Stefan T. Radev
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Stefan T. Radev has a total of 73 co-authors. Here are the top co-authors by number of shared publications:
- Andreas Voss
Heidelberg University
11 shared publications - Ullrich Köthe
Heidelberg University
7 shared publications - Paul-Christian Bürkner
Department of Statistics, TU Dortmund University, Dortmund, Germany
6 shared publications - Mischa von Krause
Department of Psychology, Heidelberg University, 69117 Heidelberg, Germany
3 shared publications - Lasse Elsemüller
Institute of Psychology, Heidelberg University
2 shared publications - Lukas Schumacher
University of Basel, Basel, Switzerland
2 shared publications
Latest Publications
- Open Access
- 9 October 2026
- Published by Center for Open Science in PsyArXiv
Neural Racing Accumulator Model Estimation With Lightweight Monotonic Flows
Race models describe speeded decision-making as a race between competing runners that accumulate evidence in favor of available choice options. Their parameters can be estimated through the model likelihood that combines the probability and cumulative density functions of each runner. However, estimating parameters of complex race models whose runners have intractable densities with (hierarchical) Bayesian inference has been challenging: Numerical density approximations tend to be slow and amortized neural posterior estimation methods require complex network architectures that can be expensive to train and, in classical settings, tie the trained estimator to a specific prior distribution and generative model. We propose a flexible neural estimation method that decouples training from the choice of prior distribution or model parameterization. We train a small neural network together with a monotonic neural spline flow to simultaneously learn the runners' conditional probability density and cumulative density functions. We then combine both functions to efficiently estimate race model likelihoods that can be used in combination with modern samplers. We demonstrate the utility of our approach on both tractable and intractable versions of the popular racing diffusion model. Based on our results, we argue that, when flexibility in choosing prior distributions or model parameterizations is needed, learning the model likelihood with lightweight neural density estimators is advantageous.
- 9 October 2026
- Published by Center for Open Science in PsyArXiv
Open AccessNeural Racing Accumulator Model Estimation With Lightweight Monotonic Flows
Race models describe speeded decision-making as a race between competing runners that accumulate evidence in favor of available choice options. Their parameters can be estimated through the model likelihood that combines the probability and cumulative density functions of each runner. However, estimating parameters of complex race models whose runners have intractable densities with (hierarchical) Bayesian inference has been challenging: Numerical density approximations tend to be slow and amortized neural posterior estimation methods require complex network architectures that can be expensive to train and, in classical settings, tie the trained estimator to a specific prior distribution and generative model. We propose a flexible neural estimation method that decouples training from the choice of prior distribution or model parameterization. We train a small neural network together with a monotonic neural spline flow to simultaneously learn the runners' conditional probability density and cumulative density functions. We then combine both functions to efficiently estimate race model likelihoods that can be used in combination with modern samplers. We demonstrate the utility of our approach on both tractable and intractable versions of the popular racing diffusion model. Based on our results, we argue that, when flexibility in choosing prior distributions or model parameterizations is needed, learning the model likelihood with lightweight neural density estimators is advantageous.
- Open Access
- 16 August 2026
- Published by Center for Open Science in PsyArXiv
Moderational Learning: A Framework for Discovering Models and IndividualDifferences from Behavioral Data
Understanding how people respond to stimuli requires models capturing both common psychological mechanisms and individual variability. Here we introduce moderational learning, a data-driven unsupervised machine-learning framework that aims to jointly discover interpretable latent dimensions representing individual variability and shared generative functions that map stimuli to responses. We implement the framework combining amortized variational inference to infer individual representations with a conditional generative decoder for predicting responses. By combining posterior analyses with symbolic regression, we identify interpretable individual-difference factors and derive explicit computational models from behavioral data. We then extend the approach to cross-domain moderational learning, allowing latent representations to be inferred from multiple behavioral tasks or data modalities. Two simulation studies show that the approach captures stimulus--response mappings and recovers underlying latent structures. As a testbed, we applied moderational learning to a behavioral dataset in which 507 participants completed four value-based decision-making tasks. In empirical analyses, moderational learning achieved higher held-out predictive performance than the comparison models and yielded exploratory within-task and cross-task latent structures, with dimensions corresponding to risk discounting, delay discounting, and bidding proclivity. Sensitivity analyses further reveal that responses in pricing tasks do not always decrease monotonically with aversive attributes. Together, these results demonstrate that moderational learning provides a data-driven framework for discovering latent behavioral structure and generative mechanisms, opening a pathway toward automated theory development in psychological science.
