Training Foundation Models as Data Compression: On Information, Model Weights and Copyright Law

The training process of foundation models as for other classes of deep learning systems is based on minimizing the reconstruction error over a training set. For this reason, they are susceptible to the memorization and subsequent reproduction of training samples. In this paper, we introduce a training-as-compressing perspective, wherein the model's weights embody a compressed representation of the training data. From a copyright standpoint, this point of view implies that the weights could be considered a reproduction or a derivative work of a potentially protected set of works. We investigate the technical and legal challenges that emerge from this framing of the copyright of outputs generated by foundation models, including their implications for practitioners and researchers. We demonstrate that adopting an information-centric approach to the problem presents a promising pathway for tackling these emerging complex legal issues.

Partial Information Decomposition for Data Interpretability and Feature Selection

In this paper, we introduce Partial Information Decomposition of Features (PIDF), a new paradigm for simultaneous data interpretability and feature selection. Contrary to traditional methods that assign a single importance value, our approach is based on three metrics per feature: the mutual information shared with the target variable, the feature's contribution to synergistic information, and the amount of this information that is redundant. In particular, we develop a novel procedure based on these three metrics, which reveals not only how features are correlated with the target but also the additional and overlapping information provided by considering them in combination with other features. We extensively evaluate PIDF using both synthetic and real-world data, demonstrating its potential applications and effectiveness, by considering case studies from genetics and neuroscience.

Large Language Models are Effective Priors for Causal Graph Discovery

Causal structure discovery from observations can be improved by integrating background knowledge provided by an expert to reduce the hypothesis space. Recently, Large Language Models (LLMs) have begun to be considered as sources of prior information given the low cost of querying them relative to a human expert. In this work, firstly, we propose a set of metrics for assessing LLM judgments for causal graph discovery independently of the downstream algorithm. Secondly, we systematically study a set of prompting designs that allows the model to specify priors about the structure of the causal graph. Finally, we present a general methodology for the integration of LLM priors in graph discovery algorithms, finding that they help improve performance on common-sense benchmarks and especially when used for assessing edge directionality. Our work highlights the potential as well as the shortcomings of the use of LLMs in this problem space.

Creative Beam Search: LLM-as-a-Judge For Improving Response Generation

Large language models are revolutionizing several areas, including artificial creativity. However, the process of generation in machines profoundly diverges from that observed in humans. In particular, machine generation is characterized by a lack of intentionality and an underlying creative process. We propose a method called Creative Beam Search that uses Diverse Beam Search and LLM-as-a-Judge to perform response generation and response validation. The results of a qualitative experiment show how our approach can provide better output than standard sampling techniques. We also show that the response validation step is a necessary complement to the response generation step.

Creative Beam Search

Large language models are revolutionizing several areas, including artificial creativity. However, the process of generation in machines profoundly diverges from that observed in humans. In particular, machine generation is characterized by a lack of intentionality and an underlying creative process. We propose a method called Creative Beam Search that uses Diverse Beam Search and LLM-as-a-Judge to perform response generation and response validation. The results of a qualitative experiment show how our approach can provide better output than standard sampling techniques. We also show that the response validation step is a necessary complement to the response generation step.

Graph Reinforcement Learning for Combinatorial Optimization: A Survey and Unifying Perspective

Graphs are a natural representation for systems based on relations between connected entities. Combinatorial optimization problems, which arise when considering an objective function related to a process of interest on discrete structures, are often challenging due to the rapid growth of the solution space. The trial-and-error paradigm of Reinforcement Learning has recently emerged as a promising alternative to traditional methods, such as exact algorithms and (meta)heuristics, for discovering better decision-making strategies in a variety of disciplines including chemistry, computer science, and statistics. Despite the fact that they arose in markedly different fields, these techniques share significant commonalities. Therefore, we set out to synthesize this work in a unifying perspective that we term Graph Reinforcement Learning, interpreting it as a constructive decision-making method for graph problems. After covering the relevant technical background, we review works along the dividing line of whether the goal is to optimize graph structure given a process of interest, or to optimize the outcome of the process itself under fixed graph structure. Finally, we discuss the common challenges facing the field and open research questions. In contrast with other surveys, the present work focuses on non-canonical graph problems for which performant algorithms are typically not known and Reinforcement Learning is able to provide efficient and effective solutions.

