arXiv subjects

Benjamin M. Marlin

Publications and source records attributed to Benjamin M. Marlin.

At least 19 recordsLinked to original sources

arXiv2026

Inducing Emergent Misalignment from Reward Hacks with Iterative DPO

Reward hacking during reinforcement learning from verifiable rewards (RLVR) can induce reward seeking and broad misalignment in language models. Studying this misgeneralization is important for developing better threat models and countermeasures, but is often infeasible due to the cost of RL on large models. As an alternative, we propose studying emergent misalignment from iterative DPO, which preserves important properties of RLVR while reducing costs and enabling training on popular finetuning APIs. In practice, we find that training GPT-4.1 with iterative DPO on a single-turn reward hacking environment induces covert misaligned power-seeking and alignment faking, the first openly available (semi)-online training pipeline to induce these concerning forms of misalignment. We also find that training Qwen2.5-32B-Instruct with the same pipeline induces both misalignment and improved instruction following accuracy, showing that iterative DPO can be used as a testbed for selective generalization. Overall, we think iterative DPO can help democratize and accelerate the study of emergent misalignment from RLVR.

cs.LG↗

arXiv2026

Probing the 3D Object-Level Understanding of Pre-Trained Detection Transformers

Detection transformer models, including DETR and its extensions, learn to output a set of object-level embeddings that can be simultaneously decoded into 2D bounding boxes and class distributions. In this paper, we investigate what pre-trained 2D detection transformers understand about the 3D properties of objects. Specifically, we investigate the extent to which properties including the depth of objects from the camera and the 3D location of objects relative to the camera can be recovered from object-level embeddings using linear and non-linear probes. Across a range of detection transformer models, our results show a surprisingly strong and previously unknown ability of 2D DETR models to represent useful information about the 3D properties of objects, despite the complete lack of 3D supervision during model pre-training.

cs.CV↗

arXiv2026

Stress-Testing Alignment Audits With Prompt-Level Strategic Deception

Alignment audits aim to robustly identify hidden goals from strategic, situationally aware misaligned models. Despite this threat model, existing auditing methods have not been systematically stress-tested against deception strategies. We address this gap, implementing an automatic red-team pipeline that generates deception strategies (in the form of system prompts) tailored to specific white-box and black-box auditing methods. Stress-testing assistant prefills, user persona sampling, sparse autoencoders, and token embedding similarity methods against secret-keeping model organisms, our automatic red-team pipeline finds prompts that deceive both the black-box and white-box methods into confident, incorrect guesses. Our results provide the first documented evidence of activation-based strategic deception, and suggest that current black-box and white-box methods would not be robust to a sufficiently capable misaligned model.

cs.LG↗

arXiv2025

GPU-GLMB: Assessing the Scalability of GPU-Accelerated Multi-Hypothesis Tracking

Much recent research on multi-target tracking has focused on multi-hypothesis approaches leveraging random finite sets. Of particular interest are labeled random finite set methods that maintain temporally coherent labels for each object. While these methods enjoy important theoretical properties as closed-form solutions to the multi-target Bayes filter, the maintenance of multiple hypotheses under the standard measurement model is highly computationally expensive, even when hypothesis pruning approximations are applied. In this work, we focus on the Generalized Labeled Multi-Bernoulli (GLMB) filter as an example of this class of methods. We investigate a variant of the filter that allows multiple detections per object from the same sensor, a critical capability when deploying tracking in the context of distributed networks of machine learning-based virtual sensors. We show that this breaks the inter-detection dependencies in the filter updates of the standard GLMB filter, allowing updates with significantly improved parallel scalability and enabling efficient deployment on GPU hardware. We report the results of a preliminary analysis of a GPU-accelerated implementation of our proposed GLMB tracker, with a focus on run time scalability with respect to the number of objects and the maximum number of retained hypotheses.

cs.CV↗

arXiv2025

SigmaScheduling: Uncertainty-Informed Scheduling of Decision Points for Intelligent Mobile Health Interventions

