Xin Qiu

Wenjie Liu, Hao Wu, Xin Qiu et al.

Modern multimodal large language models (MLLMs) adopt a unified self-attention design that processes visual and textual tokens at every Transformer layer, incurring substantial computational overhead. In this work, we revisit the necessity of such dense visual processing and show that projected visual embeddings are already well-aligned with the language space, while effective vision-language interaction occurs in only a small subset of layers. Based on these insights, we propose ViCA (Vision-only Cross-Attention), a minimal MLLM architecture in which visual tokens bypass all self-attention and feed-forward layers, interacting with text solely through sparse cross-attention at selected layers. Extensive evaluations across three MLLM backbones, nine multimodal benchmarks, and 26 pruning-based baselines show that ViCA preserves 98% of baseline accuracy while reducing visual-side computation to 4%, consistently achieving superior performance-efficiency trade-offs. Moreover, ViCA provides a regular, hardware-friendly inference pipeline that yields over 3.5x speedup in single-batch inference and over 10x speedup in multi-batch inference, reducing visual grounding to near-zero overhead compared with text-only LLMs. It is also orthogonal to token pruning methods and can be seamlessly combined for further efficiency gains. Our code is available at https://github.com/EIT-NLP/ViCA.

Jason Dai, Ding Ding, Dongjie Shi et al.

Most AI projects start with a Python notebook running on a single laptop; however, one usually needs to go through a mountain of pains to scale it to handle larger dataset (for both experimentation and production deployment). These usually entail many manual and error-prone steps for the data scientists to fully take advantage of the available hardware resources (e.g., SIMD instructions, multi-processing, quantization, memory allocation optimization, data partitioning, distributed computing, etc.). To address this challenge, we have open sourced BigDL 2.0 at https://github.com/intel-analytics/BigDL/ under Apache 2.0 license (combining the original BigDL and Analytics Zoo projects); using BigDL 2.0, users can simply build conventional Python notebooks on their laptops (with possible AutoML support), which can then be transparently accelerated on a single node (with up-to 9.6x speedup in our experiments), and seamlessly scaled out to a large cluster (across several hundreds servers in real-world use cases). BigDL 2.0 has already been adopted by many real-world users (such as Mastercard, Burger King, Inspur, etc.) in production.

Kajetan Schweighofer, Conor F. Hayes, Roberto Dailey et al.

Evolution Strategies (ES) has recently emerged as a competitive alternative to reinforcement learning (RL) for large language model (LLM) fine-tuning, offering advantages through simplicity, scalability, and inference-only training. However, recent work suggests that ES fine-tuning on new tasks may induce forgetting of prior tasks. First, this paper shows that prior task forgetting (1) is better characterized as performance drift rather than irreversible forgetting, with prior-task performance often recovering during ES training; and (2) is not a specific failure mode of ES, but can also arise for fine-tuning with RL methods. Second, it analyzes when and why such drift arises, highlighting its dependence on ES training dynamics, particularly random walk behavior in weakly constrained directions of the weight space. Third, based on these insights, it introduces Anchored Weight Decay (AWD) as a parameter-space regularization technique that constrains optimization toward the initial model parameters. AWD effectively stabilizes prior-task performance while preserving target-task performance, achieving benefits comparable to large ES population sizes at much lower computational cost. Thus, contrary to previous beliefs, the paper shows that prior-task forgetting under ES is largely avoidable, positioning ES as a promising approach for continual learning in LLMs.

Xin Qiu, Risto Miikkulainen

Population-based search has recently emerged as a possible alternative to Reinforcement Learning (RL) for black-box neural architecture search (NAS). It performs well in practice even though it is not theoretically well understood. In particular, whereas traditional population-based search methods such as evolutionary algorithms (EAs) draw much power from crossover operations, it is difficult to take advantage of them in NAS. The main obstacle is believed to be the permutation problem: The mapping between genotype and phenotype in traditional graph representations is many-to-one, leading to a disruptive effect of standard crossover. This paper presents the first theoretical analysis of the behaviors of mutation, crossover and RL in black-box NAS, and proposes a new crossover operator based on the shortest edit path (SEP) in graph space. The SEP crossover is shown theoretically to overcome the permutation problem, and as a result, have a better expected improvement compared to mutation, standard crossover and RL. Further, it empirically outperform these other methods on state-of-the-art NAS benchmarks. The SEP crossover therefore allows taking full advantage of population-based search in NAS, and the underlying theory can serve as a foundation for deeper understanding of black-box NAS methods in general.

