This systematic literature review examines the critical challenges and solutions related to scalability and maintainability in Machine Learning (ML) systems. As ML applications become increasingly complex and widespread across industries, the need to balance system scalability with long-term maintainability has emerged as a significant concern. This review synthesizes current research and practices addressing these dual challenges across the entire ML life-cycle, from data engineering to model deployment in production. We analyzed 124 papers to identify and categorize 41 maintainability challenges and 13 scalability challenges, along with their corresponding solutions. Our findings reveal intricate inter dependencies between scalability and maintainability, where improvements in one often impact the other. The review is structured around six primary research questions, examining maintainability and scalability challenges in data engineering, model engineering, and ML system development. We explore how these challenges manifest differently across various stages of the ML life-cycle. This comprehensive overview offers valuable insights for both researchers and practitioners in the field of ML systems. It aims to guide future research directions, inform best practices, and contribute to the development of more robust, efficient, and sustainable ML applications across various domains.

In this paper, we develop a systematic framework for the time-sequential compression of dynamic probabilistic occupancy grids. Our approach leverages ideas from signal compression theory to formulate an optimization problem that searches for a multi-resolution hierarchical encoder that balances the quality of the compressed map (distortion) with its description size, the latter of which relates to the bandwidth required to reliably transmit the map to other agents or to store map estimates in on-board memory. The resulting optimization problem allows for multi-resolution map compressions to be obtained that satisfy available communication or memory resources, and does not require knowledge of the occupancy map dynamics. We develop an algorithm to solve our problem, and demonstrate the utility of the proposed framework in simulation on both static (i.e., non-time varying) and dynamic (time-varying) occupancy maps.

This paper proposes a bidirectional rapidly-exploring random trees (RRT) algorithm to solve the motion planning problem for hybrid systems. The proposed algorithm, called HyRRT-Connect, propagates in both forward and backward directions in hybrid time until an overlap between the forward and backward propagation results is detected. Then, HyRRT-Connect constructs a motion plan through the reversal and concatenation of functions defined on hybrid time domains, ensuring that the motion plan satisfies the given hybrid dynamics. To address the potential discontinuity along the flow caused by tolerating some distance between the forward and backward partial motion plans, we reconstruct the backward partial motion plan by a forward-in-hybrid-time simulation from the final state of the forward partial motion plan. effectively eliminating the discontinuity. The proposed algorithm is applied to an actuated bouncing ball system and a walking robot example to highlight its computational improvement.

On-demand ride-pooling has emerged as a popular urban transportation solution, addressing the efficiency limitations of traditional ride-hailing services by grouping multiple riding requests with spatiotemporal proximity into a single vehicle. Although numerous algorithms have been developed for the Ride-pool Assignment Problem (RAP) -- a core component of ride-pooling systems, there is a lack of open-source implementations, making it difficult to benchmark these algorithms on a common dataset and objective. In this paper, we present the implementation details of a ride-pool simulator that encompasses several key ride-pool assignment algorithms, along with associated components such as vehicle routing and rebalancing. We also open-source a highly optimized and modular C++ codebase, designed to facilitate the extension of new algorithms and features. Additionally, we introduce a family of swapping-based local-search heuristics to enhance existing ride-pool assignment algorithms, achieving a better balance between performance and computational efficiency. Extensive experiments on a large-scale, real-world dataset from Manhattan, NYC reveal that while all selected algorithms perform comparably, the newly proposed Multi-Round Linear Assignment with Cyclic Exchange (LA-MR-CE) algorithm achieves a state-of-the-art service rate with significantly reduced computational time. Furthermore, an in-depth analysis suggests that a performance barrier exists for all myopic ride-pool assignment algorithms due to the system's capacity bottleneck, and incorporating future information could be key to overcoming this limitation.

Large reasoning models exhibit remarkable reasoning capabilities via long, elaborate reasoning trajectories. Supervised fine-tuning on such reasoning traces, also known as distillation, can be a cost-effective way to boost reasoning capabilities of student models. However, empirical observations reveal that these reasoning trajectories are often suboptimal, switching excessively between different lines of thought, resulting in under-thinking, over-thinking, and even degenerate responses. We introduce Retro-Search, an MCTS-inspired search algorithm, for distilling higher quality reasoning paths from large reasoning models. Retro-Search retrospectively revises reasoning paths to discover better, yet shorter traces, which can then lead to student models with enhanced reasoning capabilities with shorter, thus faster inference. Our approach can enable two use cases: self-improvement, where models are fine-tuned on their own Retro-Search-ed thought traces, and weak-to-strong improvement, where a weaker model revises stronger model's thought traces via Retro-Search. For self-improving, R1-distill-7B, fine-tuned on its own Retro-Search-ed traces, reduces the average reasoning length by 31.2% while improving performance by 7.7% across seven math benchmarks. For weak-to-strong improvement, we retrospectively revise R1-671B's traces from the OpenThoughts dataset using R1-distill-32B as the Retro-Search-er, a model 20x smaller. Qwen2.5-32B, fine-tuned on this refined data, achieves performance comparable to R1-distill-32B, yielding an 11.3% reduction in reasoning length and a 2.4% performance improvement compared to fine-tuning on the original OpenThoughts data. Our work counters recently emergent viewpoints that question the relevance of search algorithms in the era of large reasoning models, by demonstrating that there are still opportunities for algorithmic advancements, even for frontier models.

