Large Language Models (LLM) technology is rapidly advancing towards human- like dialogue. Values are fundamental drivers of human behavior, yet research on the values expressed in LLM-generated text remains limited. While prior work has begun to explore value ranking in LLMs, the crucial aspect of value correlation – the interrelationship and consistency between different values – has been largely un-examined. Drawing on established psychological theories of human value structure, this paper investigates whether LLMs exhibit human-like value correlations within a single session, reflecting a coherent “persona”. Our findings reveal that standard prompting methods fail to produce human-consistent value correlations. However, we demonstrate that a novel prompting strategy (referred to as "Value Anchoring"), significantly improves the alignment of LLM value correlations with human data. Furthermore, we analyze the mechanism by which Value Anchoring achieves this effect. These results not only deepen our understanding of value representation in LLMs but also introduce new methodologies for evaluating consistency and human-likeness in LLM responses, highlighting the importance of explicit value prompting for generating human-aligned outputs.

Large Language Models (LLMs) often do not perform well on queries that require the aggregation of information across texts. To better evaluate this setting and facilitate modeling efforts, we introduce TACT - Text And Calculations through Tables, a dataset crafted to evaluate LLMs' reasoning and computational abilities using complex instructions. TACT contains challenging instructions that demand stitching information scattered across one or more texts, and performing complex integration on this information to generate the answer. We construct this dataset by leveraging an existing dataset of texts and their associated tables. For each such tables, we formulate new queries, and gather their respective answers. We demonstrate that all contemporary LLMs perform poorly on this dataset, achieving an accuracy below 38%. To pinpoint the difficulties and thoroughly dissect the problem, we analyze model performance across three components: table-generation, Pandas command-generation, and execution. Unexpectedly, we discover that each component presents substantial challenges for current LLMs. These insights lead us to propose a focused modeling framework, which we refer to as IE as a tool. Specifically, we propose to add "tools" for each of the above steps, and implement each such tool with few-shot prompting. This approach shows an improvement over existing prompting techniques, offering a promising direction for enhancing model capabilities in these tasks.

Paul Roit, Johan Ferret, Lior Shani, Roee Aharoni, Geoffrey Cideron, Robert Dadashi, Matthieu Geist, Sertan Girgin, Leonard Hussenot, Orgad Keller, Nikola Momchev, Sabela Ramos Garea, Piotr Stanczyk, Nino Vieillard, Olivier Bachem, Gal Elidan, Avinatan Hassidim, Olivier Pietquin, Idan Szpektor. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Supervised learning typically relies on manual annotation of the true labels. When there are many potential classes, searching for the best one can be prohibitive for a human annotator. On the other hand, comparing two candidate labels is often much easier. We focus on this type of pairwise supervision and ask how it can be used effectively in learning, and in particular in active learning. We obtain several insightful results in this context. In principle, finding the best of k labels can be done with k-1 active queries. We show that there is a natural class where this approach is sub-optimal, and that there is a more comparison-efficient active learning scheme. A key element in our analysis is the “label neighborhood graph” of the true distribution, which has an edge between two classes if they share a decision boundary. We also show that in the PAC setting, pairwise comparisons cannot provide improved sample complexity in the worst case. We complement our theoretical results with experiments, clearly demonstrating the effect of the neighborhood graph on sample complexity.

Image classification models can depend on multiple different semantic attributes of the image. An explanation of the decision of the classifier needs to both discover and visualize these properties. Here we present StylEx, a method for doing this, by training a generative model to specifically explain multiple attributes that underlie classifier decisions. A natural source for such attributes is the StyleSpace of StyleGAN, which is known to generate semantically meaningful dimensions in the image. However, because standard GAN training is not dependent on the classifier, it may not represent those attributes which are important for the classifier decision, and the dimensions of StyleSpace may represent irrelevant at-tributes. To overcome this, we propose a training procedure for a StyleGAN, which incorporates the classifier model, in order to learn a classifier-specific StyleSpace. Explanatory attributes are then selected from this space. These can be used to visualize the effect of changing multiple attributes per image, thus providing image-specific explanations. We apply StylEx to multiple domains, including animals, leaves, faces and retinal images. For these, we show how an image can be modified in different ways to change its classifier output. Our results show that the method finds attributes that align well with semantic ones, generate meaningful image-specific explanations, and are human-interpretable as measured in user-studies.1