Qiyao Liang, Jinyeop Song, Yizhou Liu et al.

Weight-perturbation evolution strategies (ES) can fine-tune billion-parameter language models with surprisingly small populations (e.g., $N\!\approx\!30$), contradicting classical zeroth-order curse-of-dimensionality intuition. We also observe a second seemingly separate phenomenon: under fixed hyperparameters, the stochastic fine-tuning reward often rises, peaks, and then degrades in both ES and GRPO. We argue that both effects reflect a shared geometric property of fine-tuning landscapes: they are low-dimensional in curvature. A small set of high-curvature dimensions dominates improvement, producing (i) heterogeneous time scales that yield rise-then-decay under fixed stochasticity, as captured by a minimal quadratic stochastic-ascent model, and (ii) degenerate improving updates, where many random perturbations share similar components along these directions. Using ES as a geometric probe on fine-tuning reward landscapes of GSM8K, ARC-C, and WinoGrande across Qwen2.5-Instruct models (0.5B--7B), we show that reward-improving perturbations remain empirically accessible with small populations across scales. Together, these results reconcile ES scalability with non-monotonic training dynamics and suggest that high-dimensional fine-tuning may admit a broader class of viable optimization methods than worst-case theory implies.

Xin Qiu, Junlong Tong, Yirong Sun et al.

Large language models (LLMs) have been introduced to time series forecasting (TSF) to incorporate contextual knowledge beyond numerical signals. However, existing studies question whether LLMs provide genuine benefits, often reporting comparable performance without LLMs. We show that such conclusions stem from limited evaluation settings and do not hold at scale. We conduct a large-scale study of LLM-based TSF (LLM4TSF) across 8 billion observations, 17 forecasting scenarios, 4 horizons, multiple alignment strategies, and both in-domain and out-of-domain settings. Our results demonstrate that \emph{LLM4TS indeed improves forecasting performance}, with especially large gains in cross-domain generalization. Pre-alignment outperforming post-alignment in over 90\% of tasks. Both pretrained knowledge and model architecture of LLMs contribute and play complementary roles: pretraining is critical under distribution shifts, while architecture excels at modeling complex temporal dynamics. Moreover, under large-scale mixed distributions, a fully intact LLM becomes indispensable, as confirmed by token-level routing analysis and prompt-based improvements. Overall, Our findings overturn prior negative assessments, establish clear conditions under which LLMs are not only useful, and provide practical guidance for effective model design. We release our code at https://github.com/EIT-NLP/LLM4TSF.

Yinggan Xu, Risto Miikkulainen, Xin Qiu

Post-Training Quantization (PTQ) is essential for deploying Large Language Models (LLMs) on memory-constrained devices, yet it renders models static and difficult to fine-tune. Standard fine-tuning paradigms, including Reinforcement Learning (RL), fundamentally rely on backpropagation and high-precision weights to compute gradients. Thus they cannot be used on quantized models, where the parameter space is discrete and non-differentiable. While Evolution Strategies (ES) offer a backpropagation-free alternative, optimization of the quantized parameters can still fail due to vanishing or inaccurate gradient. This paper introduces Quantized Evolution Strategies (QES), an optimization paradigm that performs full-parameter fine-tuning directly in the quantized space. QES is based on two innovations: (1) it integrates accumulated error feedback to preserve high-precision gradient signals, and (2) it utilizes a stateless seed replay to reduce memory usage to low-precision inference levels. QES significantly outperforms the state-of-the-art zeroth-order fine-tuning method on arithmetic reasoning tasks, making direct fine-tuning for quantized models possible. It therefore opens up the possibility for scaling up LLMs entirely in the quantized space. The source code is available at https://github.com/dibbla/Quantized-Evolution-Strategies .

Yirong Sun, Yanjun Chen, Xin Qiu et al.

