Search

In medical imaging, developing generalized segmentation models that can handle multiple organs and lesions is crucial. However, the scarcity of fully annotated datasets and strict privacy regulations present significant barriers to data sharing. Federated Learning (FL) allows decentralized model training, but existing FL methods often struggle with partial labeling, leading to model divergence and catastrophic forgetting. We propose ConDistFL, a novel FL framework incorporating conditional distillation to address these challenges. ConDistFL enables effective learning from partially labeled datasets, significantly improving segmentation accuracy across distributed and non-uniform datasets. In addition to its superior segmentation performance, ConDistFL maintains computational and communication efficiency, ensuring its scalability for real-world applications. Furthermore, ConDistFL demonstrates remarkable generalizability, significantly outperforming existing FL methods in out-of-federation tests, even adapting to unseen contrast phases (e.g., non-contrast CT images) in our experiments. Extensive evaluations on 3D CT and 2D chest X-ray datasets show that ConDistFL is an efficient, adaptable solution for collaborative medical image segmentation in privacy-constrained settings.

Citizen science biodiversity data present great opportunities for ecology and conservation across vast spatial and temporal scales. However, the opportunistic nature of these data lacks the sampling structure required by modeling methodologies that address a pervasive challenge in ecological data collection: imperfect detection, i.e., the likelihood of under-observing species on field surveys. Occupancy modeling is an example of an approach that accounts for imperfect detection by explicitly modeling the observation process separately from the biological process of habitat selection. This produces species distribution models that speak to the pattern of the species on a landscape after accounting for imperfect detection in the data, rather than the pattern of species observations corrupted by errors. To achieve this benefit, occupancy models require multiple surveys of a site across which the site's status (i.e., occupied or not) is assumed constant. Since citizen science data are not collected under the required repeated-visit protocol, observations may be grouped into sites post hoc. Existing approaches for constructing sites discard some observations and/or consider only geographic distance and not environmental similarity. In this study, we compare ten approaches for site construction in terms of their impact on downstream species distribution models for 31 bird species in Oregon, using observations recorded in the eBird database. We find that occupancy models built on sites constructed by spatial clustering algorithms perform better than existing alternatives., Comment: AAAI 2025

Organizations seeking to utilize Large Language Models (LLMs) for knowledge querying and analysis often encounter challenges in maintaining an LLM fine-tuned on targeted, up-to-date information that keeps answers relevant and grounded. Retrieval Augmented Generation (RAG) has quickly become a feasible solution for organizations looking to overcome the challenges of maintaining proprietary models and to help reduce LLM hallucinations in their query responses. However, RAG comes with its own issues regarding scaling data pipelines across tiered-access and disparate data sources. In many scenarios, it is necessary to query beyond a single data silo to provide richer and more relevant context for an LLM. Analyzing data sources within and across organizational trust boundaries is often limited by complex data-sharing policies that prohibit centralized data storage, therefore, inhibit the fast and effective setup and scaling of RAG solutions. In this paper, we introduce Confidential Computing (CC) techniques as a solution for secure Federated Retrieval Augmented Generation (FedRAG). Our proposed Confidential FedRAG system (C-FedRAG) enables secure connection and scaling of a RAG workflows across a decentralized network of data providers by ensuring context confidentiality. We also demonstrate how to implement a C-FedRAG system using the NVIDIA FLARE SDK and assess its performance using the MedRAG toolkit and MIRAGE benchmarking dataset.

We present NAVCON, a large-scale annotated Vision-Language Navigation (VLN) corpus built on top of two popular datasets (R2R and RxR). The paper introduces four core, cognitively motivated and linguistically grounded, navigation concepts and an algorithm for generating large-scale silver annotations of naturally occurring linguistic realizations of these concepts in navigation instructions. We pair the annotated instructions with video clips of an agent acting on these instructions. NAVCON contains 236, 316 concept annotations for approximately 30, 0000 instructions and 2.7 million aligned images (from approximately 19, 000 instructions) showing what the agent sees when executing an instruction. To our knowledge, this is the first comprehensive resource of navigation concepts. We evaluated the quality of the silver annotations by conducting human evaluation studies on NAVCON samples. As further validation of the quality and usefulness of the resource, we trained a model for detecting navigation concepts and their linguistic realizations in unseen instructions. Additionally, we show that few-shot learning with GPT-4o performs well on this task using large-scale silver annotations of NAVCON.

