Search

A central question in multilingual language modeling is whether large language models (LLMs) develop a universal concept representation, disentangled from specific languages. In this paper, we address this question by analyzing latent representations (latents) during a word translation task in transformer-based LLMs. We strategically extract latents from a source translation prompt and insert them into the forward pass on a target translation prompt. By doing so, we find that the output language is encoded in the latent at an earlier layer than the concept to be translated. Building on this insight, we conduct two key experiments. First, we demonstrate that we can change the concept without changing the language and vice versa through activation patching alone. Second, we show that patching with the mean over latents across different languages does not impair and instead improves the models' performance in translating the concept. Our results provide evidence for the existence of language-agnostic concept representations within the investigated models., Comment: 12 pages, 10 figures, previous version published under the title "How Do Llamas Process Multilingual Text? A Latent Exploration through Activation Patching" at the ICML 2024 mechanistic interpretability workshop at https://openreview.net/forum?id=0ku2hIm4BS

In business process simulation, resource availability is typically modeled by assigning a calendar to each resource, e.g., Monday-Friday, 9:00-18:00. Resources are assumed to be always available during each time slot in their availability calendar. This assumption often becomes invalid due to interruptions, breaks, or time-sharing across processes. In other words, existing approaches fail to capture intermittent availability. Another limitation of existing approaches is that they either do not consider multitasking behavior, or if they do, they assume that resources always multitask (up to a maximum capacity) whenever available. However, studies have shown that the multitasking patterns vary across days. This paper introduces a probabilistic approach to model resource availability and multitasking behavior for business process simulation. In this approach, each time slot in a resource calendar has an associated availability probability and a multitasking probability per multitasking level. For example, a resource may be available on Fridays between 14:00-15:00 with 90\% probability, and given that they are performing one task during this slot, they may take on a second concurrent task with 60\% probability. We propose algorithms to discover probabilistic calendars and probabilistic multitasking capacities from event logs. An evaluation shows that, with these enhancements, simulation models discovered from event logs better replicate the distribution of activities and cycle times, relative to approaches with crisp calendars and monotasking assumptions.

The notion of double depth associated with quasi-Jacobi forms allows distinguishing, within the algebra of quasi-Jacobi singular forms of index zero, certain significant subalgebras (modular-type forms, elliptic-type forms, Jacobi forms). We study the stability of these subalgebras under the derivations of this algebra and through certain sequences of bidifferential operators constituting analogs of Rankin-Cohen brackets or transvectants

Today's most advanced multimodal models remain proprietary. The strongest open-weight models rely heavily on synthetic data from proprietary VLMs to achieve good performance, effectively distilling these closed models into open ones. As a result, the community is still missing foundational knowledge about how to build performant VLMs from scratch. We present Molmo, a new family of VLMs that are state-of-the-art in their class of openness. Our key innovation is a novel, highly detailed image caption dataset collected entirely from human annotators using speech-based descriptions. To enable a wide array of user interactions, we also introduce a diverse dataset mixture for fine-tuning that includes in-the-wild Q&A and innovative 2D pointing data. The success of our approach relies on careful choices for the model architecture details, a well-tuned training pipeline, and, most critically, the quality of our newly collected datasets, all of which will be released. The best-in-class 72B model within the Molmo family not only outperforms others in the class of open weight and data models but also compares favorably against proprietary systems like GPT-4o, Claude 3.5, and Gemini 1.5 on both academic benchmarks and human evaluation. We will be releasing all of our model weights, captioning and fine-tuning data, and source code in the near future. Select model weights, inference code, and demo are available at https://molmo.allenai.org.

