Recent advances in artificial intelligence have been fueled by the development of foundation models such as BERT, GPT, T5, and Vision Transformers. These models are first pretrained on vast and diverse datasets and then adapted to specific downstream tasks, often with significantly less data. However, the mechanisms behind the success of this ubiquitous pretrain-then-adapt paradigm remain underexplored, particularly the characteristics of pretraining checkpoints that lend themselves to good downstream adaptation. We introduce a Bayesian model selection criterion, called the downstream free energy, which quantifies a checkpoint's adaptability by measuring the concentration of nearby favorable parameters for the downstream task. We demonstrate that this free energy criterion can be effectively implemented without access to the downstream data or prior knowledge of the downstream task. Furthermore, we provide empirical evidence that the free energy criterion reliably correlates with improved fine-tuning performance, offering a principled approach to predicting model adaptability. The advent of foundation models has significantly reshaped the landscape of modern machine learning (Bommasani et al., 2021).

Large language models have demonstrated impressive in-context learning (ICL) capability. However, it is still unclear how the underlying transformers accomplish it, especially in more complex scenarios. Toward this goal, several recent works studied how transformers learn fixed-order Markov chains (FOMC) in context, yet natural languages are more suitably modeled by variable-order Markov chains (VOMC), i.e., context trees (CTs). In this work, we study the ICL of VOMC by viewing language modeling as a form of data compression and focus on small alphabets and low-order VOMCs. This perspective allows us to leverage mature compression algorithms, such as context-tree weighting (CTW) and prediction by partial matching (PPM) algorithms as baselines, the former of which is Bayesian optimal for a class of CTW priors. We empirically observe a few phenomena: 1) Transformers can indeed learn to compress VOMC in-context, while PPM suffers significantly; 2) The performance of transformers is not very sensitive to the number of layers, and even a two-layer transformer can learn in-context quite well; and 3) Transformers trained and tested on non-CTW priors can significantly outperform the CTW algorithm. To explain these phenomena, we analyze the attention map of the transformers and extract two mechanisms, on which we provide two transformer constructions: 1) A construction with $D+2$ layers that can mimic the CTW algorithm accurately for CTs of maximum order $D$, 2) A 2-layer transformer that utilizes the feed-forward network for probability blending. One distinction from the FOMC setting is that a counting mechanism appears to play an important role. We implement these synthetic transformer layers and show that such hybrid transformers can match the ICL performance of transformers, and more interestingly, some of them can perform even better despite the much-reduced parameter sets.

While large language models (LLMs) show impressive decision-making abilities, current methods lack a mechanism for automatic self-improvement from errors during task execution. We propose LEAP, an iterative fine-tuning framework that continually improves LLM agents using feedback from AI expert teachers. Our key insight is to equip the expert teachers with a privileged state -- information that is available during training but hidden at test time. This allows even weak experts to provide precise guidance, significantly improving the student agent's performance without access to privileged information at test time. We evaluate LEAP on diverse decision-making benchmarks, including text-based games (ALFWorld), web navigation (WebShop), and interactive coding (Intercode Bash). Our experiments show that LEAP (1) outperforms behavior cloning and ReAct baselines (2) enables weak student models (e.g., Llama3-8B) to exceed the performance of strong teacher models (GPT4-o), and (3) allows weak models to self-improve using privileged versions of themselves. We also provide a theoretical analysis showing that LEAP's success hinges on balancing privileged information with the student's realizability, which we empirically validate. Our code is available at https://leap-llm.github.io

Biological tree analysis serves as a pivotal tool in uncovering the evolutionary and differentiation relationships among organisms, genes, and cells. Its applications span diverse fields including phylogenetics, developmental biology, ecology, and medicine. Traditional tree inference methods, while foundational in early studies, face increasing limitations in processing the large-scale, complex datasets generated by modern high-throughput technologies. Recent advances in deep learning offer promising solutions, providing enhanced data processing and pattern recognition capabilities. However, challenges remain, particularly in accurately representing the inherently discrete and non-Euclidean nature of biological trees. In this review, we first outline the key biological priors fundamental to phylogenetic and differentiation tree analyses, facilitating a deeper interdisciplinary understanding between deep learning researchers and biologists. We then systematically examine the commonly used data formats and databases, serving as a comprehensive resource for model testing and development. We provide a critical analysis of traditional tree generation methods, exploring their underlying biological assumptions, technical characteristics, and limitations. Current developments in deep learning-based tree generation are reviewed, highlighting both recent advancements and existing challenges. Furthermore, we discuss the diverse applications of biological trees across various biological domains. Finally, we propose potential future directions and trends in leveraging deep learning for biological tree research, aiming to guide further exploration and innovation in this field.

In autonomous driving, accurate motion prediction is essential for safe and efficient motion planning. To ensure safety, planners must rely on reliable uncertainty information about the predicted future behavior of surrounding agents, yet this aspect has received limited attention. This paper addresses the so-far neglected problem of uncertainty modeling in trajectory prediction. We adopt a holistic approach that focuses on uncertainty quantification, decomposition, and the influence of model composition. Our method is based on a theoretically grounded information-theoretic approach to measure uncertainty, allowing us to decompose total uncertainty into its aleatoric and epistemic components. We conduct extensive experiments on the nuScenes dataset to assess how different model architectures and configurations affect uncertainty quantification and model robustness.