Automated Machine Learning (AutoML) automatically builds machine learning (ML) models on data. The de facto standard for evaluating new AutoML frameworks for tabular data is the AutoML Benchmark (AMLB). AMLB proposed to evaluate AutoML frameworks using 1-and 4-hour time budgets across 104 tasks. We argue that shorter time constraints should be considered for the benchmark because of their practical value, such as when models need to be retrained with high frequency, and to make AMLB more accessible. This work considers two ways in which to reduce the overall computation used in the benchmark: smaller time constraints and the use of early stopping. We conduct evaluations of 11 AutoML frameworks on 104 tasks with different time constraints and find the relative ranking of AutoML frameworks is fairly consistent across time constraints, but that using early-stopping leads to a greater variety in model performance. In Machine Learning (ML), manually creating good models is time-consuming and knowledge-intensive. Automated Machine Learning (AutoML) employs efficient automated search methods to create models for new data, often reducing the computational costs in the process Hutter et al. (2019); Hollmann et al. (2022). The AutoML Benchmark (AMLB, Gijsbers et al. 2024) has become the standard for the evaluation of AutoML frameworks on tabular data, greatly increasing reproducibility and comparability in AutoML research. We identified that the time budgets proposed by Gijsbers et al. (2024) were based on what seemed "practically reasonable" at the time, as signified by many frameworks' default time budget of one hour. While the authors motivate evaluating methods on two time budgets as a proxy for anytime performance, they do not motivate the particular choice of 1 hour and 4 hours. AutoML frameworks behave under different time constraints. We conduct similar experiments and analyses for frameworks with early-stopping, offering insights into its potential to reduce energy consumption in AutoML systems. However, we often see that the original benchmarking suite or time constraints (1 hour and 4 hour) are not used as proposed.

Recent advancements in Large Vision-Language Models have showcased remarkable capabilities. However, they often falter when confronted with complex reasoning tasks that humans typically address through visual aids and deliberate, step-by-step thinking. While existing methods have explored text-based slow thinking or rudimentary visual assistance, they fall short of capturing the intricate, interleaved nature of human visual-verbal reasoning processes. To overcome these limitations and inspired by the mechanisms of slow thinking in human cognition, we introduce VisuoThink, a novel framework that seamlessly integrates visuospatial and linguistic domains. VisuoThink facilitates multimodal slow thinking by enabling progressive visual-textual reasoning and incorporates test-time scaling through look-ahead tree search. Extensive experiments demonstrate that VisuoThink significantly enhances reasoning capabilities via inference-time scaling, even without fine-tuning, achieving state-of-the-art performance in tasks involving geometry and spatial reasoning.

Researchers recently extended Distributed Constraint Optimization Problems (DCOPs) to Communication-Aware DCOPs so that they are applicable in scenarios in which messages can be arbitrarily delayed. Distributed asynchronous local search and inference algorithms designed for CA-DCOPs are less vulnerable to message latency than their counterparts for regular DCOPs. However, unlike local search algorithms for (regular) DCOPs that converge to k-opt solutions (with k > 1), that is, they converge to solutions that cannot be improved by a group of k agents), local search CA-DCOP algorithms are limited to 1-opt solutions only. In this paper, we introduce Latency-Aware Monotonic Distributed Local Search-2 (LAMDLS-2), where agents form pairs and coordinate bilateral assignment replacements. LAMDLS-2 is monotonic, converges to a 2-opt solution, and is also robust to message latency, making it suitable for CA-DCOPs. Our results indicate that LAMDLS-2 converges faster than MGM-2, a benchmark algorithm, to a similar 2-opt solution, in various message latency scenarios.

Oscillator Ising machines (OIMs) represent an exemplar case of using physics-inspired non-linear dynamical systems to solve computationally challenging combinatorial optimization problems (COPs). The computational performance of such systems is highly sensitive to the underlying dynamical properties, the topology of the input graph, and their relative compatibility. In this work, we explore the concept of designing different dynamical systems that minimize the same objective function but exhibit drastically different dynamical properties. Our goal is to leverage this diversification in dynamics to reduce the sensitivity of the computational performance to the underlying graph, and subsequently, enhance the overall effectiveness of such physics-based computational methods. To this end, we introduce a novel dynamical system, the Dynamical Ising Machine (DIM), which, like the OIM, minimizes the Ising Hamiltonian but offers significantly different dynamical properties. We analyze the characteristic properties of the DIM and compare them with those of the OIM. We also show that the relative performance of each model is dependent on the input graph. Our work illustrates that using multiple dynamical systems with varying properties to solve the same COP enables an effective method that is less sensitive to the input graph, while producing robust solutions.

The emergence of large language models (LLMs) offers new possibilities for structured exploration of scientific knowledge. Rather than viewing scientific discovery as isolated ideas or content, we propose a structured approach that emphasizes the role of method combinations in shaping disruptive insights. Specifically, we investigate how knowledge units--especially those tied to methodological design--can be modeled and recombined to yield research breakthroughs. Our proposed framework addresses two key challenges. First, we introduce a contrastive learning-based mechanism to identify distinguishing features of historically disruptive method combinations within problem-driven contexts. Second, we propose a reasoning-guided Monte Carlo search algorithm that leverages the chain-of-thought capability of LLMs to identify promising knowledge recombinations for new problem statements. Empirical studies across multiple domains show that the framework is capable of modeling the structural dynamics of innovation and successfully highlights combinations with high disruptive potential. This research provides a new path for computationally guided scientific ideation grounded in structured reasoning and historical data modeling.