EvoX: Meta-Evolution for Automated Discovery

Shu Liu; Shubham Agarwal; Monishwaran Maheswaran; Mert Cemri; Zhifei Li; Qiuyang Mang; Ashwin Naren; Ethan Boneh

cs.LGFeb 26, 2026

EvoX: Meta-Evolution for Automated Discovery

Authors: Shu Liu, Shubham Agarwal, Monishwaran Maheswaran, Mert Cemri, Zhifei Li, Qiuyang Mang, Ashwin Naren, Ethan Boneh,

Organizations: UC Berkeley · Stanford University · Bespoke Labs

Abstract

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.

Figures & tables

Mathematics
Task	Human Best	AlphaEvolve	EvoX
Circle-Packing ( $\uparrow$ )	2.634	2.635	2.636
MinMaxMinDist ( $\downarrow$ )	4.16584	4.16579	4.16578
Task	Human Best	AI Best	EvoX
Systems Optimization
Cloud Transfer ( $\downarrow$ )	626.24	645.72	623.69

Table 1: EvoX across optimization tasks. We compare EvoX against strong human-designed methods and prior AI discovery systems, including AlphaEvolve, GEPA, ShinkaEvolve, and OpenEvolve, across mathematical optimization, systems performance optimization (e.g., GPU-to-model placement), and algorithm engineering tasks (e.g., Frontier-CS). For Frontier-CS, we report median scores over 172 competitive programming problems. Arrows indicate whether higher (

\uparrow

) or lower (

\downarrow

) scores are better. All open AI frameworks are evaluated under a fixed budget of 100 iterations, and the best result among them is reported as “AI Best.” AlphaEvolve results are taken directly from its original publication.

Panel of Figure 1 — Figure 1: Evolving the search strategy. ( 1 ) System architecture. EvoX has two coupled loops: an inner loop that evolves solutions, and an outer loop that evolves the search strategy that governs generation. ( 1 ) Effect of search evolution. A strategy with fixed exploration-exploitation ratio (MAP-Elites, red) stagnates, while an evolving search strategy (blue) produces discrete performance breakthroughs.

Figure 2: Evolving search strategy on the signal processing task. Starting from a uniform random sampling strategy, EvoX detects stagnation and adaptively switches to greedy search, stratified multi-objective sampling, UCB-guided structural variation, and finally local refinement. These strategy changes enable discovery of improved filtering programs and yield major gains at iterations $\sim$ 48 (+0.119), $\sim$ 70 (+0.056), and $\sim$ 96 (+0.022), achieving 34.1% higher final score than the static baseline.

				GPT-5				Gemini-3.0-Pro
Task	Stat	Human	AlphaEvolve	OpenEvolve	GEPA	Shinka	EvoX	OpenEvolve	GEPA	Shinka	EvoX
Circle Packing ( $\uparrow$ )	Mean	–	–	2.5308	2.6129	2.4642	2.6324	2.5414	2.6198	2.6216	2.6328
	Best	2.6340	2.6350	2.5414	2.6359	2.5161	2.6359	2.5414	2.6208	2.6358	2.6359
Circle Packing Rect ( $\uparrow$ )	Mean	–	–	2.2673	2.3265	2.3347	2.3525	2.3651	2.2160	2.3658	2.3658
	Best	2.3640	2.3658	2.2756	2.3537	2.3577	2.3600	2.3658	2.2507	2.3658	2.3658
heilbronn_convex ( $\uparrow$ )	Mean	–	–	0.0230	0.0259	0.0228	0.0263	0.0279	0.0228	0.0279	0.0287

Table 2: Main results for math optimization problems. We report mean and best over three runs. “

\uparrow

” / “

\downarrow

” indicate maximization / minimization. All open methods use 100 iterations under the same backbone model. Bold denotes the best LLM-based open framework result per model. Green cells indicate matches or improvements over human SOTA or closed-source AlphaEvolve results. In Circle Packing Rect, EvoX achieves 2.36583237, exceeding AlphaEvolve’s 2.36583213, although both appear equal when rounded.