Computational variant effect predictors (VEPs) are providing increasingly strong evidence to classify the pathogenicity of missense variants. Precision vs. recall analysis is useful in evaluating VEP performance, especially when adjusted for imbalanced test sets. Here, we describe VEPerform, a web-based tool for evaluating the performance of VEPs at the gene level using balanced precision vs. recall curve (BPRC) analysis.

Domain engineering in ferroelectric thin films is crucial for next-generation microelectronic and photonic technologies. Here, a method is demonstrated to precisely control domain configurations in BaTiO$_3$ thin films through low-energy He ion implantation. The approach transforms a mixed ferroelectric domain state with significant in-plane polarization into a uniform out-of-plane tetragonal phase by selectively modifying the strain state in the film's top region. This structural transition significantly improves domain homogeneity and reduces polarization imprint, leading to symmetric ferroelectric switching characteristics. The demonstrated ability to manipulate ferroelectric domains post-growth enables tailored functional properties without compromising the coherently strained bottom interface. The method's compatibility with semiconductor processing and ability to selectively modify specific regions make it particularly promising for practical implementation in integrated devices. This work establishes a versatile approach for strain-mediated domain engineering that could be extended to a wide range of ferroelectric systems, providing new opportunities for memory, sensing, and photonic applications where precise control of polarization states is essential.

The popularity of the CLEAN algorithm in radio interferometric imaging stems from its maturity, speed, and robustness. While many alternatives have been proposed in the literature, none have achieved mainstream adoption by astronomers working with data from interferometric arrays operating in the big data regime. This lack of adoption is largely due to increased computational complexity, absence of mature implementations, and the need for astronomers to tune obscure algorithmic parameters. This work introduces pfb-imaging: a flexible library that implements the scaffolding required to develop and accelerate general radio interferometric imaging algorithms. We demonstrate how the framework can be used to implement a sparsity-based image reconstruction technique known as (unconstrained) SARA in a way that scales with image size rather than data volume and features interpretable algorithmic parameters. The implementation is validated on terabyte-sized data from the MeerKAT telescope, using both a single compute node and Amazon Web Services computing instances., Comment: 29 pages, submitted to Astronomy & Computing

Quantifying the uncertainty in the factual parametric knowledge of Large Language Models (LLMs), especially in a black-box setting, poses a significant challenge. Existing methods, which gauge a model's uncertainty through evaluating self-consistency in responses to the original query, do not always capture true uncertainty. Models might respond consistently to the origin query with a wrong answer, yet respond correctly to varied questions from different perspectives about the same query, and vice versa. In this paper, we propose a novel method, DiverseAgentEntropy, for evaluating a model's uncertainty using multi-agent interaction under the assumption that if a model is certain, it should consistently recall the answer to the original query across a diverse collection of questions about the same original query. We further implement an abstention policy to withhold responses when uncertainty is high. Our method offers a more accurate prediction of the model's reliability and further detects hallucinations, outperforming other self-consistency-based methods. Additionally, it demonstrates that existing models often fail to consistently retrieve the correct answer to the same query under diverse varied questions even when knowing the correct answer.

Understanding the properties of excited states of complex molecules is crucial for many chemical and physical processes. Calculating these properties is often significantly more resource-intensive than calculating their ground state counterparts. We present a quantum machine learning model that predicts excited-state properties from the molecular ground state for different geometric configurations. The model comprises a symmetry-invariant quantum neural network and a conventional neural network and is able to provide accurate predictions with only a few training data points. The proposed procedure is fully NISQ compatible. This is achieved by using a quantum circuit that requires a number of parameters linearly proportional to the number of molecular orbitals, along with a parameterized measurement observable, thereby reducing the number of necessary measurements. We benchmark the algorithm on three different molecules by evaluating its performance in predicting excited state transition energies and transition dipole moments. We show that, in many instances, the procedure is able to outperform various classical models that rely solely on classical features., Comment: 10 + 4 pages, 7 + 3 figures

The even distribution and optimization of tasks across resources and workstations is a critical process in manufacturing aimed at maximizing efficiency, productivity, and profitability, known as Robotic Assembly Line Balancing (RALB). With the increasing complexity of manufacturing required by mass customization, traditional computational approaches struggle to solve RALB problems efficiently. To address these scalability challenges, we investigate applying quantum computing, particularly quantum annealing, to the real-world based problem. We transform the integer programming formulation into a quadratic unconstrained binary optimization problem, which is then solved using a hybrid quantum-classical algorithm on the D-Wave Advantage 4.1 quantum computer. In a case study, the quantum solution is compared to an exact solution, demonstrating the potential for quantum computing to enhance manufacturing productivity and reduce costs. Nevertheless, limitations of quantum annealing, including hardware constraints and problem-specific challenges, suggest that continued advancements in quantum technology will be necessary to improve its applicability to RALB manufacturing optimization., Comment: 7 pages, 3 figures, 2 tables