Achieving human-like memory recall in artificial systems remains a challenging frontier in computer vision. Humans demonstrate remarkable ability to recall images after a single exposure, even after being shown thousands of images. However, this capacity diminishes significantly when confronted with non-natural stimuli such as random textures. In this paper, we present a method inspired by human memory processes to bridge this gap between artificial and biological memory systems. Our approach focuses on encoding images to mimic the high-level information retained by the human brain, rather than storing raw pixel data. By adding noise to images before encoding, we introduce variability akin to the non-deterministic nature of human memory encoding. Leveraging pre-trained models' embedding layers, we explore how different architectures encode images and their impact on memory recall. Our method achieves impressive results, with 97% accuracy on natural images and near-random performance (52%) on textures. We provide insights into the encoding process and its implications for machine learning memory systems, shedding light on the parallels between human and artificial intelligence memory mechanisms., Comment: 5 pages, 7 figures

We consider $G$ a semisimple Lie group with finite center and $K$ a maximal compact subgroup of $G$. We study the regularity of $K$-finite matrix coefficients of unitary representations of $G$. More precisely, we find the optimal value $\kappa(G)$ such that all such coefficients are $\kappa(G)$-H\"older continuous. The proof relies on analysis of spherical functions of the symmetric Gelfand pair $(G,K)$, using stationary phase estimates from Duistermaat, Kolk and Varadarajan. If $U$ is a compact form of $G$, then $(U,K)$ is a compact symmetric pair. Using the same tools, we study the regularity of $K$-finite coefficients of unitary representations of $U$, improving on previous results obtained by the author.

This paper addresses the following problem: Given a process model and an event log containing trace prefixes of ongoing cases of a process, map each case to its corresponding state (i.e., marking) in the model. This state computation operation is a building block of other process mining operations, such as log animation and short-term simulation. An approach to this state computation problem is to perform a token-based replay of each trace prefix against the model. However, when a trace prefix does not strictly follow the behavior of the process model, token replay may produce a state that is not reachable from the initial state of the process. An alternative approach is to first compute an alignment between the trace prefix of each ongoing case and the model, and then replay the aligned trace prefix. However, (prefix-)alignment is computationally expensive. This paper proposes a method that, given a trace prefix of an ongoing case, computes its state in constant time using an index that represents states as n-grams. An empirical evaluation shows that the proposed approach has an accuracy comparable to that of the prefix-alignment approach, while achieving a throughput of hundreds of thousands of traces per second.

Simulation is a common approach to predict the effect of business process changes on quantitative performance. The starting point of Business Process Simulation (BPS) is a process model enriched with simulation parameters. To cope with the typically large parameter spaces of BPS models, several methods have been proposed to automatically discover BPS models from event logs. Virtually all these approaches neglect the data perspective of business processes. Yet, the data attributes manipulated by a business process often determine which activities are performed, how many times, and when. This paper addresses this gap by introducing a data-aware BPS modeling approach and a method to discover data-aware BPS models from event logs. The BPS modeling approach supports three types of data attributes (global, case-level, and event-level) as well as deterministic and stochastic attribute update rules and data-aware branching conditions. An empirical evaluation shows that the proposed method accurately discovers the type of each data attribute and its associated update rules, and that the resulting BPS models more closely replicate the process execution control flow relative to data-unaware BPS models.

Representing and quantifying Minimal Residual Disease (MRD) in Acute Myeloid Leukemia (AML), a type of cancer that affects the blood and bone marrow, is essential in the prognosis and follow-up of AML patients. As traditional cytological analysis cannot detect leukemia cells below 5\%, the analysis of flow cytometry dataset is expected to provide more reliable results. In this paper, we explore statistical learning methods based on optimal transport (OT) to achieve a relevant low-dimensional representation of multi-patient flow cytometry measurements (FCM) datasets considered as high-dimensional probability distributions. Using the framework of OT, we justify the use of the K-means algorithm for dimensionality reduction of multiple large-scale point clouds through mean measure quantization by merging all the data into a single point cloud. After this quantization step, the visualization of the intra and inter-patients FCM variability is carried out by embedding low-dimensional quantized probability measures into a linear space using either Wasserstein Principal Component Analysis (PCA) through linearized OT or log-ratio PCA of compositional data. Using a publicly available FCM dataset and a FCM dataset from Bordeaux University Hospital, we demonstrate the benefits of our approach over the popular kernel mean embedding technique for statistical learning from multiple high-dimensional probability distributions. We also highlight the usefulness of our methodology for low-dimensional projection and clustering patient measurements according to their level of MRD in AML from FCM. In particular, our OT-based approach allows a relevant and informative two-dimensional representation of the results of the FlowSom algorithm, a state-of-the-art method for the detection of MRD in AML using multi-patient FCM.