Despite seminal advances in reinforcement learning in recent years, many domains where the rewards are sparse, e.g. given only at task completion, remain quite challenging. In such cases, it can be beneficial to tackle the task both from its beginning and end, and make the two ends meet. Existing approaches that do so, however, are not effective in the common scenario where the strategy needed near the end goal is very different from the one that is effective earlier on. In this work we propose a novel RL approach for such settings. In short, we first train a backward-looking agent with a simple relaxed goal, and then augment the state representation of the forward-looking agent with straightforward hint features. This allows the learned forward agent to leverage information from backward plans, without mimicking their policy. We demonstrate the efficacy of our approach on the challenging game of Sokoban, where we substantially surpass learned solvers that generalize across levels, and are competitive with SOTA performance of the best highly-crafted systems. Impressively, we achieve these results while learning from a small number of practice levels and using simple RL techniques.

We study conversational domain exploration (CODEX), where the user’s goal is to enrich her knowledge of a given domain by conversing with an informative bot. Such conversations should be well grounded in high-quality domain knowledge as well as engaging and open-ended. A CODEX bot should be proactive and introduce relevant information even if not directly asked for by the user. The bot should also appropriately pivot the conversation to undiscovered regions of the domain. To address these dialogue characteristics, we introduce a novel approach termed dynamic composition that decouples candidate content generation from the flexible composition of bot responses. This allows the bot to control the source, correctness and quality of the offered content, while achieving flexibility via a dialogue manager that selects the most appropriate contents in a compositional manner. We implemented a CODEX bot based on dynamic composition and integrated it into the Google Assistant . As an example domain, the bot conversed about the NBA basketball league in a seamless experience, such that users were not aware whether they were conversing with the vanilla system or the one augmented with our CODEX bot. Results are positive and offer insights into what makes for a good conversation. To the best of our knowledge, this is the first real user experiment of open-ended dialogues as part of a commercial assistant system.

Complex classifiers may exhibit “embarassing” failures in cases where humans can easily provide a justified classification. Avoiding such failures is obviously of key importance. In this work, we focus on one such setting, where a label is perfectly predictable if the input contains certain features, or rules, and otherwise it is predictable by a linear classifier. We define a hypothesis class that captures this notion and determine its sample complexity. We also give evidence that efficient algorithms cannot achieve this sample complexity. We then derive a simple and efficient algorithm and show that its sample complexity is close to optimal, among efficient algorithms. Experiments on synthetic and sentiment analysis data demonstrate the efficacy of the method, both in terms of accuracy and interpretability.

Heavy tailed and contaminated data are common in various applications of machine learning. A standard technique to handle regression tasks that involve such data, is to use robust losses, e.g., the popular Huber’s loss. In structured problems, however, where there are multiple labels and structural constraints on the labels are imposed (or learned), robust optimization is challenging, and more often than not the loss used is simply the negative log-likelihood of a Gaussian Markov random field. Heavy tailed and contaminated data are common in various applications of machine learning. A standard technique to handle regression tasks that involve such data, is to use robust losses, e.g., the popular Huber’s loss. In structured problems, however, where there are multiple labels and structural constraints on the labels are imposed (or learned), robust optimization is challenging, and more often than not the loss used is simply the negative log-likelihood of a Gaussian Markov random field. In this work, we analyze robust alternatives. Theoretical understanding of such problems is quite limited, with guarantees on optimization given only for special cases and non-structured settings. The core of the difficulty is the non-convexity of the objective function, implying that standard optimization algorithms may converge to sub-optimal critical points. Our analysis focuses on loss functions that arise from elliptical distributions, which appealingly include most loss functions proposed in the literature as special cases. We show that, even though these problems are non-convex, they can be optimized efficiently. Concretely, we prove that at the limit of infinite training data, due to algebraic properties of the problem, all stationary points are globally optimal. Finally, we demonstrate the empirical appeal of using these losses for regression on synthetic and real-life data.