We present a scalable, robust approach to creating quantum programs of arbitrary size and complexity. The approach is based on the true abstraction of the problem. The quantum program is expressed in terms of a high-level model together with constraints and objectives on the final program. Advanced synthesis algorithms transform the model into a low-level quantum program that meets the user's specification and is directed at a stipulated hardware. This separation of description from implementation is essential for scale. The technology adapts electronic design automation methods to quantum computing, finding feasible implementations in a virtually unlimited functional space. The results show clear superiority over the compilation and transpilation methods used today. We expect that this technological approach will take over and prevail as quantum software become more demanding, complex, and essential.

This paper introduces Mayfly, a federated analytics approach enabling aggregate queries over ephemeral on-device data streams without central persistence of sensitive user data. Mayfly minimizes data via on-device windowing and contribution bounding through SQL-programmability, anonymizes user data via streaming differential privacy (DP), and mandates immediate in-memory cross-device aggregation on the server -- ensuring only privatized aggregates are revealed to data analysts. Deployed for a sustainability use case estimating transportation carbon emissions from private location data, Mayfly computed over 4 million statistics across more than 500 million devices with a per-device, per-week DP $\varepsilon = 2$ while meeting strict data utility requirements. To achieve this, we designed a new DP mechanism for Group-By-Sum workloads leveraging statistical properties of location data, with potential applicability to other domains., Comment: 22 pages, 7 figures

Model merging combines multiple expert models - finetuned from a base foundation model on diverse tasks and domains - into a single, more capable model. However, most existing model merging approaches assume that all experts are available simultaneously. In reality, new tasks and domains emerge progressively over time, requiring strategies to integrate the knowledge of expert models as they become available: a process we call temporal model merging. The temporal dimension introduces unique challenges not addressed in prior work, raising new questions such as: when training for a new task, should the expert model start from the merged past experts or from the original base model? Should we merge all models at each time step? Which merging techniques are best suited for temporal merging? Should different strategies be used to initialize the training and deploy the model? To answer these questions, we propose a unified framework called TIME - Temporal Integration of Model Expertise - which defines temporal model merging across three axes: (1) Initialization Phase, (2) Deployment Phase, and (3) Merging Technique. Using TIME, we study temporal model merging across model sizes, compute budgets, and learning horizons on the FoMo-in-Flux benchmark. Our comprehensive suite of experiments across TIME allows us to uncover key insights for temporal model merging, offering a better understanding of current challenges and best practices for effective temporal model merging., Comment: Technical Report. Code at https://github.com/ExplainableML/fomo_in_flux

A digital twin is a virtual representation that accurately replicates its physical counterpart, fostering bi-directional real-time data exchange throughout the entire process lifecycle. For Laser Directed Energy Deposition of Wire (DED-LB/w) additive manufacturing processes, digital twins may help to control the residual stress design in build parts. This study focuses on providing faster-than-real-time and highly accurate surrogate models for the formation of residual stresses by employing neural ordinary differential equations. The approach enables accurate prediction of temperatures and altered structural properties like stress tensor components. The developed surrogates can ultimately facilitate on-the-fly re-optimization of the ongoing manufacturing process to achieve desired structural outcomes. Consequently, this building block contributes significantly to realizing digital twins and the first-time-right paradigm in additive manufacturing., Comment: Presented at the ICCE 2024, Darmstadt

In invasive breast cancer, cell clusters of varying sizes and shapes are embedded in the fibrous extracellular matrix (ECM). Although the prevailing view attributes this structure to increasing disorder resulting from loss of function and dedifferentiation, our findings reveal that it arises through a process of active self-organization driven by cancer cell motility. Simulations and histological analyses of tumours from over 2,000 breast cancer patients reveal that motile, aligned cancer cells within clusters move as active nematic aggregates through the surrounding highly aligned ECM fibres, which form a confining, passive nematic phase. Cellular motion leads to cluster splitting and coalescence. The degree of cluster activity, combined with heterogeneity in cell motility, is reflected in specific scaling behaviours for cluster shape, size distribution, and the distance between cluster boundaries and nematic defects in ECM alignment. Increased activity estimates correlate with tumour progression and are associated with a poorer prognosis for patients.