We propose an original technique for the HF radar estimation of the main sea state parameters by exploiting the amplitude modulation of the radar signal time series. While the classical method for ocean wave measurement is based on the second-order ocean Doppler spectrum, this alternative approach uses the slow amplitude modulation of the Bragg wave in the radar signal, which is more robust to noise than the second-order echo. We apply this method to an annual set of HF radar data in the vicinity of Toulon (Mediterranean coast of France) and compare it with the classical second-order method, using a nearby buoy and the WWIII model as ground truth. The Doppler spectrum-based method is found to be more accurate in calculating the significant wave height and the peak wave frequency while the modulation approach provides coarser estimates but can achieve longer ranges and higher temporal coverage., Comment: 12 pages, 16 figures, 2 tables

Language Models (LMs) have achieved impressive performance on various linguistic tasks, but their relationship to human language processing in the brain remains unclear. This paper examines the gaps and overlaps between LMs and the brain at different levels of analysis, emphasizing the importance of looking beyond input-output behavior to examine and compare the internal processes of these systems. We discuss how insights from neuroscience, such as sparsity, modularity, internal states, and interactive learning, can inform the development of more biologically plausible language models. Furthermore, we explore the role of scaling laws in bridging the gap between LMs and human cognition, highlighting the need for efficiency constraints analogous to those in biological systems. By developing LMs that more closely mimic brain function, we aim to advance both artificial intelligence and our understanding of human cognition.

The Alderney Race has been identified as a future site for the development of tidal energy, due to its bidirectional strong current reaching 5 m/s during spring tides. This hydrodynamics is very difficult to measure by in situ or remote sensing means. High-frequency coastal radars can provide a synoptic and near-real-time view of such a complex circulation, but the classical processing algorithms are not adapted to the extreme situation of strongly sheared currents. We propose an improved high-resolution direction-finding technique for the azimuthal processing of such radar data. It uses phased-array systems and combines the advantages of the usual beam-forming technique to eliminate many problems related to the distortion of Doppler spectra by extreme currents. The method is evaluated with a unique data set of radar measurements at two radar frequencies (13 and 24.5 MHz) and three spatial resolutions (200, 750, and 1500 m). The radar-based surface currents are analyzed in the light of a high-resolution numerical model and also compared with in situ measurements. While high azimuthal resolution can be achieved in this way, it is shown that the typical range resolutions of 750 and 1500 m are insufficient to account for the strong spatial variations of the surface current at some specific times and locations., Comment: 26 pages, 15 figures, 3 tables

We address the issue of resilience of the Atlantic Meridional Overturning Circulation (AMOC) given the many indications that this dynamical system is in a multi-stable regime. A novel approach to resilience based on rare event techniques is presented which leads to a measure capturing `resistance to change` and `ability to return' aspects in a probabilistic way. The application of this measure to a conceptual model demonstrates its suitability for assessing AMOC resilience but also shows its potential use in many other non-autonomous dynamical systems. This framework is then extended to compute the probability that the AMOC undergoes a transition conditioned on an external forcing. Such conditional probability can be estimated by exploiting the information available when computing the resilience of this system. This allows us to provide a probabilistic view on safe operating spaces by defining a conditional safe operating space as a subset of the parameter space of the (possibly transient) imposed forcing., Comment: 35 pages, 8 figures, 1 table