			GPT-5				Gemini-3.0-Pro
Task	Stat	Human	OpenEvolve	GEPA	Shinka	EvoX	OpenEvolve	GEPA	Shinka	EvoX
EPLB ( $\uparrow$ )	Mean	–	0.1300	0.1338	0.1185	0.1358	0.1272	0.1268	0.1206	0.1392
	Best	0.1265	0.1272	0.1445	0.1272	0.1453	0.1272	0.1272	0.1272	0.1453
PRISM ( $\uparrow$ )	Mean	–	25.15	26.19	26.26	27.67	26.24	26.16	26.25	26.26
	Best	21.89	26.23	26.23	26.26	30.52	26.24	26.19	26.26	26.26
LLM-SQL ( $\uparrow$ )	Mean	–	0.7053	0.7127	0.7123	0.7231	0.7251	0.7129	0.7210	0.7278

Table 3: Main results for system problems. We report mean and best scores over three runs. “

\uparrow

” denotes maximization and “

\downarrow

” minimization. All methods (except human / prior AI) are run for 100 iterations using GPT-5 and Gemini-3.0-Pro. Bold denotes the best LLM-based open framework result per model. Green cells indicate matches or improvements over human SOTA results.

Panel of Figure 3 — Figure 3: Algorithm and Research Challenges. EvoX achieves the highest average private performance with GPT-5 on 10 different ALE-Bench-Lite problems and highest median score across 172 different Frontier-CS challenges.

Panel of Figure 4 — Figure 4: Search strategy evolution and cost-quality tradeoffs on the Heilbronn triangle task. (a) Dashed lines show fixed strategies, while solid lines show EvoX initialized from each strategy and allowed to evolve. Regardless of initialization, EvoX continues improving beyond the fixed strategy. (b) Cost-quality tradeoff under the GPT-5 model. To exceed a score of 0.031, EvoX and GEPA both require less than $1 in LLM generation cost, compared to ShinkaEvolve ($ 7.6) and OpenEvolve ($15.4).

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Problem	Objective	Description
Circle Packing (Square)	$\max\,r$	Place $n$ equal-radius circles inside the unit square without overlap; maximize the common radius subject to non-overlap and boundary constraints.
Circle Packing (Rectangle)	$\max\,r$	Pack $n$ equal-radius circles into a rectangle with fixed aspect ratio, maximizing the radius under geometric constraints.
Heilbronn Triangle	$\max\,\min$ area	Place $n$ points in $[0,1]^{2}$ to maximize the area of the smallest triangle formed by any three points.
Heilbronn Convex	$\max\,\min$ area	Generalization of the Heilbronn triangle problem: maximize the minimum area of convex hulls formed by any subset of $k>3$ points.
Min–Max–Min Distance (2 variants)	$\max\,\frac{d_{\min}}{d_{\max}}$	Place points to maximize uniformity, measured by the ratio between minimum and maximum pairwise distances.
Third Autocorrelation Inequality	$\min\,C_{3}$	Construct witness functions that minimize the third autocorrelation constant in additive combinatorics.

Appendix

Table 4: Mathematical optimization benchmarks. Summary of optimization tasks used to evaluate EvoX, including geometric, combinatorial, and signal processing problems.

Problem	Objective	Description
EPLB (Expert Placement Load Balancing)	$\max$ throughput ratio	Place and replicate experts in MoE models across GPUs based on activation frequency, minimizing load imbalance and computational skew.
PRISM (Global Model Scheduling)	$\max$ throughput	Schedule multiple deep learning models across distributed GPUs under memory, compute, and latency SLO constraints in a multi-tenant setting.
LLM–SQL	$\max$ accuracy	Optimize LLM-driven SQL generation via prompting or post-processing; a query is correct only if it exactly matches ground-truth execution results.
Cloudcast (Multi-Region Data Transfer)	$\min$ cost	Design cost-efficient data transfer strategies across cloud regions with heterogeneous bandwidths, latencies, and egress pricing under deadline constraints.
Transaction Scheduling	$\max$ commits/sec	Schedule database transactions under contention to maximize throughput while minimizing aborts from conflicts, deadlocks, or timeouts.
Telemetry Repair	$\max$ reconstruction accuracy	Reconstruct missing or corrupted telemetry signals using partial observations, exploiting temporal and spatial correlations.

Appendix

Table 5: System performance optimization benchmarks. Benchmarks drawn from ADRS-Bench [ 7 ] , capturing realistic optimization problems from production systems.

Problem / Category	Count	Description
ALE-Bench-Lite: NP-hard optimization from AtCoder Heuristic Contests
ahc008 (Territory)	1	Control agents on a grid to place fences isolating pets; $\max$ satisfaction.
ahc011 (Sliding Puzzle)	1	Rearrange tiles so line patterns form a large connected tree; $\max$ connectivity.
ahc015 (Candy Clustering)	1	Cluster candies of the same flavor using global tilt operations; $\max$ clustering score.
ahc016 (Graph Classification)	1	Design reference graphs and classify noisy, permuted query graphs; $\min$ classification error.
ahc024 (Map Compression)	1	Compress a grid map preserving district adjacency; $\min$ map size.