We present an efficient reduction that converts any machine learning algorithm into an interactive protocol, enabling collaboration with another party (e.g., a human) to achieve consensus on predictions and improve accuracy. This approach imposes calibration conditions on each party, which are computationally and statistically tractable relaxations of Bayesian rationality. These conditions are sensible even in prior-free settings, representing a significant generalization of Aumann's classic "agreement theorem." In our protocol, the model first provides a prediction. The human then responds by either agreeing or offering feedback. The model updates its state and revises its prediction, while the human may adjust their beliefs. This iterative process continues until the two parties reach agreement. Initially, we study a setting that extends Aumann's Agreement Theorem, where parties aim to agree on a one-dimensional expectation by iteratively sharing their current estimates. Here, we recover the convergence theorem of Aaronson'05 under weaker assumptions. We then address the case where parties hold beliefs over distributions with d outcomes, exploring two feedback mechanisms. The first involves vector-valued estimates of predictions, while the second adopts a decision-theoretic approach: the human, needing to take an action from a finite set based on utility, communicates their utility-maximizing action at each round. In this setup, the number of rounds until agreement remains independent of d. Finally, we generalize to scenarios with more than two parties, where computational complexity scales linearly with the number of participants. Our protocols rely on simple, efficient conditions and produce predictions that surpass the accuracy of any individual party's alone.

Coping with prolonged periods of low availability of wind and solar power, also referred to as "Dunkelflaute", emerges as a key challenge for realizing a decarbonized European energy system fully based on renewable energy sources. Here, we investigate the role of long-duration electricity storage and geographical balancing in dealing with such variable renewable energy droughts. To this end, we combine renewable availability time series analysis and power sector modeling, using 36 historic weather years. We find that extreme drought events define long-duration storage operation and investment. The most extreme event in Europe occurred in the winter of 1996/97. Assuming policy-relevant interconnection, long-duration storage of 351 TWh or 7% of yearly electricity demand would be required to deal with this event. As it affects many countries simultaneously, a storage capacity of 159 TWh or 3% of yearly electricity demand remains required even in the extreme case of unconstrained geographical balancing. Before and during Dunkelflaute events, we find complex interactions of long-duration storage with other flexibility options. Sensitivity analyses illustrate that firm zero-emission generation technologies would only moderately reduce long-duration storage needs. Thus, policymakers and system planners should prepare for a rapid expansion of long-duration storage capacity to safeguard the renewable energy transition in Europe. We further argue that including multiple weather years is required for weather-resilient energy system modeling, particularly those with pronounced renewable energy droughts.

Covert communication in wireless networks ensures that transmissions remain undetectable to adversaries, making it a potential enabler for privacy and security in sensitive applications. However, to meet the high performance and connectivity demands of sixth-generation (6G) networks, future wireless systems will require larger antenna arrays, higher operating frequencies, and advanced antenna architectures. This shift changes the propagation model from far-field planar-wave to near-field spherical-wave which necessitates a redesign of existing covert communication systems. Unlike far-field beamforming, which relies only on direction, near-field beamforming depends on both distance and direction, providing additional degrees of freedom for system design. In this paper, we aim to utilize those freedoms by proposing near-field Frequency Diverse Array (FDA)-based transmission strategies that manipulate the beampattern in both distance and angle, thereby establishing a non-covert region around the legitimate user. Our approach takes advantage of near-field properties and FDA technology to significantly reduce the area vulnerable to detection by adversaries while maintaining covert communication with the legitimate receiver. Numerical simulations show that our methods outperform conventional phased arrays by shrinking the non-covert region and allowing the covert region to expand as the number of antennas increases.

Large-scale multimodal representation learning successfully optimizes for zero-shot transfer at test time. Yet the standard pretraining paradigm (contrastive learning on large amounts of image-text data) does not explicitly encourage representations to support few-shot adaptation. In this work, we propose a simple, but carefully designed extension to multimodal pretraining which enables representations to accommodate additional context. Using this objective, we show that vision-language models can be trained to exhibit significantly increased few-shot adaptation: across 21 downstream tasks, we find up to four-fold improvements in test-time sample efficiency, and average few-shot adaptation gains of over 5%, while retaining zero-shot generalization performance across model scales and training durations. In particular, equipped with simple, training-free, metric-based adaptation mechanisms, our representations easily surpass more complex and expensive optimization-based schemes, vastly simplifying generalization to new domains.