Osseointegration is one of the key conditions for long term successful dental implantation. To this end, titanium alloys undergo plethora of surface treatments able to sustain osteogenic differentiation. For these surface treatments, femtosecond laser (FSL) can generate precise and reproducible surface patterns on titanium, avoiding thermal damage and chemical pollution. We recently identify that laser-induced periodic surface structure (LIPSS) with radial orientation, generated on model (flat) titanium surface, has a high osteogenic potential. However, nano-texturing is time consuming. In the present study, we aimed to reduce the texturing time of radial LIPSS, as well as processing of large commercially available dental screws by ways of laser beam engineering. Our objectives were to maintain at least osteogenic properties demonstrated on model surfaces by adjusting laser beam diameters and to demonstrate maintenance of performance with a dental screw texturing process not exceeding 1 minute.We first textured model surfaces with radial LIPSS by laser beams of different diameters, with surface impacts of 24$\mu$m, 80$\mu$m or 180$\mu$m, named as R24, R80 or R180 respectively. Osteogenic performance of human mesenchymal stem cells (hMSCs) were compared; seeded on polished control surfaces and textured surfaces and subjected to osteogenic evaluation by cell/matrix imaging, qRT-PCR and mineral deposition quantification. All textured surfaces showed greater osteogenic potential than the control surfaces, with significantly higher efficacy on R180 surface. Therefore, R180 pattern with large beam impacts was chosen for texturing on a dental screw and its osteogenic activity was compared to that of a non-textured screw. Interestingly, R180 required only 40 seconds to be textured on whole screws, on which it preserved a high osteogenic potential. Thus, by using FSL technology, we have improved the osteogenic potential of a topographic pattern while optimizing and scaling up its processing time on a medical device.

We propose a method for estimating disparity confidence intervals in stereo matching problems. Confidence intervals provide complementary information to usual confidence measures. To the best of our knowledge, this is the first method creating disparity confidence intervals based on the cost volume. This method relies on possibility distributions to interpret the epistemic uncertainty of the cost volume. Our method has the benefit of having a white-box nature, differing in this respect from current state-of-the-art deep neural networks approaches. The accuracy and size of confidence intervals are validated using the Middlebury stereo datasets as well as a dataset of satellite images. This contribution is freely available on GitHub.

While discounted payoff games and classic games that reduce to them, like parity and mean-payoff games, are symmetric, their solutions are not. We have taken a fresh view on the properties that optimal solutions need to have, and devised a novel way to converge to them, which is entirely symmetric. We achieve this by building a constraint system that uses every edge to define an inequation, and update the objective function by taking a single outgoing edge for each vertex into account. These edges loosely represent strategies of both players, where the objective function intuitively asks to make the inequation to these edges sharp. In fact, where they are not sharp, there is an `error' represented by the difference between the two sides of the inequation, which is 0 where the inequation is sharp. Hence, the objective is to minimise the sum of these errors. For co-optimal strategies, and only for them, it can be achieved that all selected inequations are sharp or, equivalently, that the sum of these errors is zero. While no co-optimal strategies have been found, we step-wise improve the error by improving the solution for a given objective function or by improving the objective function for a given solution. This also challenges the gospel that methods for solving payoff games are either based on strategy improvement or on value iteration., Comment: arXiv admin note: substantial text overlap with arXiv:2310.01008

Property $(TTT)$ was introduced by Ozawa as a strengthening of Kazhdan's property $(T)$ and Burger and Monod's property $(TT)$. In this paper, we improve Ozawa's result by showing that any simple algebraic group of rank $\geq 2$ over a local field has property $(TTT)$. We also show that lattices in a second countable locally compact group inherits property $(TTT)$. Finally, we study to what extent Lie groups with infinite center fail to have properties $(TT)$ and $(TTT)$.