Large Audio Language Models (LALMs) excel at semantic and paralinguistic tasks, yet their ability to perceive the fundamental physical attributes of audio such as pitch, loudness, and spatial location remains under-explored. To bridge this gap, we introduce SonicBench, a psychophysically grounded benchmark that systematically evaluates 12 core physical attributes across five perceptual dimensions. Unlike previous datasets, SonicBench uses a controllable generation toolbox to construct stimuli for two complementary paradigms: recognition (absolute judgment) and comparison (relative judgment). This design allows us to probe not only sensory precision but also relational reasoning capabilities, a domain where humans typically exhibit greater proficiency. Our evaluation reveals a substantial deficiency in LALMs' foundational auditory understanding; most models perform near random guessing and, contrary to human patterns, fail to show the expected advantage on comparison tasks. Furthermore, explicit reasoning yields minimal gains. However, our linear probing analysis demonstrates crucially that frozen audio encoders do successfully capture these physical cues (accuracy at least 60%), suggesting that the primary bottleneck lies in the alignment and decoding stages, where models fail to leverage the sensory signals they have already captured.

Elliot Meyerson, Giuseppe Paolo, Roberto Dailey et al.

LLMs have achieved remarkable breakthroughs in reasoning, insights, and tool use, but chaining these abilities into extended processes at the scale of those routinely executed by humans, organizations, and societies has remained out of reach. The models have a persistent error rate that prevents scale-up: for instance, recent experiments in the Towers of Hanoi benchmark domain showed that the process inevitably becomes derailed after at most a few hundred steps. Thus, although LLM research is often still benchmarked on tasks with relatively few dependent logical steps, there is increasing attention on the ability (or inability) of LLMs to perform long range tasks. This paper describes MAKER, the first system that successfully solves a task with over one million LLM steps with zero errors, and, in principle, scales far beyond this level. The approach relies on an extreme decomposition of a task into subtasks, each of which can be tackled by focused microagents. The high level of modularity resulting from the decomposition allows error correction to be applied at each step through an efficient multi-agent voting scheme. This combination of extreme decomposition and error correction makes scaling possible. Thus, the results suggest that instead of relying on continual improvement of current LLMs, massively decomposed agentic processes (MDAPs) may provide a way to efficiently solve problems at the level of organizations and societies.

Xin Qiu, Yulu Gan, Conor F. Hayes et al. · pku

Fine-tuning pre-trained large language models (LLMs) for down-stream tasks is a critical step in the AI deployment pipeline. Reinforcement learning (RL) is arguably the most prominent fine-tuning method, contributing to the birth of many state-of-the-art LLMs. In contrast, evolution strategies (ES), which once showed comparable performance to RL on models with a few million parameters, was neglected due to the pessimistic perception of its scalability to larger models. In this work, we report the first successful attempt to scale up ES for fine-tuning the full parameters of LLMs, showing the surprising fact that ES can search efficiently over billions of parameters and outperform existing RL fine-tuning methods in multiple respects, including sample efficiency, tolerance to long-horizon rewards, robustness to different base LLMs, less tendency to reward hacking, and more stable performance across runs. It therefore serves as a basis to unlock a new direction in LLM fine-tuning beyond what current RL techniques provide. The source codes are provided at: https://github.com/VsonicV/es-fine-tuning-paper.

Luqing Luo, Wenjin Gui, Yunfei Liu et al.

Deep understanding of electromagnetic signals is fundamental to dynamic spectrum management, intelligent transportation, autonomous driving and unmanned vehicle perception. The field faces challenges because electromagnetic signals differ greatly from text and images, showing high heterogeneity, strong background noise and complex joint time frequency structure, which prevents existing general models from direct use. Electromagnetic communication and sensing tasks are diverse, current methods lack cross task generalization and transfer efficiency, and the scarcity of large high quality datasets blocks the creation of a truly general multitask learning framework. To overcome these issue, we introduce EMind, an electromagnetic signals foundation model that bridges large scale pretraining and the unique nature of this modality. We build the first unified and largest standardized electromagnetic signal dataset covering multiple signal types and tasks. By exploiting the physical properties of electromagnetic signals, we devise a length adaptive multi-signal packing method and a hardware-aware training strategy that enable efficient use and representation learning from heterogeneous multi-source signals. Experiments show that EMind achieves strong performance and broad generalization across many downstream tasks, moving decisively from task specific models to a unified framework for electromagnetic intelligence. The code is available at: https://github.com/GabrielleTse/EMind.

Xin Qiu, Junlong Tong, Yirong Sun et al.