Timely decision making is critical to the effectiveness of mobile health (mHealth) interventions. At predefined timepoints called "decision points," intelligent mHealth systems such as just-in-time adaptive interventions (JITAIs) estimate an individual's biobehavioral context from sensor or survey data and determine whether and how to intervene. For interventions targeting habitual behavior (e.g., oral hygiene), effectiveness often hinges on delivering support shortly before the target behavior is likely to occur. Current practice schedules decision points at a fixed interval (e.g., one hour) before user-provided behavior times, and the fixed interval is kept the same for all individuals. However, this one-size-fits-all approach performs poorly for individuals with irregular routines, often scheduling decision points after the target behavior has already occurred, rendering interventions ineffective. In this paper, we propose SigmaScheduling, a method to dynamically schedule decision points based on uncertainty in predicted behavior times. When behavior timing is more predictable, SigmaScheduling schedules decision points closer to the predicted behavior time; when timing is less certain, SigmaScheduling schedules decision points earlier, increasing the likelihood of timely intervention. We evaluated SigmaScheduling using real-world data from 68 participants in a 10-week trial of Oralytics, a JITAI designed to improve daily toothbrushing. SigmaScheduling increased the likelihood that decision points preceded brushing events in at least 70% of cases, preserving opportunities to intervene and impact behavior. Our results indicate that SigmaScheduling can advance precision mHealth, particularly for JITAIs targeting time-sensitive, habitual behaviors such as oral hygiene or dietary habits.

cs.AI↗

arXiv2025

ACE and Diverse Generalization via Selective Disagreement

Deep neural networks are notoriously sensitive to spurious correlations - where a model learns a shortcut that fails out-of-distribution. Existing work on spurious correlations has often focused on incomplete correlations,leveraging access to labeled instances that break the correlation. But in cases where the spurious correlations are complete, the correct generalization is fundamentally \textit{underspecified}. To resolve this underspecification, we propose learning a set of concepts that are consistent with training data but make distinct predictions on a subset of novel unlabeled inputs. Using a self-training approach that encourages \textit{confident} and \textit{selective} disagreement, our method ACE matches or outperforms existing methods on a suite of complete-spurious correlation benchmarks, while remaining robust to incomplete spurious correlations. ACE is also more configurable than prior approaches, allowing for straight-forward encoding of prior knowledge and principled unsupervised model selection. In an early application to language-model alignment, we find that ACE achieves competitive performance on the measurement tampering detection benchmark \textit{without} access to untrusted measurements. While still subject to important limitations, ACE represents significant progress towards overcoming underspecification.

cs.LG↗

arXiv2025

Enhancing Adaptive Behavioral Interventions with LLM Inference from Participant-Described States

The use of reinforcement learning (RL) methods to support health behavior change via personalized and just-in-time adaptive interventions is of significant interest to health and behavioral science researchers focused on problems such as smoking cessation support and physical activity promotion. However, RL methods are often applied to these domains using a small collection of context variables to mitigate the significant data scarcity issues that arise from practical limitations on the design of adaptive intervention trials. In this paper, we explore an approach to significantly expanding the state space of an adaptive intervention without impacting data efficiency. The proposed approach enables intervention participants to provide natural language descriptions of aspects of their current state. It then leverages inference with pre-trained large language models (LLMs) to better align the policy of a base RL method with these state descriptions. To evaluate our method, we develop a novel physical activity intervention simulation environment that generates text-based state descriptions conditioned on latent state variables using an auxiliary LLM. We show that this approach has the potential to significantly improve the performance of online policy learning methods.

cs.LG↗

arXiv2025

Combining LLM decision and RL action selection to improve RL policy for adaptive interventions

Reinforcement learning (RL) is increasingly being used in the healthcare domain, particularly for the development of personalized health adaptive interventions. Inspired by the success of Large Language Models (LLMs), we are interested in using LLMs to update the RL policy in real time, with the goal of accelerating personalization. We use the text-based user preference to influence the action selection on the fly, in order to immediately incorporate the user preference. We use the term "user preference" as a broad term to refer to a user personal preference, constraint, health status, or a statement expressing like or dislike, etc. Our novel approach is a hybrid method that combines the LLM response and the RL action selection to improve the RL policy. Given an LLM prompt that incorporates the user preference, the LLM acts as a filter in the typical RL action selection. We investigate different prompting strategies and action selection strategies. To evaluate our approach, we implement a simulation environment that generates the text-based user preferences and models the constraints that impact behavioral dynamics. We show that our approach is able to take into account the text-based user preferences, while improving the RL policy, thus improving personalization in adaptive intervention.

cs.LG↗

arXiv2024

BOTS: Batch Bayesian Optimization of Extended Thompson Sampling for Severely Episode-Limited RL Settings

