI Evaluation
EvoX is additionally tested on ARC-AGI-2, a reasoning-oriented benchmark that differs from conventional evolutionary optimization.
- ARC-AGI-2 evaluates abstract and compositional reasoning across programmatic problem-solving instances.
- OpenEvolve and EvoX are compared under a matched budget of 30 LLM iterations per task.
- ARC-AGI evaluation assumes strict train–test separation, unlike evolutionary frameworks that adapt during inference.
- The ARC-AGI results indicate that EvoX maintains performance improvements on reasoning-oriented tasks outside traditional optimization settings.
B.2. Full Results
Full results compare EvoX with recent evolutionary systems while describing the search-strategy interface and prompt constraints used in evaluation.
- B.2. Full Results: The evaluation includes comparisons with CodeEvolve and ThetaEvolve as additional prior approaches for LLM-driven optimization.
- B.2. Full Results: EvoX maintains consistent performance advantages over both fixed-policy and adaptive-policy baselines across tasks.
- B.2. Full Results: The mathematical and system result tables report mean and best scores, with directional arrows indicating whether maximization or minimization is preferred.
- Search Strategy: The optimizer constructs each next-generation input by selecting a parent, one variation operator, and an optional inspiration set.
- Search Strategy: Search-strategy quality is measured by how effectively it improves the best solution during optimization.
- Search Strategy: The strategy interface separates adding evaluated programs from sampling the parent, operator, and inspiration set.
- Prompt Constraints: Stored program metadata and predefined label constants are read-only, while generation supports free-form, local-refinement, or structural-variation operators.
- Prompt Constraints: Local refinement requests small structure-preserving edits, whereas structural variation encourages larger exploratory changes.
D. Analysis of Search Evolution
Across Signal Processing, Circle Packing, and Heilbronn Triangle, search-strategy evolution tracks score improvements and phase-specific outcomes over 100 iterations.
- Signal Processing improves from 0.499 to 0.743 over the 100-iteration run.The table reports the improvement achieved within each phase window.
- Circle Packing improves from 0.364 to 1.0004 over the 100-iteration run.The table reports phase-window improvements alongside the evolving variation operator.
- Heilbronn Triangle reaches a final minimum area of approximately 0.0365, with its best result at iteration 91.The table also reports SOTA 1.0 as a reference value.
D.1. Case Study: Search Evolution in Circle Packing
The Circle Packing case study shows EvoX moving from heuristic exploration through diversity-inducing search toward refinement and qualitatively different optimization mechanisms. Its score rises from 0.364 to 1.0004 as the evolved strategy changes with search progress.
- Circle Packing requires dense, stable, non-overlapping configurations that converge reliably within the evaluation budget.Candidates are evaluated using a normalized packing score under constraint satisfaction.
- EvoX raises the Circle Packing score from 0.364 to 1.0004 by evolving search strategies alongside candidate solutions.The case study describes the sequence of evolved strategies across the run.
- Exploration-Dominated Phase: During exploration, EvoX discovers stable geometric constructions and achieves rapid early gains of +0.59.These gains primarily come from viable placement schemes rather than precise optimization dynamics.
- Diversity-Inducing Strategies: As structural redundancy emerges, usage-penalized and tiered sampling increase exposure to structurally diverse programs.These strategies discourage repeated parent selection and reduce over-exploitation of early heuristics.
- Mechanism Discovery via Structural Variation: Under stagnation-aware sampling, structural variation discovers constrained numerical optimization with SLSQP instead of sequential local displacement rules.This marks a qualitative shift in how solutions are constructed.
- Strategy evolution enables transitions from geometric heuristics to coordinated global updates that static strategies fail to realize.The case study frames this as optimization-mechanism discovery rather than merely improved sampling efficiency.
- Refinement-Dominated Phase: EvoX shifts toward local refinement after SLSQP solutions emerge because large structural edits destabilize high-quality configurations.Refinement adjusts convergence parameters, constraint tolerances, and step-size dynamics.