- 16 August 2026
- Published by Center for Open Science in PsyArXiv
Open AccessModerational Learning: A Framework for Discovering Models and IndividualDifferences from Behavioral Data
Understanding how people respond to stimuli requires models capturing both common psychological mechanisms and individual variability. Here we introduce moderational learning, a data-driven unsupervised machine-learning framework that aims to jointly discover interpretable latent dimensions representing individual variability and shared generative functions that map stimuli to responses. We implement the framework combining amortized variational inference to infer individual representations with a conditional generative decoder for predicting responses. By combining posterior analyses with symbolic regression, we identify interpretable individual-difference factors and derive explicit computational models from behavioral data. We then extend the approach to cross-domain moderational learning, allowing latent representations to be inferred from multiple behavioral tasks or data modalities. Two simulation studies show that the approach captures stimulus--response mappings and recovers underlying latent structures. As a testbed, we applied moderational learning to a behavioral dataset in which 507 participants completed four value-based decision-making tasks. In empirical analyses, moderational learning achieved higher held-out predictive performance than the comparison models and yielded exploratory within-task and cross-task latent structures, with dimensions corresponding to risk discounting, delay discounting, and bidding proclivity. Sensitivity analyses further reveal that responses in pricing tasks do not always decrease monotonically with aversive attributes. Together, these results demonstrate that moderational learning provides a data-driven framework for discovering latent behavioral structure and generative mechanisms, opening a pathway toward automated theory development in psychological science.
- 18 June 2026
- Published by American Psychological Association (APA) in Psychological Methods
Scaling cognitive modeling to big data: A deep learning approach to studying individual differences in evidence accumulation model parameters.
Recent advances in Bayesian modeling and deep learning have enabled scalable estimation of cognitive process models. In this article, we present a fully Bayesian workflow that leverages amortized inference with neural networks to rapidly estimate individual parameters and compare models from big behavioral data. Using data from a large online implicit association test sample (N > 5,000,000), we investigate how latent parameters, such as drift rate, boundary separation, nondecision times, and their variabilities, relate to key socioeconomic variables. Our exploratory findings reveal small but consistent associations of cognitive model parameters with socioeconomic covariates. Notably, trial-by-trial variability in drift rate, often ignored in prior work, emerged as the strongest predictor across all socioeconomic covariates. Our primary contribution lies in illustrating how deep learning-based Bayesian estimation and model comparison can be applied to mine robust insights from large and noisy behavioral data sets. We discuss limitations and implications for modeling individual differences in large-scale data sets and provide an open pipeline for future use. This work exemplifies how the emerging field of behavioral data science can extend cognitive modeling to new domains and support data-driven hypothesis generation targeting the cognitive underpinnings of individual differences. (PsycInfo Database Record (c) 2026 APA, all rights reserved).
- 18 June 2026
- Published by American Psychological Association (APA) in Psychological Methods
Scaling cognitive modeling to big data: A deep learning approach to studying individual differences in evidence accumulation model parameters.
Recent advances in Bayesian modeling and deep learning have enabled scalable estimation of cognitive process models. In this article, we present a fully Bayesian workflow that leverages amortized inference with neural networks to rapidly estimate individual parameters and compare models from big behavioral data. Using data from a large online implicit association test sample (N > 5,000,000), we investigate how latent parameters, such as drift rate, boundary separation, nondecision times, and their variabilities, relate to key socioeconomic variables. Our exploratory findings reveal small but consistent associations of cognitive model parameters with socioeconomic covariates. Notably, trial-by-trial variability in drift rate, often ignored in prior work, emerged as the strongest predictor across all socioeconomic covariates. Our primary contribution lies in illustrating how deep learning-based Bayesian estimation and model comparison can be applied to mine robust insights from large and noisy behavioral data sets. We discuss limitations and implications for modeling individual differences in large-scale data sets and provide an open pipeline for future use. This work exemplifies how the emerging field of behavioral data science can extend cognitive modeling to new domains and support data-driven hypothesis generation targeting the cognitive underpinnings of individual differences. (PsycInfo Database Record (c) 2026 APA, all rights reserved).