In this paper, we consider the parameter estimation of K-Gaussians, given convex combinations of their realizations. In the remote sensing literature, this setting is known as the normal compositional model (NCM) and has shown promising gains in modeling hyperspectral images. Current NCM parameter estimation techniques are based on Bayesian methodology and are computationally slow and sensitive to their prior assumptions. Here, we introduce a deterministic variant of the NCM, named DNCM, which assumes that the unknown mixing coefficients are nonrandom. This leads to a standard Gaussian model with a simple estimation procedure, which we denote by K-Gaussians. Its iterations are provided in closed form and do not require any sampling schemes or simplifying structural assumptions. We illustrate the performance advantages of K-Gaussians using synthetic and real images, in terms of accuracy and computational costs in comparison to state of the art. We also demonstrate the use of our algorithm in hyperspectral target detection on a real image with known targets.

Modern retrieval systems are often driven by an underlying machine learning model. The goal of such systems is to identify and possibly rank the few most relevant items for a given query or context. Thus, such systems are typically evaluated using a ranking-based performance metric such as the area under the precision-recall curve, the F score, precision at fixed recall, etc. Obviously, it is desirable to train such systems to optimize the metric of interest. In practice, due to the scalability limitations of existing approaches for optimizing such objectives, large-scale retrieval systems are instead trained to maximize classification accuracy, in the hope that performance as measured via the true objective will also be favorable. In this work we present a unified framework that, using straightforward building block bounds, allows for highly scalable optimization of a wide range of ranking-based objectives. We demonstrate the advantage of our approach on several real-life retrieval problems that are significantly larger than those considered in the literature, while achieving substantial improvement in performance over the accuracy-objective baseline.

User modeling in advertising and recommendation has typically focused on myopic predictors of user responses. In this work, we consider the long-term decision problem associated with user interaction. We propose a concise specification of long-term interaction dynamics by combining factored dynamic Bayesian networks with logistic predictors of user responses, allowing state-of-the-art prediction models to be seamlessly extended. We show how to solve such models at scale by providing a constraint generation approach for approximate linear programming that overcomes the variable coupling and non-linearity induced by the logistic regression predictor. The efficacy of the approach is demonstrated on advertising domains with up to 2^54 states and 2^39 actions.

Neural networks have recently re-emerged as a powerful hypothesis class, yielding impressive classification accuracy in multiple domains. However, their training is a non convex optimization problem. Here we address this difficulty by turning to "improper learning" of neural nets. In other words, we learn a classifier that is not a neural net but is competitive with the best neural net model given a sufficient number of training examples. Our approach relies on a novel kernel which integrates over the set of possible neural models. It turns out that the corresponding integral can be evaluated in closed form via a simple recursion. The learning problem is then an SVM with this kernel, and a global optimum can thus be found efficiently. We also provide sample complexity results which depend on the stability of the optimal neural net.

A formidable challenge in uncertainty modeling in general, and when learning Bayesian networks in particular, is the discovery of unknown hidden variables. Few works that tackle this task are typically limited to discrete or Gaussian domains, or to tree structures. We propose a novel general purpose approach for discovering hidden variables in flexible non-Gaussian domains using the powerful class of Gaussian copula networks. Briefly, we define the concept of a hypothetically optimal predictor of variable and show it can be used to discover useful hidden variables in the expressive framework of copula networks. Our approach leads to performance and compactness advantages over competitors in a variety of domains.

In this paper we address target detection in correlated non-Gaussian noise. We introduce a powerful class of multivariate complex valued distribution that allows us to specify flexible non-Gaussian marginals, as well as correlation between the variables, while preserving circular symmetry. For noise belonging to this class, we study the fundamental problem of signal detection under different settings, and develop the needed (generalized) likelihood ratio tests. We also consider the problem of estimation of the noise parameters, and derive the maximum likelihood formulations. We compare the performance of the proposed methods using numerical simulations on synthetic data, and demonstrate the importance of using both correlations and non-Gaussiantiy.

Probabilistic modeling of temporal phenomena is of central importance in a variety of fields ranging from neuroscience to economics to speech recognition. While the task has received extensive attention in recent decades, learning temporal models for multivariate real-valued data that is non-Gaussian is still a formidable challenge. Recently, the power of copulas, a framework for representing complex multi-modal and heavy-tailed distributions, was fused with the formalism of Bayesian networks to allow for flexible modeling of high-dimensional distributions. In this work we introduce Dynamic Copula Bayesian Networks, a generalization aimed at capturing the distribution of rich temporal sequences. We apply our model to three markedly different real-life domains and demonstrate substantial quantitative and qualitative advantage.