Policy evaluation is an important instrument for the comparison of different algorithms in Reinforcement Learning (RL). Yet even a precise knowledge of the value function $V^{\pi}$ corresponding to a policy $\pi$ does not provide reliable information on how far is the policy $\pi$ from the optimal one. We present a novel model-free upper value iteration procedure $({\sf UVIP})$ that allows us to estimate the suboptimality gap $V^{\star}(x) - V^{\pi}(x)$ from above and to construct confidence intervals for $V^\star$. Our approach relies on upper bounds to the solution of the Bellman optimality equation via martingale approach. We provide theoretical guarantees for ${\sf UVIP}$ under general assumptions and illustrate its performance on a number of benchmark RL problems.

Many scientific fields collect longitudinal multivariate count data where the total number of counts is arbitrary (e.g., multinomial observations). These data are often called count compositional as the information in the data relates to the relative frequencies of the categories (Silverman et al., 2018). These data occur frequently in molecular biology (Espinoza et al., 2020), microbiome studies (Silverman et al., 2018; Joseph et al., 2020; Äijö et al., 2018), natural language processing (Linderman et al., 2015), biomedicine (Fokianos and Kedem, 2003), and social sciences (Cargnoni et al., 1997). Although the counting process used to collect these data is often modeled as multinomial, other sources of noise in the system being studied often lead to extra-multinomial variation. While some account for this extra-multinomial variability with multinomial-Dirichlet models (Mosimann, 1962), multinomial logistic-normal models are often superior, as they can account for both positive and negative covariation between multinomial categories (Aitchison and Shen, 1980; Cargnoni et al., 1997; Joseph et al., 2020; Silverman et al., 2018). Moreover, under suitable transformation (i.e., link function), the logistic-normal is multivariate Gaussian.

Restless multi-armed bandits (RMAB) has been widely used to model constrained sequential decision making problems, where the state of each restless arm evolves according to a Markov chain and each state transition generates a scalar reward. However, the success of RMAB crucially relies on the availability and quality of reward signals. Unfortunately, specifying an exact reward function in practice can be challenging and even infeasible. In this paper, we introduce Pref-RMAB, a new RMAB model in the presence of preference signals, where the decision maker only observes pairwise preference feedback rather than scalar reward from the activated arms at each decision epoch. Preference feedback, however, arguably contains less information than the scalar reward, which makes Pref-RMAB seemingly more difficult. To address this challenge, we present a direct online preference learning (DOPL) algorithm for Pref-RMAB to efficiently explore the unknown environments, adaptively collect preference data in an online manner, and directly leverage the preference feedback for decision-makings. We prove that DOPL yields a sublinear regret. To our best knowledge, this is the first algorithm to ensure $\tilde{\mathcal{O}}(\sqrt{T\ln T})$ regret for RMAB with preference feedback. Experimental results further demonstrate the effectiveness of DOPL.

The Gaussian process (GP) is a Bayesian nonparametric paradigm that is widely adopted for uncertainty quantification (UQ) in a number of safety-critical applications, including robotics, healthcare, as well as surveillance. The consistency of the resulting uncertainty values however, hinges on the premise that the learning function conforms to the properties specified by the GP model, such as smoothness, periodicity and more, which may not be satisfied in practice, especially with data arriving on the fly. To combat against such model mis-specification, we propose to wed the GP with the prevailing conformal prediction (CP), a distribution-free post-processing framework that produces it prediction sets with a provably valid coverage under the sole assumption of data exchangeability. However, this assumption is usually violated in the online setting, where a prediction set is sought before revealing the true label. To ensure long-term coverage guarantee, we will adaptively set the key threshold parameter based on the feedback whether the true label falls inside the prediction set. Numerical results demonstrate the merits of the online GP-CP approach relative to existing alternatives in the long-term coverage performance.

When making treatment selection decisions, it is essential to include a causal effect estimation analysis to compare potential outcomes under different treatments or controls, assisting in optimal selection. However, merely estimating individual treatment effects may not suffice for truly optimal decisions. Our study addressed this issue by incorporating additional criteria, such as the estimations' uncertainty, measured by the conditional value-at-risk, commonly used in portfolio and insurance management. For continuous outcomes observable before and after treatment, we incorporated a specific prediction condition. We prioritized treatments that could yield optimal treatment effect results and lead to post-treatment outcomes more desirable than pretreatment levels, with the latter condition being called the prediction criterion. With these considerations, we propose a comprehensive methodology for multitreatment selection. Our approach ensures satisfaction of the overlap assumption, crucial for comparing outcomes for treated and control groups, by training propensity score models as a preliminary step before employing traditional causal models. To illustrate a practical application of our methodology, we applied it to the credit card limit adjustment problem. Analyzing a fintech company's historical data, we found that relying solely on counterfactual predictions was inadequate for appropriate credit line modifications. Incorporating our proposed additional criteria significantly enhanced policy performance.