In-context learning is a remarkable capability of transformers, referring to their ability to adapt to specific tasks based on a short history or context. Previous research has found that task-specific information is locally encoded within models, though their emergence and functionality remain unclear due to opaque pre-training processes. In this work, we investigate the formation of task vectors in a controlled setting, using models trained from scratch on synthetic datasets. Our findings confirm that task vectors naturally emerge under certain conditions, but the tasks may be relatively weakly and/or non-locally encoded within the model. To promote strong task vectors encoded at a prescribed location within the model, we propose an auxiliary training mechanism based on a task vector prompting loss (TVP-loss). This method eliminates the need to search for task-correlated encodings within the trained model and demonstrably improves robustness and generalization.

We propose a novel cooperative multi-agent reinforcement learning (MARL) approach for networked agents. In contrast to previous methods that rely on complete state information or joint observations, our agents must learn how to reach shared objectives under partial observability. During training, they collect individual rewards and approximate a team value function through local communication, resulting in cooperative behavior. To describe our problem, we introduce the networked dynamic partially observable Markov game framework, where agents communicate over a switching topology communication network. Our distributed method, DNA-MARL, uses a consensus mechanism for local communication and gradient descent for local computation. DNA-MARL increases the range of the possible applications of networked agents, being well-suited for real world domains that impose privacy and where the messages may not reach their recipients. We evaluate DNA-MARL across benchmark MARL scenarios. Our results highlight the superior performance of DNA-MARL over previous methods.

Generative AI systems like foundation models (FMs) must align well with human values to ensure their behavior is helpful and trustworthy. While Reinforcement Learning from Human Feedback (RLHF) has shown promise for optimizing model performance using human judgments, existing RLHF pipelines predominantly rely on immediate feedback, which can fail to accurately reflect the downstream impact of an interaction on users' utility. We demonstrate that feedback based on evaluators' foresight estimates of downstream consequences systematically induces Goodhart's Law dynamics, incentivizing misaligned behaviors like sycophancy and deception and ultimately degrading user outcomes. To alleviate this, we propose decoupling evaluation from prediction by refocusing RLHF on hindsight feedback. Our theoretical analysis reveals that conditioning evaluator feedback on downstream observations mitigates misalignment and improves expected human utility, even when these observations are simulated by the AI system itself. To leverage this insight in a practical alignment algorithm, we introduce Reinforcement Learning from Hindsight Simulation (RLHS), which first simulates plausible consequences and then elicits feedback to assess what behaviors were genuinely beneficial in hindsight. We apply RLHS to two widely-employed online and offline preference optimization methods -- Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO) -- and show empirically that misalignment is significantly reduced with both methods. Through an online human user study, we show that RLHS consistently outperforms RLHF in helping users achieve their goals and earns higher satisfaction ratings, despite being trained solely with simulated hindsight feedback. These results underscore the importance of focusing on long-term consequences, even simulated ones, to mitigate misalignment in RLHF.

Classifying genome sequences based on metadata has been an active area of research in comparative genomics for decades with many important applications across the life sciences. Established methods for classifying genomes can be broadly grouped into sequence alignment-based and alignment-free models. Conventional alignment-based models rely on genome similarity measures calculated based on local sequence alignments or consistent ordering among sequences. However, such methods are computationally expensive when dealing with large ensembles of even moderately sized genomes. In contrast, alignment-free (AF) approaches measure genome similarity based on summary statistics in an unsupervised setting and are efficient enough to analyze large datasets. However, both alignment-based and AF methods typically assume fixed scoring rubrics that lack the flexibility to assign varying importance to different parts of the sequences based on prior knowledge. In this study, we integrate AI and network science approaches to develop a comparative genomic analysis framework that addresses these limitations. Our approach, termed the Genome Misclassification Network Analysis (GMNA), simultaneously leverages misclassified instances, a learned scoring rubric, and label information to classify genomes based on associated metadata and better understand potential drivers of misclassification. We evaluate the utility of the GMNA using Naive Bayes and convolutional neural network models, supplemented by additional experiments with transformer-based models, to construct SARS-CoV-2 sampling location classifiers using over 500,000 viral genome sequences and study the resulting network of misclassifications. We demonstrate the global health potential of the GMNA by leveraging the SARS-CoV-2 genome misclassification networks to investigate the role human mobility played in structuring geographic clustering of SARS-CoV-2.