Large-scale models are at the forefront of time series (TS) forecasting, dominated by two paradigms: fine-tuning text-based Large Language Models (LLM4TS) and training Time Series Foundation Models (TSFMs) from scratch. Both approaches share a foundational assumption that scaling up model capacity and data volume leads to improved performance. However, we observe a \textit{\textbf{scaling paradox}} in TS models, revealing a puzzling phenomenon that larger models do \emph{NOT} achieve better performance. Through extensive experiments on two model families across four scales (100M to 1.7B parameters) and diverse data (up to 6B observations), we rigorously confirm that the scaling paradox is a pervasive issue. We then diagnose its root cause by analyzing internal representations, identifying a phenomenon we call \textit{few-layer dominance}: only a small subset of layers are functionally important, while the majority are redundant, under-utilized, and can even distract training. Based on this discovery, we propose a practical method to automatically identify and retain only these dominant layers. In our models, retaining only 21\% of the parameters achieves up to a 12\% accuracy improvement and a 2.7$\times$ inference speedup. We validate the universality of our method on 8 prominent SOTA models (LLM4TS and TSFMs, 90M to 6B), showing that retaining less than 30\% of layers achieves comparable or superior accuracy in over 95\% of tasks.

Xin Qiu, Risto Miikkulainen

With the widespread application of Large Language Models (LLMs) to various domains, concerns regarding the trustworthiness of LLMs in safety-critical scenarios have been raised, due to their unpredictable tendency to hallucinate and generate misinformation. Existing LLMs do not have an inherent functionality to provide the users with an uncertainty/confidence metric for each response it generates, making it difficult to evaluate trustworthiness. Although several studies aim to develop uncertainty quantification methods for LLMs, they have fundamental limitations, such as being restricted to classification tasks, requiring additional training and data, considering only lexical instead of semantic information, and being prompt-wise but not response-wise. A new framework is proposed in this paper to address these issues. Semantic density extracts uncertainty/confidence information for each response from a probability distribution perspective in semantic space. It has no restriction on task types and is "off-the-shelf" for new models and tasks. Experiments on seven state-of-the-art LLMs, including the latest Llama 3 and Mixtral-8x22B models, on four free-form question-answering benchmarks demonstrate the superior performance and robustness of semantic density compared to prior approaches.

Elliot Meyerson, Xin Qiu

Decomposing hard problems into subproblems often makes them easier and more efficient to solve. With large language models (LLMs) crossing critical reliability thresholds for a growing slate of capabilities, there is an increasing effort to decompose systems into sets of LLM-based agents, each of whom can be delegated sub-tasks. However, this decomposition (even when automated) is often intuitive, e.g., based on how a human might assign roles to members of a human team. How close are these role decompositions to optimal? This position paper argues that asymptotic analysis with LLM primitives is needed to reason about the efficiency of such decomposed systems, and that insights from such analysis will unlock opportunities for scaling them. By treating the LLM forward pass as the atomic unit of computational cost, one can separate out the (often opaque) inner workings of a particular LLM from the inherent efficiency of how a set of LLMs are orchestrated to solve hard problems. In other words, if we want to scale the deployment of LLMs to the limit, instead of anthropomorphizing LLMs, asymptotic analysis with LLM primitives should be used to reason about and develop more powerful decompositions of large problems into LLM agents.

Xinzhu Liang, Joseph M. Lukens, Sanjaya Lohani et al.

This work systematically compares parallel implementations of consistent (asymptotically unbiased) Bayesian deep learning algorithms: sequential Monte Carlo sampler (SMC$_\parallel$) or Markov chain Monte Carlo (MCMC$_\parallel$). We provide a proof of convergence for SMC$_\parallel$ showing that it theoretically achieves the same level of convergence as a single monolithic SMC sampler, while the reduced communication lowers wall-clock time. It is well-known that the first samples from MCMC need to be discarded to eliminate initialization bias, and that the number of discarded samples must grow like the logarithm of the number of parallel chains to control that bias for MCMC$_\parallel$. A systematic empirical numerical study on MNIST, CIFAR, and IMDb, reveals that parallel implementations of both methods perform comparably to non-parallel implementations in terms of performance and total cost, and also comparably to each other. However, both methods still require a large wall-clock time, and suffer from catastrophic non-convergence if they aren't run for long enough.

Sangjun Park, Elliot Meyerson, Xin Qiu et al.

Metacognition is a critical component of intelligence, specifically regarding the awareness of one's own knowledge. While humans rely on shared internal memory for both answering questions and reporting their knowledge state, this dependency in LLMs remains underexplored. This study proposes a framework to measure metacognitive ability $d_{\rm{type2}}'$ using a dual-prompt method, followed by the introduction of Evolution Strategy for Metacognitive Alignment (ESMA) to bind a model's internal knowledge to its explicit behaviors. ESMA demonstrates robust generalization across diverse untrained settings, indicating a enhancement in the model's ability to reference its own knowledge. Furthermore, parameter analysis attributes these improvements to a sparse set of significant modifications.