The empirical success of the belief propagation approximate inference algorithm has inspired numerous theoretical and algorithmic advances. Yet, for continuous non-Gaussian domains performing belief propagation remains a challenging task: recent innovations such as nonparametric or kernel belief propagation, while useful, come with a substantial computational cost and offer little theoretical guarantees, even for tree structured models. In this work we present Nonparanormal BP for performing efficient inference on distributions parameterized by a Gaussian copulas network and any univariate marginals. For tree structured networks, our approach is guaranteed to be exact for this powerful class of non-Gaussian models. Importantly, the method is as efficient as standard Gaussian BP, and its convergence properties do not depend on the complexity of the univariate marginals, even when a nonparametric representation is used.

We present the Copula Bayesian Network model for representing multivariate continuous distributions. Our approach builds on a novel copula-based parameterization of a conditional density that, joined with a graph that encodes independencies, offers great flexibility in modeling high-dimensional densities, while maintaining control over the form of the univariate marginals. We demonstrate the advantage of our framework for generalization over standard Bayesian networks as well as tree structured copula models for varied real-life domains that are of substantially higher dimension than those typically considered in the copula literature.

With the increased availability of data for complex domains, it is desirable to learn Bayesian network structures that are sufficiently expressive for generalization while also allowing for tractable inference. While the method of thin junction trees can, in principle, be used for this purpose, its fully greedy nature makes it prone to overfitting, particularly when data is scarce. In this work we present a novel method for learning Bayesian networks of bounded treewidth that employs global structure modifications and that is polynomial in the size of the graph and the treewidth bound. At the heart of our method is a triangulated graph that we dynamically update in a way that facilitates the addition of chain structures that increase the bound on the model's treewidth by at most one. We demonstrate the effectiveness of our ``treewidth-friendly'' method on several real-life datasets. Importantly, we also show that by using global operators, we are able to achieve better generalization even when learning Bayesian networks of unbounded treewidth.

Discriminative tasks, including object categorization and detection, are central components of high-level computer vision. Sometimes, however, we are interested in more refined aspects of the object in an image, such as pose or particular regions. In this paper we develop a method (LOOPS) for learning a shape and image feature model that can be trained on a particular object class, and used to outline instances of the class in novel images. Furthermore, while the training data consists of uncorresponded outlines, the resulting LOOPS model contains a set of landmark points that appear consistently across instances, and can be accurately localized in an image. Our model achieves state-of-the-art results in precisely outlining objects that exhibit large deformations and articulations in cluttered natural images. These localizations can then be used to address a range of tasks, including descriptive classification, search, and clustering.

We consider the important challenge of recognizing a variety of deformable object classes in images. Of fundamental importance and particular difficulty in this setting is the problem of "outlining" an object, rather than simply deciding on its presence or absence. A major obstacle in learning a model that will allow us to address this task is the need for hand-segmented training images. In this paper we present a novel landmark-based, piecewise-linear model of the shape of an object class. We then formulate a learning approach that allows us to learn this model with minimal user supervision. We circumvent the need for hand-segmentation by transferring the shape "essence" of an object from drawings to complex images. We show that our method is able to automatically and effectively learn and localize a variety of object classes.

We consider the problem of learning classifiers for structurally incomplete data, where some ob jects have a subset of features inherently absent due to complex relationships between the features. The common approach for handling missing features is to begin with a preprocessing phase that completes the missing features, and then use a standard classification procedure. In this paper we show how incomplete data can be classified directly without any completion of the missing features using a max-margin learning framework. We formulate this task using a geometrically-inspired ob jective function, and discuss two optimization approaches: The linearly separable case is written as a set of convex feasibility problems, and the non-separable case has a non-convex ob jective that we optimize iteratively. By avoiding the pre-processing phase in which the data is completed, these approaches offer considerable computational savings. More importantly, we show that by elegantly handling complex patterns of missing values, our approach is both competitive with other methods when the values are missing at random and outperforms them when the missing values have non-trivial structure. We demonstrate our results on two real-world problems: edge prediction in metabolic pathways, and automobile detection in natural images.