- Open Access
- 6 April 2026
- Published by Springer Nature in Computational Brain & Behavior
- p. 1-25
Amortized Bayesian Workflow for Modeling Congruency Effects Using the Diffusion Model for Conflict Tasks
The congruency effect, characterized by faster and more accurate performance in congruent compared to incongruent trials, has been consistently observed in various conflict tasks. However, the search for common underlying attentional control processes reveals significant differences between conflict tasks, as evident from different time courses of the congruency effect (i.e., delta functions) and a lack of correlations in performance across different conflict tasks. To address these issues, the Diffusion Model for Conflict Tasks (DMC) has been employed to explicitly model congruency effects and explain different delta function shapes. However, estimating the parameters of the DMC using standard iterative methods faces computational challenges. In this study, we leverage Amortized Bayesian Inference (ABI) with deep learning to address these limitations. We use different prior distributions, levels of model complexity, and automated hyperparameter optimization within a principled workflow. ABI produces well-calibrated posteriors across a wide range of trial counts, estimating individual-level parameters as accurately as iterative methods, while requiring only a fraction of their runtime once the initial training phase is complete. Additionally, we selectively manipulated irrelevant information strength in a flanker task to assess the sensitivity of DMC parameters to experimental effects. As expected, the amplitude of the automatic process was influenced by stimulus spacing in a flanker paradigm. However, the effect on the peak latency parameter depended on the choice of model parameterization and prior hyperparameters. Taken together, we demonstrate the viability of ABI as a scalable, accurate alternative to conventional estimation techniques and provide a reproducible workflow for future cognitive modeling studies.
- 6 April 2026
- Published by Springer Nature in Computational Brain & Behavior
- p. 1-25
Open AccessAmortized Bayesian Workflow for Modeling Congruency Effects Using the Diffusion Model for Conflict Tasks
The congruency effect, characterized by faster and more accurate performance in congruent compared to incongruent trials, has been consistently observed in various conflict tasks. However, the search for common underlying attentional control processes reveals significant differences between conflict tasks, as evident from different time courses of the congruency effect (i.e., delta functions) and a lack of correlations in performance across different conflict tasks. To address these issues, the Diffusion Model for Conflict Tasks (DMC) has been employed to explicitly model congruency effects and explain different delta function shapes. However, estimating the parameters of the DMC using standard iterative methods faces computational challenges. In this study, we leverage Amortized Bayesian Inference (ABI) with deep learning to address these limitations. We use different prior distributions, levels of model complexity, and automated hyperparameter optimization within a principled workflow. ABI produces well-calibrated posteriors across a wide range of trial counts, estimating individual-level parameters as accurately as iterative methods, while requiring only a fraction of their runtime once the initial training phase is complete. Additionally, we selectively manipulated irrelevant information strength in a flanker task to assess the sensitivity of DMC parameters to experimental effects. As expected, the amplitude of the automatic process was influenced by stimulus spacing in a flanker paradigm. However, the effect on the peak latency parameter depended on the choice of model parameterization and prior hyperparameters. Taken together, we demonstrate the viability of ABI as a scalable, accurate alternative to conventional estimation techniques and provide a reproducible workflow for future cognitive modeling studies.