We develop and compare two algorithms for computing first-order right-hand factors in the ring of linear Mahler operators$\ell_r M^r + \dots + \ell_1 M + \ell_0$where $\ell_0, \dots, \ell_r$ are polynomials in~$x$ and $Mx = x^b M$ for some integer $b \geq 2$. In other words, we give algorithms for finding all formal infinite product solutions of linear functional equations$\ell_r(x) f(x^{b^r}) + \dots + \ell_1(x) f(x^b) + \ell_0(x) f(x) = 0$. The first of our algorithms is adapted from Petkov\v{s}ek's classical algorithm forthe analogous problem in the case of linear recurrences. The second one proceeds by computing a basis of generalized power series solutions of the functional equation and by using Hermite-Pad{\'e} approximants to detect those linear combinations of the solutions that correspond to first-order factors. We present implementations of both algorithms and discuss their use in combination with criteria from the literature to prove the differential transcendence of power series solutions of Mahler equations., Comment: Dedicated to the memory of Marko Petkov\v{s}ek

We present new two-party protocols for the Unbalanced Private Set Union (UPSU) problem.Here, the Sender holds a set of data points, and the Receiver holds another (possibly much larger) set, and they would like for the Receiver to learn the union of the two sets and nothing else. Furthermore, the Sender's computational cost, along with the communication complexity, should be smaller when the Sender has a smaller set.While the UPSU problem has numerous applications and has seen considerable recent attention in the literature, our protocols are the first where the Sender's computational cost and communication volume are linear in the size of the Sender's set only, and do not depend on the size of the Receiver's set.Our constructions combine linearly homomorphic encryption (LHE) withfully homomorphic encryption (FHE). The first construction uses multi-point polynomial evaluation (MEv) on FHE, and achieves optimal linear cost for the Sender, but has higher quadratic computational cost for the Receiver. In the second construction we explore another trade-off: the Receiver computes fast polynomial Euclidean remainder in FHE while the Sender computes a fast MEv, in LHE only. This reduces the Receiver's cost to quasi-linear, with a modest increase in the computational cost for the Sender.Preliminary experimental results using HElib indicate that, for example, a Sender holding 1000 elements can complete our first protocol using less than 2s of computation time and less than 10MB of communication volume, independently of the Receiver's set size.

We consider a natural generalization of chordal graphs, in which every minimal separator induces a subgraph with independence number at most $2$. Such graphs can be equivalently defined as graphs that do not contain the complete bipartite graph $K_{2,3}$ as an induced minor, that is, graphs from which $K_{2,3}$ cannot be obtained by a sequence of edge contractions and vertex deletions. We develop a polynomial-time algorithm for recognizing these graphs. Our algorithm relies on a characterization of $K_{2,3}$-induced minor-free graphs in terms of excluding particular induced subgraphs, called Truemper configurations., Comment: 18 pages

Despite the high measurement fidelity that can now be reached, the dispersive qubit readout of circuit quantum electrodynamics is plagued by a loss of its quantum nondemolition character and a decrease in fidelity with increased measurement strength. In this work, we elucidate the nature of this dynamical process, which we refer to as transmon ionization. We develop a comprehensive framework which provides a physical picture of the origin of transmon ionization. This framework consists of three complementary levels of descriptions: a fully quantized transmon-resonator model, a semiclassical model where the resonator is treated as a classical drive on the transmon, and a fully classical model. Crucially, all three approaches preserve the full cosine potential of the transmon and lead to similar predictions. This framework identifies the multiphoton resonances responsible for transmon ionization. It also allows one to efficiently compute numerical estimates of the photon number threshold for ionization, which are in remarkable agreement with recent experimental results. The tools developed within this work are both conceptually and computationally simple, and we expect them to become an integral part of the theoretical underpinning of all circuit QED experiments., Comment: This version is close to the published version

We propose a non-commutative algorithm for multiplying 2x2 matrices using 7 coefficient products. This algorithm reaches simultaneously a better accuracy in practice compared to previously known such fast algorithms, and a time complexity bound with the best currently known leading term (obtained via alternate basis sparsification). To build this algorithm, we consider matrix and tensor norms bounds governing the stability and accuracy of numerical matrix multiplication. First, we reduce those bounds by minimizing a growth factor along the unique orbit of Strassen's 2x2-matrix multiplication tensor decomposition. Second, we develop heuristics for minimizing the number of operations required to realize a given bilinear formula, while further improving its accuracy. Third, we perform an alternate basis sparsification that improves on the time complexity constant and mostly preserves the overall accuracy.