Introduction: Substance use disorders (SUDs) have emerged as a pressing public health crisis in the United States, with adolescent substance use often leading to SUDs in adulthood. Effective strategies are needed to prevent this progression. To help in filling this need, we develop a novel and the first-ever absolute risk prediction model for cannabis use disorder (CUD) for adolescent or young adult cannabis users. Methods: We train a Bayesian machine learning model that provides a personalized CUD absolute risk for adolescent or young adult cannabis users using data from the National Longitudinal Study of Adolescent to Adult Health. Model performance is assessed using 5-fold cross-validation (CV) with area under the curve (AUC) and ratio of the expected to observed number of cases (E/O). External validation of the final model is conducted using two independent datasets. Results: The proposed model has five risk factors: biological sex, delinquency, and scores on personality traits of conscientiousness, neuroticism, and openness. For predicting CUD risk within five years of first cannabis use, AUC and E/O, computed via 5-fold CV, were 0.68 and 0.95, respectively. For the same type of prediction in external validation, AUC values were 0.64 and 0.75, with E/O values of 0.98 and 1, indicating good discrimination and calibration performances of the model. Discussion and Conclusion: The proposed model is the first absolute risk prediction model for an SUD. It can aid clinicians in identifying adolescent/youth substance users at a high risk of developing CUD in future for clinically appropriate interventions.

Fixpoints are ubiquitous in computer science and when dealing with quantitative semantics and verification one is commonly led to consider least fixpoints of (higher-dimensional) functions over the nonnegative reals. We show how to approximate the least fixpoint of such functions, focusing on the case in which they are not known precisely, but represented by a sequence of approximating functions that converge to them. We concentrate on monotone and non-expansive functions, for which uniqueness of fixpoints is not guaranteed and standard fixpoint iteration schemes might get stuck at a fixpoint that is not the least. Our main contribution is the identification of an iteration scheme, a variation of Mann iteration with a dampening factor, which, under suitable conditions, is shown to guarantee convergence to the least fixpoint of the function of interest. We then argue that these results are relevant in the context of model-based reinforcement learning for Markov decision processes (MDPs), showing that the proposed iteration scheme instantiates to MDPs and allows us to derive convergence to the optimal expected return. More generally, we show that our results can be used to iterate to the least fixpoint almost surely for systems where the function of interest can be approximated with given probabilistic error bounds, as it happens for probabilistic systems, such as simple stochastic games, that can be explored via sampling.

A key challenge in training Large Language Models (LLMs) is properly aligning them with human preferences. Reinforcement Learning with Human Feedback (RLHF) uses pairwise comparisons from human annotators to train reward functions and has emerged as a popular alignment method. However, input datasets in RLHF are not necessarily balanced in the types of questions and answers that are included. Therefore, we want RLHF algorithms to perform well even when the set of alternatives is not uniformly distributed. Drawing on insights from social choice theory, we introduce robustness to approximate clones, a desirable property of RLHF algorithms which requires that adding near-duplicate alternatives does not significantly change the learned reward function. We first demonstrate that the standard RLHF algorithm based on regularized maximum likelihood estimation (MLE) fails to satisfy this property. We then propose the weighted MLE, a new RLHF algorithm that modifies the standard regularized MLE by weighting alternatives based on their similarity to other alternatives. This new algorithm guarantees robustness to approximate clones while preserving desirable theoretical properties.

The development of neural sequence models, such as Transformers [Vaswani et al., 2017], along with the improvements in large-scale self-supervised learning, has opened the door to universal language understanding and generation. This achievement is largely motivated by pre-training: we separate common components from many neural network-based systems, and then train them on huge amounts of unlabeled data using self-supervision. These pre-trained models serve as foundation models that can be easily adapted to different tasks via fine-tuning or prompting. As a result, the paradigm of NLP has been enormously changed. In many cases, large-scale supervised learning for specific tasks is no longer required, and instead, we only need to adapt pre-trained foundation models.

We consider a high-dimensional mean estimation problem over a binary hidden Markov model, which illuminates the interplay between memory in data, sample size, dimension, and signal strength in statistical inference. In this model, an estimator observes n samples of a d -dimensional parameter vector \theta_{*}\in\mathbb{R} {d}, multiplied by a random sign S_i ( 1\le i\le n), and corrupted by isotropic standard Gaussian noise. As \delta varies, this model smoothly interpolates two well-studied models: the Gaussian Location Model for which \delta 0 and the Gaussian Mixture Model for which \delta 1/2 . We then provide an upper bound to the case of estimating \delta, assuming a (possibly inaccurate) knowledge of \theta_{*} . The bound is proved to be tight when \theta_{*} is an accurately known constant.

We propose that inherent stochasticity enables synaptic plasticity to carry out probabilistic inference by sampling from a posterior distribution of synaptic parameters. This view provides a viable alternative to existing models that propose convergence of synaptic weights to maximum likelihood parameters. It explains how priors on weight distributions and connection probabilities can be merged optimally with learned experience. In simulations we show that our model for synaptic plasticity allows spiking neural networks to compensate continuously for unforeseen disturbances. Furthermore it provides a normative mathematical framework to better understand the permanent variability and rewiring observed in brain networks.