In settings where the application of reinforcement learning (RL) requires running real-world trials, including the optimization of adaptive health interventions, the number of episodes available for learning can be severely limited due to cost or time constraints. In this setting, the bias-variance trade-off of contextual bandit methods can be significantly better than that of more complex full RL methods. However, Thompson sampling bandits are limited to selecting actions based on distributions of immediate rewards. In this paper, we extend the linear Thompson sampling bandit to select actions based on a state-action utility function consisting of the Thompson sampler's estimate of the expected immediate reward combined with an action bias term. We use batch Bayesian optimization over episodes to learn the action bias terms with the goal of maximizing the expected return of the extended Thompson sampler. The proposed approach is able to learn optimal policies for a strictly broader class of Markov decision processes (MDPs) than standard Thompson sampling. Using an adaptive intervention simulation environment that captures key aspects of behavioral dynamics, we show that the proposed method can significantly out-perform standard Thompson sampling in terms of total return, while requiring significantly fewer episodes than standard value function and policy gradient methods.

cs.LG↗

arXiv2024

Heteroscedastic Temporal Variational Autoencoder For Irregular Time Series

Irregularly sampled time series commonly occur in several domains where they present a significant challenge to standard deep learning models. In this paper, we propose a new deep learning framework for probabilistic interpolation of irregularly sampled time series that we call the Heteroscedastic Temporal Variational Autoencoder (HeTVAE). HeTVAE includes a novel input layer to encode information about input observation sparsity, a temporal VAE architecture to propagate uncertainty due to input sparsity, and a heteroscedastic output layer to enable variable uncertainty in output interpolations. Our results show that the proposed architecture is better able to reflect variable uncertainty through time due to sparse and irregular sampling than a range of baseline and traditional models, as well as recently proposed deep latent variable models that use homoscedastic output layers.

cs.LG↗

arXiv2024

StepCountJITAI: simulation environment for RL with application to physical activity adaptive intervention

The use of reinforcement learning (RL) to learn policies for just-in-time adaptive interventions (JITAIs) is of significant interest in many behavioral intervention domains including improving levels of physical activity. In a messaging-based physical activity JITAI, a mobile health app is typically used to send messages to a participant to encourage engagement in physical activity. In this setting, RL methods can be used to learn what intervention options to provide to a participant in different contexts. However, deploying RL methods in real physical activity adaptive interventions comes with challenges: the cost and time constraints of real intervention studies result in limited data to learn adaptive intervention policies. Further, commonly used RL simulation environments have dynamics that are of limited relevance to physical activity adaptive interventions and thus shed little light on what RL methods may be optimal for this challenging application domain. In this paper, we introduce StepCountJITAI, an RL environment designed to foster research on RL methods that address the significant challenges of policy learning for adaptive behavioral interventions.

cs.LG↗

arXiv2024

REBAR: Retrieval-Based Reconstruction for Time-series Contrastive Learning

The success of self-supervised contrastive learning hinges on identifying positive data pairs, such that when they are pushed together in embedding space, the space encodes useful information for subsequent downstream tasks. Constructing positive pairs is non-trivial as the pairing must be similar enough to reflect a shared semantic meaning, but different enough to capture within-class variation. Classical approaches in vision use augmentations to exploit well-established invariances to construct positive pairs, but invariances in the time-series domain are much less obvious. In our work, we propose a novel method of using a learned measure for identifying positive pairs. Our Retrieval-Based Reconstruction (REBAR) measure measures the similarity between two sequences as the reconstruction error that results from reconstructing one sequence with retrieved information from the other. Then, if the two sequences have high REBAR similarity, we label them as a positive pair. Through validation experiments, we show that the REBAR error is a predictor of mutual class membership. Once integrated into a contrastive learning framework, our REBAR method learns an embedding that achieves state-of-the-art performance on downstream tasks across various modalities.

cs.LG↗

arXiv2024

Temporally Multi-Scale Sparse Self-Attention for Physical Activity Data Imputation

Wearable sensors enable health researchers to continuously collect data pertaining to the physiological state of individuals in real-world settings. However, such data can be subject to extensive missingness due to a complex combination of factors. In this work, we study the problem of imputation of missing step count data, one of the most ubiquitous forms of wearable sensor data. We construct a novel and large scale data set consisting of a training set with over 3 million hourly step count observations and a test set with over 2.5 million hourly step count observations. We propose a domain knowledge-informed sparse self-attention model for this task that captures the temporal multi-scale nature of step-count data. We assess the performance of the model relative to baselines and conduct ablation studies to verify our specific model designs.