Language models (LMs) have become ubiquitous in both NLP research and in commercial product offerings. As their commercial importance has surged, the most powerful models have become closed off, gated behind proprietary interfaces, with important details of their training data, architectures, and development undisclosed. Given the importance of these details in scientifically studying these models, including their biases and potential risks, we believe it is essential for the research community to have access to powerful, truly open LMs. To this end, we have built OLMo, a competitive, truly Open Language Model, to enable the scientific study of language models. Unlike most prior efforts that have only released model weights and inference code, we release OLMo alongside open training data and training and evaluation code. We hope this release will empower the open research community and inspire a new wave of innovation.

Information about pretraining corpora used to train the current best-performing language models is seldom discussed: commercial models rarely detail their data, and even open models are often released without accompanying training data or recipes to reproduce them. As a result, it is challenging to conduct and advance scientific research on language modeling, such as understanding how training data impacts model capabilities and limitations. To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials. We extensively document Dolma, including its design principles, details about its construction, and a summary of its contents. We present analyses and experimental results on intermediate states of Dolma to share what we have learned about important data curation practices. Finally, we open-source our data curation toolkit to enable reproduction of our work as well as support further research in large-scale data curation., Comment: Accepted at ACL 2024; Dataset: https://hf.co/datasets/allenai/dolma; Code: https://github.com/allenai/dolma

Recent simulations performed with the Community Earth System Model (CESM) have suggested a crucial role of sea-ice processes in AMOC hysteresis behaviour under varying surface freshwater forcing. Here, we further investigate this issue using additional CESM simulations and a novel conceptual ocean-sea-ice box model. The CESM simulations show that the presence of sea ice gives rise to the existence of statistical equilibrium states with a weak AMOC strength. This is confirmed in the conceptual model, which captures the same AMOC hysteresis behaviour as in the CESM simulation and where steady states are computed versus forcing parameters. In the conceptual model, transition probabilities between the different equilibrium states are determined using rare event techniques. The transition probabilities from a strong AMOC state to a weak AMOC state increase when considering sea-ice processes and indicate that sea ice promotes these transitions. On the other hand, sea ice strongly reduces the probabilities of the reverse transition from a weak AMOC state to a strong AMOC state and this implies that sea ice also limits AMOC recovery. The results here indicate that sea-ice processes play a dominant role in AMOC hysteresis width and influence transition probabilities between the different equilibrium states.

The Atlantic Meridional Overturning Circulation (AMOC) is an important component of the global climate, known to be a tipping element, as it could collapse under global warming. The main objective of this study is to compute the probability that the AMOC collapses within a specified time window, using a rare-event algorithm called Trajectory-Adaptive Multilevel Splitting (TAMS). However, the efficiency and accuracy of TAMS depend on the choice of the score function. Although the definition of the optimal score function, called ``committor function" is known, it is impossible in general to compute it a priori. Here, we combine TAMS with a Next-Generation Reservoir Computing technique that estimates the committor function from the data generated by the rare-event algorithm. We test this technique in a stochastic box model of the AMOC for which two types of transition exist, the so-called F(ast)-transitions and S(low)-transitions. Results for the F-transtions compare favorably with those in the literature where a physically-informed score function was used. We show that coupling a rare-event algorithm with machine learning allows for a correct estimation of transition probabilities, transition times, and even transition paths for a wide range of model parameters. We then extend these results to the more difficult problem of S-transitions in the same model. In both cases of F-transitions and S-transitions, we also show how the Next-Generation Reservoir Computing technique can be interpreted to retrieve an analytical estimate of the committor function., Comment: 15 pages, 9 figures