Do Agents Dream of Electric Sheep?: Improving Generalization in Reinforcement Learning through Generative Learning

The Overfitted Brain hypothesis suggests dreams happen to allow generalization in the human brain. Here, we ask if the same is true for reinforcement learning agents as well. Given limited experience in a real environment, we use imagination-based reinforcement learning to train a policy on dream-like episodes, where non-imaginative, predicted trajectories are modified through generative augmentations. Experiments on four ProcGen environments show that, compared to classic imagination and offline training on collected experience, our method can reach a higher level of generalization when dealing with sparsely rewarded environments.

Dynamics of Moral Behavior in Heterogeneous Populations of Learning Agents

Growing concerns about safety and alignment of AI systems highlight the importance of embedding moral capabilities in artificial agents. A promising solution is the use of learning from experience, i.e., Reinforcement Learning. In multi-agent (social) environments, complex population-level phenomena may emerge from interactions between individual learning agents. Many of the existing studies rely on simulated social dilemma environments to study the interactions of independent learning agents. However, they tend to ignore the moral heterogeneity that is likely to be present in societies of agents in practice. For example, at different points in time a single learning agent may face opponents who are consequentialist (i.e., caring about maximizing some outcome over time) or norm-based (i.e., focusing on conforming to a specific norm here and now). The extent to which agents' co-development may be impacted by such moral heterogeneity in populations is not well understood. In this paper, we present a study of the learning dynamics of morally heterogeneous populations interacting in a social dilemma setting. Using a Prisoner's Dilemma environment with a partner selection mechanism, we investigate the extent to which the prevalence of diverse moral agents in populations affects individual agents' learning behaviors and emergent population-level outcomes. We observe several types of non-trivial interactions between pro-social and anti-social agents, and find that certain classes of moral agents are able to steer selfish agents towards more cooperative behavior.

(Ir)rationality and Cognitive Biases in Large Language Models

Do large language models (LLMs) display rational reasoning? LLMs have been shown to contain human biases due to the data they have been trained on; whether this is reflected in rational reasoning remains less clear. In this paper, we answer this question by evaluating seven language models using tasks from the cognitive psychology literature. We find that, like humans, LLMs display irrationality in these tasks. However, the way this irrationality is displayed does not reflect that shown by humans. When incorrect answers are given by LLMs to these tasks, they are often incorrect in ways that differ from human-like biases. On top of this, the LLMs reveal an additional layer of irrationality in the significant inconsistency of the responses. Aside from the experimental results, this paper seeks to make a methodological contribution by showing how we can assess and compare different capabilities of these types of models, in this case with respect to rational reasoning.

Information-Theoretic State Variable Selection for Reinforcement Learning

Identifying the most suitable variables to represent the state is a fundamental challenge in Reinforcement Learning (RL). These variables must efficiently capture the information necessary for making optimal decisions. In order to address this problem, in this paper, we introduce the Transfer Entropy Redundancy Criterion (TERC), an information-theoretic criterion, which determines if there is \textit{entropy transferred} from state variables to actions during training. We define an algorithm based on TERC that provably excludes variables from the state that have no effect on the final performance of the agent, resulting in more sample efficient learning. Experimental results show that this speed-up is present across three different algorithm classes (represented by tabular Q-learning, Actor-Critic, and Proximal Policy Optimization (PPO)) in a variety of environments. Furthermore, to highlight the differences between the proposed methodology and the current state-of-the-art feature selection approaches, we present a series of controlled experiments on synthetic data, before generalizing to real-world decision-making tasks. We also introduce a representation of the problem that compactly captures the transfer of information from state variables to actions as Bayesian networks.