- Open Access
- 27 March 2026
- Published by Center for Open Science in OSF Preprints
Scaling Cognitive Modeling to Big Data: A Deep Learning Approach to Studying Individual Differences in Evidence Accumulation Model Parameters
Recent advances in Bayesian modeling and deep learning have enabled scalable estimationof cognitive process models. In this paper, we present a fully Bayesian workflow thatleverages amortized inference with neural networks to rapidly estimate individualparameters and compare models from big behavioral data. Using data from a large onlineimplicit association test (IAT) sample (N > 5, 000, 000), we investigate how latentparameters, such as drift rate, boundary separation, non-decision times, and theirvariabilities, relate to key socioeconomic variables. Our exploratory findings reveal smallbut consistent associations of cognitive parameters with socioeconomic covariates. Notably,trial-by-trial variability in drift rate, often ignored in prior work, emerged as the strongestpredictor across all socioeconomic covariates. Our primary contribution lies in illustratinghow deep learning-based Bayesian estimation and model comparison can be applied tomine robust insights from large and noisy behavioral datasets. We discuss limitations andimplications for modeling individual differences in large-scale datasets and provide an openpipeline for future use. This work exemplifies how the emerging field of behavioral datascience can extend cognitive modeling to new domains and support data-driven hypothesisgeneration targeting the cognitive underpinnings of individual differences
- 27 March 2026
- Published by Center for Open Science in OSF Preprints
Open AccessScaling Cognitive Modeling to Big Data: A Deep Learning Approach to Studying Individual Differences in Evidence Accumulation Model Parameters
Recent advances in Bayesian modeling and deep learning have enabled scalable estimationof cognitive process models. In this paper, we present a fully Bayesian workflow thatleverages amortized inference with neural networks to rapidly estimate individualparameters and compare models from big behavioral data. Using data from a large onlineimplicit association test (IAT) sample (N > 5, 000, 000), we investigate how latentparameters, such as drift rate, boundary separation, non-decision times, and theirvariabilities, relate to key socioeconomic variables. Our exploratory findings reveal smallbut consistent associations of cognitive parameters with socioeconomic covariates. Notably,trial-by-trial variability in drift rate, often ignored in prior work, emerged as the strongestpredictor across all socioeconomic covariates. Our primary contribution lies in illustratinghow deep learning-based Bayesian estimation and model comparison can be applied tomine robust insights from large and noisy behavioral datasets. We discuss limitations andimplications for modeling individual differences in large-scale datasets and provide an openpipeline for future use. This work exemplifies how the emerging field of behavioral datascience can extend cognitive modeling to new domains and support data-driven hypothesisgeneration targeting the cognitive underpinnings of individual differences
- Open Access
- 24 January 2026
- Published by Center for Open Science in PsyArXiv
Amortized Bayesian Workflow for Modeling Congruency Effects Using the Diffusion Model for Conflict Tasks
The congruency effect, characterized by faster and more accurate performance in congruent compared to incongruent trials, has been consistently observed in various conflict tasks. However, the search for common underlying attentional control processes reveals significant differences between conflict tasks, as evident from different time courses of the congruency effect (i.e., delta functions) and a lack of correlations in performance across different conflict tasks. To address these issues, the Diffusion Model for Conflict Tasks (DMC) has been employed to explicitly model congruency effects and explain different delta function shapes. However, estimating the parameters of the DMC using standard iterative methods faces computational challenges. In this study, we leverage Amortized Bayesian Inference (ABI) with deep learning to address these limitations. We use different prior distributions, levels of model complexity, and automated hyperparameter optimization within a principled workflow. ABI produces well-calibrated posteriors across a wide range of trial counts, estimating individual-level parameters as accurately as iterative methods, while requiring only a fraction of their runtime once the initial training phase is complete. Additionally, we selectively manipulated irrelevant information strength in a flanker task to assess the sensitivity of DMC parameters to experimental effects. As expected, the amplitude of the automatic process was influenced by stimulus spacing in a flanker paradigm. However, the effect on the peak latency parameter depended on the choice of model parameterization and prior hyperparameters. Taken together, we demonstrate the viability of ABI as a scalable, accurate alternative to conventional estimation techniques and provide a reproducible workflow for future cognitive modeling studies.
- 24 January 2026
- Published by Center for Open Science in PsyArXiv
Open AccessAmortized Bayesian Workflow for Modeling Congruency Effects Using the Diffusion Model for Conflict Tasks
The congruency effect, characterized by faster and more accurate performance in congruent compared to incongruent trials, has been consistently observed in various conflict tasks. However, the search for common underlying attentional control processes reveals significant differences between conflict tasks, as evident from different time courses of the congruency effect (i.e., delta functions) and a lack of correlations in performance across different conflict tasks. To address these issues, the Diffusion Model for Conflict Tasks (DMC) has been employed to explicitly model congruency effects and explain different delta function shapes. However, estimating the parameters of the DMC using standard iterative methods faces computational challenges. In this study, we leverage Amortized Bayesian Inference (ABI) with deep learning to address these limitations. We use different prior distributions, levels of model complexity, and automated hyperparameter optimization within a principled workflow. ABI produces well-calibrated posteriors across a wide range of trial counts, estimating individual-level parameters as accurately as iterative methods, while requiring only a fraction of their runtime once the initial training phase is complete. Additionally, we selectively manipulated irrelevant information strength in a flanker task to assess the sensitivity of DMC parameters to experimental effects. As expected, the amplitude of the automatic process was influenced by stimulus spacing in a flanker paradigm. However, the effect on the peak latency parameter depended on the choice of model parameterization and prior hyperparameters. Taken together, we demonstrate the viability of ABI as a scalable, accurate alternative to conventional estimation techniques and provide a reproducible workflow for future cognitive modeling studies.