Xinzhu Liang, Joseph M. Lukens, Sanjaya Lohani et al.

This work introduces a new method designed for Bayesian deep learning called scalable Bayesian Monte Carlo (SBMC). The method is comprised of a model and an algorithm. The model interpolates between a point estimator and the posterior. The algorithm is a parallel implementation of sequential Monte Carlo sampler (SMC$_\parallel$) or Markov chain Monte Carlo (MCMC$_\parallel$). We collectively refer to these consistent (asymptotically unbiased) algorithms as Bayesian Monte Carlo (BMC), and any such algorithm can be used in our SBMC method. The utility of the method is demonstrated on practical examples: MNIST, CIFAR, IMDb. A systematic numerical study reveals that for the same wall-clock time as state-of-the-art (SOTA) methods like deep ensembles (DE), SBMC achieves comparable or better accuracy and substantially improved uncertainty quantification (UQ)--in particular, epistemic UQ. This is demonstrated on the downstream task of estimating the confidence in predictions, which can be used for reliability assessment or abstention decisions.

Elliot Meyerson, Xin Qiu, Risto Miikkulainen

This paper characterizes the inherent power of evolutionary algorithms. This power depends on the computational properties of the genetic encoding. With some encodings, two parents recombined with a simple crossover operator can sample from an arbitrary distribution of child phenotypes. Such encodings are termed \emph{expressive encodings} in this paper. Universal function approximators, including popular evolutionary substrates of genetic programming and neural networks, can be used to construct expressive encodings. Remarkably, this approach need not be applied only to domains where the phenotype is a function: Expressivity can be achieved even when optimizing static structures, such as binary vectors. Such simpler settings make it possible to characterize expressive encodings theoretically: Across a variety of test problems, expressive encodings are shown to achieve up to super-exponential convergence speed-ups over the standard direct encoding. The conclusion is that, across evolutionary computation areas as diverse as genetic programming, neuroevolution, genetic algorithms, and theory, expressive encodings can be a key to understanding and realizing the full power of evolution.

Xin Qiu, Risto Miikkulainen

As neural network classifiers are deployed in real-world applications, it is crucial that their failures can be detected reliably. One practical solution is to assign confidence scores to each prediction, then use these scores to filter out possible misclassifications. However, existing confidence metrics are not yet sufficiently reliable for this role. This paper presents a new framework that produces a quantitative metric for detecting misclassification errors. This framework, RED, builds an error detector on top of the base classifier and estimates uncertainty of the detection scores using Gaussian Processes. Experimental comparisons with other error detection methods on 125 UCI datasets demonstrate that this approach is effective. Further implementations on two probabilistic base classifiers and two large deep learning architecture in vision tasks further confirm that the method is robust and scalable. Third, an empirical analysis of RED with out-of-distribution and adversarial samples shows that the method can be used not only to detect errors but also to understand where they come from. RED can thereby be used to improve trustworthiness of neural network classifiers more broadly in the future.

Santiago Gonzalez, Xin Qiu, Risto Miikkulainen

Evolutionary computation can be used to optimize several different aspects of neural network architectures. For instance, the TaylorGLO method discovers novel, customized loss functions, resulting in improved performance, faster training, and improved data utilization. A likely reason is that such functions discourage overfitting, leading to effective regularization. This paper demonstrates theoretically that this is indeed the case for TaylorGLO. Learning rule decomposition reveals that evolved loss functions balance two factors: the pull toward zero error, and a push away from it to avoid overfitting. This is a general principle that may be used to understand other regularization techniques as well (as demonstrated in this paper for label smoothing). The theoretical analysis leads to a constraint that can be utilized to find more effective loss functions in practice; the mechanism also results in networks that are more robust (as demonstrated in this paper with adversarial inputs). The analysis in this paper thus constitutes a first step towards understanding regularization, and demonstrates the power of evolutionary neural architecture search in general.

Risto Miikkulainen, Olivier Francon, Elliot Meyerson et al.