Smart contracts form the core of Web3 applications. Contracts mediate the transfer of cryptocurrency, making them irresistible targets for hackers. We introduce ASP, a system aimed at easing the construction of provably secure contracts. The Asp system consists of three closely-linked components: a programming language, a defensive compiler, and a proof checker. The language semantics guarantee that Asp contracts are free of commonly exploited vulnerabilities such as arithmetic overflow and reentrancy. The defensive compiler enforces the semantics and translates Asp to Solidity, the most popular contract language. Deductive proofs establish functional correctness and freedom from critical vulnerabilities such as unauthorized access.

Generalist vision language models (VLMs) have made significant strides in computer vision, but they fall short in specialized fields like healthcare, where expert knowledge is essential. In traditional computer vision tasks, creative or approximate answers may be acceptable, but in healthcare, precision is paramount.Current large multimodal models like Gemini and GPT-4o are insufficient for medical tasks due to their reliance on memorized internet knowledge rather than the nuanced expertise required in healthcare. VLMs are usually trained in three stages: vision pre-training, vision-language pre-training, and instruction fine-tuning (IFT). IFT has been typically applied using a mixture of generic and healthcare data. In contrast, we propose that for medical VLMs, a fourth stage of specialized IFT is necessary, which focuses on medical data and includes information from domain expert models. Domain expert models developed for medical use are crucial because they are specifically trained for certain clinical tasks, e.g. to detect tumors and classify abnormalities through segmentation and classification, which learn fine-grained features of medical data$-$features that are often too intricate for a VLM to capture effectively especially in radiology. This paper introduces a new framework, VILA-M3, for medical VLMs that utilizes domain knowledge via expert models. Through our experiments, we show an improved state-of-the-art (SOTA) performance with an average improvement of ~9% over the prior SOTA model Med-Gemini and ~6% over models trained on the specific tasks. Our approach emphasizes the importance of domain expertise in creating precise, reliable VLMs for medical applications.

Simulating large, strongly interacting fermionic systems remains a major challenge for existing numerical methods. In this work, we present, for the first time, the application of neural quantum states - specifically, hidden fermion determinant states (HFDS) - to simulate the strongly interacting limit of the Fermi-Hubbard model, namely the $t-J$ model, across the entire doping regime. We demonstrate that HFDS achieve energies competitive with matrix product states (MPS) on lattices as large as $8 \times 8$ sites while using several orders of magnitude fewer parameters, suggesting the potential for efficient application to even larger system sizes. This remarkable efficiency enables us to probe low-energy physics across the full doping range, providing new insights into the competition between kinetic and magnetic interactions and the nature of emergent quasiparticles. Starting from the low-doping regime, where magnetic polarons dominate the low energy physics, we track their evolution with increasing doping through analyses of spin and polaron correlation functions. Our findings demonstrate the potential of determinant-based neural quantum states with inherent fermionic sign structure, opening the way for simulating large-scale fermionic systems at any particle filling.

Language model users often issue queries that lack specification, where the context under which a query was issued -- such as the user's identity, the query's intent, and the criteria for a response to be useful -- is not explicit. For instance, a good response to a subjective query like "What book should I read next?" would depend on the user's preferences, and a good response to an open-ended query like "How do antibiotics work against bacteria?" would depend on the user's expertise. This makes evaluation of responses to such queries an ill-posed task, as evaluators may make arbitrary judgments about the response quality. To remedy this, we present contextualized evaluations, a protocol that synthetically constructs context surrounding an underspecified query and provides it during evaluation. We find that the presence of context can 1) alter conclusions drawn from evaluation, even flipping win rates between model pairs, 2) nudge evaluators to make fewer judgments based on surface-level criteria, like style, and 3) provide new insights about model behavior across diverse contexts. Specifically, our procedure uncovers an implicit bias towards WEIRD contexts in models' "default" responses and we find that models are not equally sensitive to following different contexts, even when they are provided in prompts., Comment: Code & data available at https://github.com/allenai/ContextEval

Recent progress with CMOS detector development has opened new parameter space for high cadence time resolved imaging of transients and fast proper motion solar system objects. Using computer simulations for a ground-based 1.23 m telescope, this research note illustrates the gain of a new generation of fast readout low noise qCMOS sensors over CCDs and makes the case for high precision monitoring of asteroid orbits that can potentially shed light on the hypothetical existence of low mass primordial black holes, as well as for other applications requiring high speed imaging., Comment: 3 pages, 1 figure. Accepted by RNAAS

This article presents a generic hybrid numerical method to price a wide range of options on one or several assets, as well as assets with stochastic drift or volatility. In particular for equity and interest rate hybrid with local volatility.