Appendix

Table 6: Algorithmic and research problem benchmarks. We evaluate on 182 open-ended problems: 10 NP-hard optimization problems from ALE-Bench-Lite [ 15 ] and 172 problems from FrontierCS [ 20 ] .

Backbone	OpenEvolve	EvoX
GPT-5	41%	48%
Gemini-3-Pro	45%	51%

Appendix

Table 7: EvoX performance on ARC-AGI benchmarks. Values denote final accuracy.

						GPT-5					Gemini-3.0-Pro
Task	Stat	Human	AlphaEvolve	ThetaEvolve	CodeEvolve	Random	OpenEvolve	GEPA	Shinka	EvoX	Random	OpenEvolve	GEPA	Shinka	EvoX
Circle Packing ( $\uparrow$ )	Mean	–	–	–	–	2.4905	2.5308	2.6129	2.4642	2.6324	2.5564	2.5414	2.6198	2.6216	2.6328
	Best	2.6340	2.6350	2.6360	2.6360	2.5000	2.5414	2.6285	2.5078	2.6360	2.5853	2.5414	2.6208	2.6358	2.6360
Circle Packing Rect ( $\uparrow$ )	Mean	–	–	–	–	2.2105	2.2673	2.3265	2.3347	2.3525	2.3432	2.3651	2.2160	2.3658	2.3658
	Best	2.3640	2.3658	–	2.3658	2.2783	2.2756	2.3537	2.3577	2.3600	2.3652	2.3658	2.2507	2.3658	2.3658
heilbronn_convex ( $\uparrow$ )	Mean	–	–	–	–	0.0185	0.0230	0.0259	0.0228	0.0263	0.0285	0.0279	0.0228	0.0279	0.0287

Appendix

Table 8: Main results for math optimization problems. We report mean and best over three runs. “

\uparrow

” / “

\downarrow

” indicate maximization / minimization.

			GPT-5					Gemini-3.0-Pro
Task	Stat	Human	Random	OpenEvolve	GEPA	Shinka	EvoX	Random	OpenEvolve	GEPA	Shinka	EvoX
EPLB ( $\uparrow$ )	Mean	–	0.1338	0.1300	0.1338	0.1185	0.1358	0.1280	0.1272	0.1268	0.1206	0.1392
	Best	0.1265	0.1445	0.1272	0.1445	0.1272	0.1453	0.1285	0.1272	0.1272	0.1272	0.1453
PRISM ( $\uparrow$ )	Mean	–	26.23	25.15	26.19	26.26	27.67	26.23	26.24	26.16	26.25	26.26
	Best	21.89	26.23	26.23	26.23	26.26	30.52	26.23	26.24	26.19	26.26	26.26
LLM-SQL ( $\uparrow$ )	Mean	–	0.7201	0.7053	0.7127	0.7123	0.7231	0.7380	0.7251	0.7129	0.7210	0.7278

Appendix

Table 9: Main results for system problems. We report mean and best scores over three runs. “

\uparrow

” denotes maximization and “

\downarrow

” minimization.

Task	Strategy	Search Mechanism	Variation Operator	$\boldsymbol{\Delta}$	$\boldsymbol{W}$
Signal Processing	Random sampling	Parents and inspiration programs sampled uniformly from the population	Free-form variation	+0.03	11
	Greedy sampling	Parents sampled from top-performing candidates	Free-form variation	+0.01	10
	Stratified + multi-objective sampling	Parents selected across objective-specific score rankings	Structural variation	+0.12	19
	UCB-based selection	Usage-based parent selection with penalty for overuse	Structural variation	+0.06	16
	Top-tier refinement	Parents restricted to high-score candidates	Local refinement	+0.03	14

Appendix

Table 10: Search strategy evolution on Signal Processing (Gemini, 100 iterations). Initial score 0.499

\rightarrow

final 0.743.

\Delta

: improvement achieved within the phase window;

W

: window length. Variation operator

\pi\in\{\textit{local refinement},\textit{structural variation},\textit{free-form variation}\}

, where free-form variation is the default.