- 19 June 2025
- Published by Springer Nature in Statistics and Computing
- Vol. 35 (5) , 1-33
Expert-elicitation method for non-parametric joint priors using normalizing flows
- 19 June 2025
- Published by Springer Nature in Statistics and Computing
- Vol. 35 (5) , 1-33
Expert-elicitation method for non-parametric joint priors using normalizing flows
- Open Access
Real-Time Wide-Field Fluorescence Lifetime Imaging via Single-Snapshot Acquisition for Biomedical Applications
Fluorescence lifetime imaging (FLI) is a powerful tool for investigating molecular processes, microenvironmental parameters, and molecular interactions across tissue to (sub-)cellular levels. Despite its established value in numerous biomedical applications, conventional FLI techniques are hindered by long acquisition times. This limitation restricts their use in real-time scenarios, such as monitoring fast biological processes, studying live organisms, and in environments that require rapid imaging and immediate inference, such as clinical image-guided interventions. Here, we present a novel FLI approach that combines a large-format time-gated SPAD array with dual-gate acquisition capability, alongside a rapid lifetime determination algorithm. This integration allows for real-time fluorescence lifetime estimation through single-snapshot acquisitions, eliminating the need for traditional, time-consuming time-resolved data collection. We demonstrate the scalability and versatility of this method by achieving real-time FLI across challenging biomedical applications, ranging from capturing fast neural dynamics at the microscopic scale, performing multimodal 3D volumetric FLI of tumor organoids at the mesoscopic scale, to macroscale FLI in both direct and highly scattering regimes. Furthermore, we validate its utility in fluorescence lifetime-guided surgical procedures using tissue-mimicking phantoms. Overall, this new methodology significantly enhances the temporal and spatial capabilities of FLI, opening the door to the assessment of fast dynamic biomedical signals. It also enables the seamless integration of FLI into clinical workflows, particularly in applications like fluorescence-guided surgery.
Open AccessReal-Time Wide-Field Fluorescence Lifetime Imaging via Single-Snapshot Acquisition for Biomedical Applications
Fluorescence lifetime imaging (FLI) is a powerful tool for investigating molecular processes, microenvironmental parameters, and molecular interactions across tissue to (sub-)cellular levels. Despite its established value in numerous biomedical applications, conventional FLI techniques are hindered by long acquisition times. This limitation restricts their use in real-time scenarios, such as monitoring fast biological processes, studying live organisms, and in environments that require rapid imaging and immediate inference, such as clinical image-guided interventions. Here, we present a novel FLI approach that combines a large-format time-gated SPAD array with dual-gate acquisition capability, alongside a rapid lifetime determination algorithm. This integration allows for real-time fluorescence lifetime estimation through single-snapshot acquisitions, eliminating the need for traditional, time-consuming time-resolved data collection. We demonstrate the scalability and versatility of this method by achieving real-time FLI across challenging biomedical applications, ranging from capturing fast neural dynamics at the microscopic scale, performing multimodal 3D volumetric FLI of tumor organoids at the mesoscopic scale, to macroscale FLI in both direct and highly scattering regimes. Furthermore, we validate its utility in fluorescence lifetime-guided surgical procedures using tissue-mimicking phantoms. Overall, this new methodology significantly enhances the temporal and spatial capabilities of FLI, opening the door to the assessment of fast dynamic biomedical signals. It also enables the seamless integration of FLI into clinical workflows, particularly in applications like fluorescence-guided surgery.
- Open Access
- 8 October 2024
- Published by Springer Nature in Computational Brain & Behavior
- Vol. 8 (2) , 191-210
Validation and Comparison of Non-stationary Cognitive Models: A Diffusion Model Application
Cognitive processes undergo various fluctuations and transient states across different temporal scales. Superstatistics are emerging as a flexible framework for incorporating such non-stationary dynamics into existing cognitive model classes. In this work, we provide the first experimental validation of superstatistics and formal comparison of four non-stationary diffusion decision models in a specifically designed perceptual decision-making task. Task difficulty and speed-accuracy trade-off were systematically manipulated to induce expected changes in model parameters. To validate our models, we assess whether the inferred parameter trajectories align with the patterns and sequences of the experimental manipulations. To address computational challenges, we present novel deep learning techniques for amortized Bayesian estimation and comparison of models with time-varying parameters. Our findings indicate that transition models incorporating both gradual and abrupt parameter shifts provide the best fit to the empirical data. Moreover, we find that the inferred parameter trajectories closely mirror the sequence of experimental manipulations. Posterior re-simulations further underscore the ability of the models to faithfully reproduce critical data patterns. Accordingly, our results suggest that the inferred non-stationary dynamics may reflect actual changes in the targeted psychological constructs. We argue that our initial experimental validation paves the way for the widespread application of superstatistics in cognitive modeling and beyond.