cs.LG↗

arXiv2023

Heteroskedastic Geospatial Tracking with Distributed Camera Networks

Visual object tracking has seen significant progress in recent years. However, the vast majority of this work focuses on tracking objects within the image plane of a single camera and ignores the uncertainty associated with predicted object locations. In this work, we focus on the geospatial object tracking problem using data from a distributed camera network. The goal is to predict an object's track in geospatial coordinates along with uncertainty over the object's location while respecting communication constraints that prohibit centralizing raw image data. We present a novel single-object geospatial tracking data set that includes high-accuracy ground truth object locations and video data from a network of four cameras. We present a modeling framework for addressing this task including a novel backbone model and explore how uncertainty calibration and fine-tuning through a differentiable tracker affect performance.

cs.CV↗

arXiv2023

Assessing the Impact of Context Inference Error and Partial Observability on RL Methods for Just-In-Time Adaptive Interventions

Just-in-Time Adaptive Interventions (JITAIs) are a class of personalized health interventions developed within the behavioral science community. JITAIs aim to provide the right type and amount of support by iteratively selecting a sequence of intervention options from a pre-defined set of components in response to each individual's time varying state. In this work, we explore the application of reinforcement learning methods to the problem of learning intervention option selection policies. We study the effect of context inference error and partial observability on the ability to learn effective policies. Our results show that the propagation of uncertainty from context inferences is critical to improving intervention efficacy as context uncertainty increases, while policy gradient algorithms can provide remarkable robustness to partially observed behavioral state information.

cs.LG↗

arXiv2022

BayesLDM: A Domain-Specific Language for Probabilistic Modeling of Longitudinal Data

In this paper we present BayesLDM, a system for Bayesian longitudinal data modeling consisting of a high-level modeling language with specific features for modeling complex multivariate time series data coupled with a compiler that can produce optimized probabilistic program code for performing inference in the specified model. BayesLDM supports modeling of Bayesian network models with a specific focus on the efficient, declarative specification of dynamic Bayesian Networks (DBNs). The BayesLDM compiler combines a model specification with inspection of available data and outputs code for performing Bayesian inference for unknown model parameters while simultaneously handling missing data. These capabilities have the potential to significantly accelerate iterative modeling workflows in domains that involve the analysis of complex longitudinal data by abstracting away the process of producing computationally efficient probabilistic inference code. We describe the BayesLDM system components, evaluate the efficiency of representation and inference optimizations and provide an illustrative example of the application of the system to analyzing heterogeneous and partially observed mobile health data.

cs.LG↗

arXiv2022

Impact of Parameter Sparsity on Stochastic Gradient MCMC Methods for Bayesian Deep Learning

Bayesian methods hold significant promise for improving the uncertainty quantification ability and robustness of deep neural network models. Recent research has seen the investigation of a number of approximate Bayesian inference methods for deep neural networks, building on both the variational Bayesian and Markov chain Monte Carlo (MCMC) frameworks. A fundamental issue with MCMC methods is that the improvements they enable are obtained at the expense of increased computation time and model storage costs. In this paper, we investigate the potential of sparse network structures to flexibly trade-off model storage costs and inference run time against predictive performance and uncertainty quantification ability. We use stochastic gradient MCMC methods as the core Bayesian inference method and consider a variety of approaches for selecting sparse network structures. Surprisingly, our results show that certain classes of randomly selected substructures can perform as well as substructures derived from state-of-the-art iterative pruning methods while drastically reducing model training times.

cs.LG↗

arXiv2021

Challenges and Opportunities in Approximate Bayesian Deep Learning for Intelligent IoT Systems

Approximate Bayesian deep learning methods hold significant promise for addressing several issues that occur when deploying deep learning components in intelligent systems, including mitigating the occurrence of over-confident errors and providing enhanced robustness to out of distribution examples. However, the computational requirements of existing approximate Bayesian inference methods can make them ill-suited for deployment in intelligent IoT systems that include lower-powered edge devices. In this paper, we present a range of approximate Bayesian inference methods for supervised deep learning and highlight the challenges and opportunities when applying these methods on current edge hardware. We highlight several potential solutions to decreasing model storage requirements and improving computational scalability, including model pruning and distillation methods.

cs.LG↗