Predictive process monitoring is an evolving research field that studies how to train and use predictive models for operational decision-making. One of the problems studied in this field is that of predicting the sequence of upcoming activities in a case up to its completion, a.k.a. the case suffix. The prediction of case suffixes provides input to estimate short-term workloads and execution times under different resource schedules. Existing methods to address this problem often generate suffixes wherein some activities are repeated many times, whereas this pattern is not observed in the data. Closer examination shows that this shortcoming stems from the approach used to sample the successive activity instances to generate a case suffix. Accordingly, the paper introduces a sampling approach aimed at reducing repetitions of activities in the predicted case suffixes. The approach, namely Daemon action, strikes a balance between exploration and exploitation when generating the successive activity instances. We enhance a deep learning approach for case suffix predictions using this sampling approach, and experimentally show that the enhanced approach outperforms the unenhanced ones with respect to control-flow accuracy measures.

Transmission system operators employ reserves to deal with unexpected variations of demand and generation to guarantee the security of supply. The French transmission system operator RTE dynamically sizes the required margins using a probabilistic approach relying on continuous forecasts of the main drivers of the uncertainties of the system imbalance and a 1 % risk threshold. However, this criterion does not specify which means to activate upward/downward and when to face a deficit of available margins versus the required margins. Thus, this work presents a strategy using a probabilistic unit commitment with a stochastic optimization-based approach, including the fixed and variable costs of units and the costs of lost load and production. The abstract problem is formulated with a multi-stage stochastic program and approximated with a heuristic called two-stage stochastic model predictive control. It solves a sequence of two-stage stochastic programs to conduct the central dispatch. An implementation is conducted by solving an approximated version with a single two-stage stochastic program. This method is tested on a real case study comprising nuclear and fossil-based units with French electrical consumption and renewable production.

Despite decades of research, the creative process remains to be fully understood, and most theories and empirical evidence focus on adults' creativity. Without understanding children's creative processes, the generalizability of these theories is questionable, which is crucial for teaching, learning, and parenting. However, studying children's creative processes is challenging because of their still-developing meta-cognition, which is an essential component of the creative process. In this study, we conducted interviews with middle childhood-aged students (n = 30) after they completed four different types of divergent thinking tasks, examining the creative process during and beyond idea generation. Using inductive thematic analysis, we identified four main themes with associated subthemes, with Memory (experiences, knowledge, references, and personal and social interests) and Cognitive Strategies (evaluative processes and hypothetical thinking) as the primary themes. These data offer insight into how the main themes Memory and Cognitive Strategies reflect associative and executive processes, respectively, and the interplay between these in children's divergent thinking. Findings from this study revealed that children were able to explain their process of generating an idea in the vast majority of responses. Additionally, this work provides insight into improving creativity, methods of assessment, and our understanding of creativity development. [This is the online version of an article published in "Creativity Research Journal."]

This brief explores data science methods that workforce programs can use to predict participant success. With access to vast amounts of data on their programs, workforce training providers can leverage their management information systems (MIS) to understand and improve their programs' outcomes. By predicting which participants are at greater risk of dropping out of their program and why, providers can segment their caseloads so that participants receive services better tailored to their needs. Within the data science field, machine learning (ML) has gained popularity for its ability to extract hidden patterns without being explicitly guided by a data analyst. While these data science methods hold promise, are the added costs and complexity worth it? The authors explore the tradeoffs by answering the following questions: (1) What factors are important in predicting a participant's outcome in a program?; (2) Are participant outcomes predictable using simple methods, like creating basic risk indicators in Management Information Systems (MIS)? For example, how well does an indicator for prior education predict participant outcomes?; and (3) What is the added value and cost of incorporating regression and more complex machine learning methods?