How can we test AI performance? This question seems trivial, but it isn't. Standard benchmarks often have problems such as in-distribution and small-size test sets, oversimplified metrics, unfair comparisons, and short-term outcome pressure. As a consequence, good performance on standard benchmarks does not guarantee success in real-world scenarios. To address these problems, we present Touchstone, a large-scale collaborative segmentation benchmark of 9 types of abdominal organs. This benchmark is based on 5,195 training CT scans from 76 hospitals around the world and 5,903 testing CT scans from 11 additional hospitals. This diverse test set enhances the statistical significance of benchmark results and rigorously evaluates AI algorithms across various out-of-distribution scenarios. We invited 14 inventors of 19 AI algorithms to train their algorithms, while our team, as a third party, independently evaluated these algorithms on three test sets. In addition, we also evaluated pre-existing AI frameworks--which, differing from algorithms, are more flexible and can support different algorithms--including MONAI from NVIDIA, nnU-Net from DKFZ, and numerous other open-source frameworks. We are committed to expanding this benchmark to encourage more innovation of AI algorithms for the medical domain., Comment: Accepted to NeurIPS-2024

This work examines the hypothesis that intravoxel incoherent motion MRI (IVIM) can quantify local cerebral blood flow (qCBF), infarct volume, and define the ischemic penumbra for determination of the perfusion-diffusion mismatch (PWI/DWI) volume in a setting of acute ischemic stroke. Eight experiments were conducted in a pre-clinical middle cerebral artery occlusion (MCAO) model. IVIM and dynamic susceptibility contrast (DSC) imaging were acquired 2.5hr post-MCAO. IVIM was post-processed using software written in-house to produce parametric images of local qCBF, Water Transport Time (WTT), diffusion, and subsequently, PWI/DWI mismatch. These IVIM image parameters were compared with delay-and-dispersion-corrected local-AIF DSC perfusion image parameters including Tmax, qCBF, mean transit time (MTT), and mean diffusivity for DSC PWI/DWI mismatch. Final infarct volume was measured 4hrs post-occlusion. Early (2.5hr post-occlusion) DSC qCBF and IVIM qCBF in the diffusion negative MCA territory correlated strongly (slope=1.00, p=0.01,R2=0.69,Lins CCC=0.71), and both DSC and IVIM qCBF values negatively correlated with final infarct volume (R2=0.78,R2=0.61 respectively). The volume of hypoperfusion measured at 2.5 hours from DSC qCBF and from IVIM qCBF both predicted final infarct volume with good sensitivity and correlation (slope=2.08, R2=0.67, slope=2.50,R2=0.68 respectively). IVIM PWI/DWI ratio was correlated with infarct growth (R2=0.70) and WTT correlated with MTT (slope=0.82,R2=0.60). IVIM qCBF correlated strongly with local-AIF DSC qCBF and IVIM PWI/DWI correlated strongly with infarct growth. Both DSC and IVIM quantitative perfusion image acquired early after occlusion were able to predict final infarct volume, and IVIM simultaneous PWI/DWI ratio predicted infarct growth., Comment: Incomplete analysis

Who ends up in the teacher workforce is greatly influenced by who is admitted into teacher education programs (TEPs). To better understand how the preferences of teacher education faculty might shape admissions of STEM teacher candidates, we surveyed faculty who teach content or methods courses to STEM teacher candidates across five universities. Faculty reported that they most value information collected from individual interviews with applicants and data on the number of STEM courses taken in college and their performance in these courses, and least value data on university admissions tests, high school GPA, and teacher licensure test scores. When we investigate faculty members' revealed preferences through a conjoint analysis, we find that faculty most value applicants who have worked with students from diverse backgrounds and applicants from a marginalized racial or ethnic community, and least value whether they received high grades in math and/or science courses. Finally, we find significant variation in these perceptions across respondents in different faculty roles, who teach different courses, and from different institutions: for example, Arts and Sciences faculty tend to value TEP applicants' performance in college STEM courses relatively more than STEM education faculty, while STEM education faculty tend to value applicants' race and ethnicity relatively more than Arts and Sciences faculty.