Task	Strategy	Search Mechanism	Operator Bias	$\boldsymbol{\Delta}$	$\boldsymbol{W}$
Circle Packing	Random sampling	Parents and contexts sampled uniformly from the population	Free-form variation	+0.59	21
	Usage-penalized sampling	Parent selection penalizing recently sampled candidates	Structural variation	+0.04	17
	Tiered sampling	Parents selected from elite, mid, and lower score bands	Mixed (refinement + structural)	+0.004	10
	Stagnation-aware sampling	Parent selection conditioned on recent parent-child scores	Structural variation	+0.003	10
	Elite refinement	Parents restricted to high-score candidates	Local refinement	0	20
	Quantile-biased refinement	Parents sampled from upper quantiles of the score distribution	Local refinement	$\sim$ 2e $-$ 5	10

Appendix

Table 11: Search strategy evolution on Circle Packing (Gemini, 100 iterations). Initial score 0.364

\rightarrow

final 1.0004.

\Delta

: improvement achieved within the phase window;

W

: window length. Variation operator

\pi\in\{\textit{local refinement},\textit{structural variation},\textit{free-form variation}\}

, where free-form variation is the default.

Task	Strategy	Search Mechanism	Operator Bias	$\boldsymbol{\Delta}$	$\boldsymbol{W}$
Heilbronn Triangle	Stratified sampling	Parents selected from top and mid score bands	Free-form variation	+0.853	26
	Exploration-biased sampling	Parents sampled uniformly with occasional high-score selection	Structural variation	0	10
	Usage-penalized sampling	Parent selection penalizing recently sampled candidates	Mixed (refinement + structural)	+0.026	13
	Tiered refinement	Parents selected from top and mid score bands	Local refinement	+0.100	16
	Visit-weighted sampling	Parent selection weighted by historical sampling frequency	Free-form variation	0	10
	Refinement-focused sampling	Parents sampled from high-score candidates	Local refinement	+0.021	10

Appendix

Table 12: Search strategy evolution on Heilbronn Triangle (Gemini, 100 iterations). Initial score 0.0

\rightarrow

final

\texttt{min\_area}\approx 0.0365

(SOTA 1.0); best at iteration 91.

\Delta

: improvement achieved within the phase window;

W

: window length. Variation operator

\pi\in\{\textit{local refinement},\textit{structural variation},\textit{free-form variation}\}

, where free-form variation is the default.

Evolutionary Optimization

Evolutionary optimization searches for strong solutions through population-based variation, selection, and adaptation.

See all 285 papers

AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization
Feb 23, 2026Mert Cemri, Shubham Agrawal, Akshat Gupta +9Large Language Model-Guided Optimization LLMs for Evolutionary Optimization
University of California, Berkeley · Bespoke Labs
Evolutionary Multi-Task Optimization for LLM-Guided Program Discovery
May 21, 2026Halil Alperen Gozeten, Xuechen Zhang, Emrullah Ildiz +3Evolutionary Optimization Evolutionary Multitask Optimization
University of Michigan - Ann Arbor · University of California San Diego
TurboEvolve: Towards Fast and Robust LLM-Driven Program Evolution
Apr 12, 2026Yang Yang, Zining Zhong, Jindong Li +5Large Language Model-Guided Optimization Evolutionary Optimization
Artificial Intelligence Thrust, Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China.

Feb 23, 2026J/K move · Enter open · S save

AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization

Mert Cemri, Shubham Agrawal, Akshat Gupta, Shu Liu+8

University of California, Berkeley · Bespoke Labs

Large Language Model-Guided Optimization LLMs for Evolutionary Optimization Evolutionary Optimization LLM-Based Program Synthesis Zeroth-Order Optimization

Open paper AskPDF arXiv

The paradigm of automated program generation is shifting from one-shot generation to inference-time search, where Large Language Models (LLMs) function as semantic mutation operators within evolutionary loops. While effective, these systems are currently governed by static schedules that fail to account for the non-stationary dynamics of the search process. This rigidity results in substantial computational waste, as resources are indiscriminately allocated to stagnating populations while promising frontiers remain under-exploited. We introduce AdaEvolve, a framework that reformulates LLM-driven evolution as a hierarchical adaptive optimization problem. AdaEvolve uses an "accumulated improvement signal" to unify decisions across three levels: Local Adaptation, which dynamically modulates the exploration intensity within a population of solution candidates; Global Adaptation, which routes the global resource budget via bandit-based scheduling across different solution candidate populations; and Meta-Guidance which generates novel solution tactics based on the previously generated solutions and their corresponding improvements when the progress stalls. We demonstrate that AdaEvolve consistently outperforms the open-sourced baselines across 185 different open-ended optimization problems including combinatorial, systems optimization and algorithm design problems.

Abstract

Evolutionary Optimization

Explore similar work

AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization

Evolutionary Multi-Task Optimization for LLM-Guided Program Discovery

TurboEvolve: Towards Fast and Robust LLM-Driven Program Evolution