In general, the problem of computing a maximum a posteriori (MAP) assignment in a Markov random field (MRF) is computationally intractable. However, in certain subclasses of MRF, an optimal or close-to-optimal assignment can be found very efficiently using combinatorial optimization algorithms: certain MRFs with mutual exclusion constraints can be solved using bipartite matching, and MRFs with regular potentials can be solved using minimum cut methods. However, these solutions do not apply to the many MRFs that contain such tractable components as sub-networks, but also other non-complying potentials. In this paper, we present a new method, called C O M P O S E, for exploiting combinatorial optimization for sub-networks within the context of a max-product belief propagation algorithm. C O M P O S E uses combinatorial optimization for computing exact maxmarginals for an entire sub-network; these can then be used for inference in the context of the network as a whole. We describe highly efficient methods for computing max-marginals for subnetworks corresponding both to bipartite matchings and to regular networks. We present results on both synthetic and real networks encoding correspondence problems between images, which involve both matching constraints and pairwise geometric constraints. We compare to a range of current methods, showing that the ability of C O M P O S E to transmit information globally across the network leads to improved convergence, decreased running time, and higher-scoring assignments.

MOTIVATION: A key aspect of transcriptional regulation is the binding of transcription factors to sequence-specific binding sites that allow them to modulate the expression of nearby genes. Given models of such binding sites, one can scan regulatory regions for putative binding sites and construct a genome-wide regulatory network. In such genome-wide scans, it is crucial to control the amount of false positive predictions. Recently, several works demonstrated the benefits of modeling dependencies between positions within the binding site. Yet, computing the statistical significance of putative binding sites in this scenario remains a challenge. RESULTS: We present a general, accurate and efficient method for computing p-values of putative binding sites that is applicable to a large class of probabilistic binding site and background models. We demonstrate the accuracy of the method on synthetic and real-life data. AVAILABILITY: The procedure for scanning DNA sequences and computing the statistical significance of putative binding site scores is available upon request at http://compbio.cs.huji.ac.il/CIS/ CONTACT: [email protected].

A central challenge in learning probabilistic graphical models is dealing with domains that involve hidden variables. The common approach for learning model parameters in such domains is the expectation maximization (EM) algorithm. This algorithm, however, can easily get trapped in sub-optimal local maxima. Learning the model structure is even more challenging. The structural EM algorithm can adapt the structure in the presence of hidden variables, but usually performs poorly without prior knowledge about the cardinality and location of the hidden variables. In this work, we present a general approach for learning Bayesian networks with hidden variables that overcomes these problems. The approach builds on the information bottleneck framework of Tishby et al. (1999). We start by proving formal correspondence between the information bottleneck objective and the standard parametric EM functional. We then use this correspondence to construct a learning algorithm that combines an information-theoretic smoothing term with a continuation procedure. Intuitively, the algorithm bypasses local maxima and achieves superior solutions by following a continuous path from a solution of, an easy and smooth, target function, to a solution of the desired likelihood function. As we show, our algorithmic framework allows learning of the parameters as well as the structure of a network. In addition, it also allows us to introduce new hidden variables during model selection and learn their cardinality. We demonstrate the performance of our procedure on several challenging real-life data sets.

The availability of whole genome sequences and high-throughput genomic assays opens the door for in silico analysis of transcription regulation. This includes methods for discovering and characterizing the binding sites of DNA-binding proteins, such as transcription factors. A common representation of transcription factor binding sites is a position specific score matrix (PSSM). This representation makes the strong assumption that binding site positions are independent of each other. In this work, we explore Bayesian network representations of binding sites that provide different tradeoffs between complexity (number of parameters) and the richness of dependencies between positions. We develop the formal machinery for learning such models from data and for estimating the statistical significance of putative binding sites. We then evaluate the ramifications of these richer representations in characterizing binding site motifs and predicting their genomic locations. We show that these richer representations improve over the PSSM model in both tasks.

A serious problem in learning probabilistic models is the presence of hid(cid:173) den variables. These variables are not observed, yet interact with several of the observed variables. As such, they induce seemingly complex de(cid:173) pendencies among the latter. In recent years, much attention has been devoted to the development of algorithms for learning parameters, and in some cases structure, in the presence of hidden variables. In this pa(cid:173) per, we address the related problem of detecting hidden variables that interact with the observed variables. This problem is of interest both for improving our understanding of the domain and as a preliminary step that guides the learning procedure towards promising models. A very natural approach is to search for "structural signatures" of hidden variables - substructures in the learned network that tend to suggest the presence of a hidden variable. We make this basic idea concrete, and show how to integrate it with structure-search algorithms. We evaluate this method on several synthetic and real-life datasets, and show that it performs surpris(cid:173) ingly well.