- 8 October 2024
- Published by Springer Nature in Computational Brain & Behavior
- Vol. 8 (2) , 191-210
Open AccessValidation and Comparison of Non-stationary Cognitive Models: A Diffusion Model Application
Cognitive processes undergo various fluctuations and transient states across different temporal scales. Superstatistics are emerging as a flexible framework for incorporating such non-stationary dynamics into existing cognitive model classes. In this work, we provide the first experimental validation of superstatistics and formal comparison of four non-stationary diffusion decision models in a specifically designed perceptual decision-making task. Task difficulty and speed-accuracy trade-off were systematically manipulated to induce expected changes in model parameters. To validate our models, we assess whether the inferred parameter trajectories align with the patterns and sequences of the experimental manipulations. To address computational challenges, we present novel deep learning techniques for amortized Bayesian estimation and comparison of models with time-varying parameters. Our findings indicate that transition models incorporating both gradual and abrupt parameter shifts provide the best fit to the empirical data. Moreover, we find that the inferred parameter trajectories closely mirror the sequence of experimental manipulations. Posterior re-simulations further underscore the ability of the models to faithfully reproduce critical data patterns. Accordingly, our results suggest that the inferred non-stationary dynamics may reflect actual changes in the targeted psychological constructs. We argue that our initial experimental validation paves the way for the widespread application of superstatistics in cognitive modeling and beyond.
- 6 May 2024
- Published by American Psychological Association (APA) in Psychological Methods
- Vol. 31 (5) , 917-945
A deep learning method for comparing Bayesian hierarchical models.
Bayesian model comparison (BMC) offers a principled approach to assessing the relative merits of competing computational models and propagating uncertainty into model selection decisions. However, BMC is often intractable for the popular class of hierarchical models due to their high-dimensional nested parameter structure. To address this intractability, we propose a deep learning method for performing BMC on any set of hierarchical models which can be instantiated as probabilistic programs. Since our method enables amortized inference, it allows efficient re-estimation of posterior model probabilities and fast performance validation prior to any real-data application. In a series of extensive validation studies, we benchmark the performance of our method against the state-of-the-art bridge sampling method and demonstrate excellent amortized inference across all BMC settings. We then showcase our method by comparing four hierarchical evidence accumulation models that have previously been deemed intractable for BMC due to partly implicit likelihoods. Additionally, we demonstrate how transfer learning can be leveraged to enhance training efficiency. We provide reproducible code for all analyses and an open-source implementation of our method. (PsycInfo Database Record (c) 2024 APA, all rights reserved).
- 6 May 2024
- Published by American Psychological Association (APA) in Psychological Methods
- Vol. 31 (5) , 917-945
A deep learning method for comparing Bayesian hierarchical models.
Bayesian model comparison (BMC) offers a principled approach to assessing the relative merits of competing computational models and propagating uncertainty into model selection decisions. However, BMC is often intractable for the popular class of hierarchical models due to their high-dimensional nested parameter structure. To address this intractability, we propose a deep learning method for performing BMC on any set of hierarchical models which can be instantiated as probabilistic programs. Since our method enables amortized inference, it allows efficient re-estimation of posterior model probabilities and fast performance validation prior to any real-data application. In a series of extensive validation studies, we benchmark the performance of our method against the state-of-the-art bridge sampling method and demonstrate excellent amortized inference across all BMC settings. We then showcase our method by comparing four hierarchical evidence accumulation models that have previously been deemed intractable for BMC due to partly implicit likelihoods. Additionally, we demonstrate how transfer learning can be leveraged to enhance training efficiency. We provide reproducible code for all analyses and an open-source implementation of our method. (PsycInfo Database Record (c) 2024 APA, all rights reserved).