Several models have been developed to predict how the COVID-19 pandemic spreads, and how it could be contained with non-pharmaceutical interventions (NPIs) such as social distancing restrictions and school and business closures. This paper demonstrates how evolutionary AI could be used to facilitate the next step, i.e. determining most effective intervention strategies automatically. Through evolutionary surrogate-assisted prescription (ESP), it is possible to generate a large number of candidate strategies and evaluate them with predictive models. In principle, strategies can be customized for different countries and locales, and balance the need to contain the pandemic and the need to minimize their economic impact. While still limited by available data, early experiments suggest that workplace and school restrictions are the most important and need to be designed carefully. It also demonstrates that results of lifting restrictions can be unreliable, and suggests creative ways in which restrictions can be implemented softly, e.g. by alternating them over time. As more data becomes available, the approach can be increasingly useful in dealing with COVID-19 as well as possible future pandemics.

Olivier Francon, Santiago Gonzalez, Babak Hodjat et al.

There is now significant historical data available on decision making in organizations, consisting of the decision problem, what decisions were made, and how desirable the outcomes were. Using this data, it is possible to learn a surrogate model, and with that model, evolve a decision strategy that optimizes the outcomes. This paper introduces a general such approach, called Evolutionary Surrogate-Assisted Prescription, or ESP. The surrogate is, for example, a random forest or a neural network trained with gradient descent, and the strategy is a neural network that is evolved to maximize the predictions of the surrogate model. ESP is further extended in this paper to sequential decision-making tasks, which makes it possible to evaluate the framework in reinforcement learning (RL) benchmarks. Because the majority of evaluations are done on the surrogate, ESP is more sample efficient, has lower variance, and lower regret than standard RL approaches. Surprisingly, its solutions are also better because both the surrogate and the strategy network regularize the decision-making behavior. ESP thus forms a promising foundation to decision optimization in real-world problems.

Xin Qiu, Elliot Meyerson, Risto Miikkulainen

Neural Networks (NNs) have been extensively used for a wide spectrum of real-world regression tasks, where the goal is to predict a numerical outcome such as revenue, effectiveness, or a quantitative result. In many such tasks, the point prediction is not enough: the uncertainty (i.e. risk or confidence) of that prediction must also be estimated. Standard NNs, which are most often used in such tasks, do not provide uncertainty information. Existing approaches address this issue by combining Bayesian models with NNs, but these models are hard to implement, more expensive to train, and usually do not predict as accurately as standard NNs. In this paper, a new framework (RIO) is developed that makes it possible to estimate uncertainty in any pretrained standard NN. The behavior of the NN is captured by modeling its prediction residuals with a Gaussian Process, whose kernel includes both the NN's input and its output. The framework is evaluated in twelve real-world datasets, where it is found to (1) provide reliable estimates of uncertainty, (2) reduce the error of the point predictions, and (3) scale well to large datasets. Given that RIO can be applied to any standard NN without modifications to model architecture or training pipeline, it provides an important ingredient for building real-world NN applications.

Jason Dai, Yiheng Wang, Xin Qiu et al.

This paper presents BigDL (a distributed deep learning framework for Apache Spark), which has been used by a variety of users in the industry for building deep learning applications on production big data platforms. It allows deep learning applications to run on the Apache Hadoop/Spark cluster so as to directly process the production data, and as a part of the end-to-end data analysis pipeline for deployment and management. Unlike existing deep learning frameworks, BigDL implements distributed, data parallel training directly on top of the functional compute model (with copy-on-write and coarse-grained operations) of Spark. We also share real-world experience and "war stories" of users that have adopted BigDL to address their challenges(i.e., how to easily build end-to-end data analysis and deep learning pipelines for their production data).

Xin Qiu, Risto Miikkulainen

Conversion rate optimization means designing web interfaces such that more visitors perform a desired action (such as register or purchase) on the site. One promising approach, implemented in Sentient Ascend, is to optimize the design using evolutionary algorithms, evaluating each candidate design online with actual visitors. Because such evaluations are costly and noisy, several challenges emerge: How can available visitor traffic be used most efficiently? How can good solutions be identified most reliably? How can a high conversion rate be maintained during optimization? This paper proposes a new technique to address these issues. Traffic is allocated to candidate solutions using a multi-armed bandit algorithm, using more traffic on those evaluations that are most useful. In a best-arm identification mode, the best candidate can be identified reliably at the end of evolution, and in a campaign mode, the overall conversion rate can be optimized throughout the entire evolution process. Multi-armed bandit algorithms thus improve performance and reliability of machine discovery in noisy real-world environments.