The task of detecting moral values in text has significant implications in various fields, including natural language processing, social sciences, and ethical decision-making. Previously proposed supervised models often suffer from overfitting, leading to hyper-specialized moral classifiers that struggle to perform well on data from different domains. To address this issue, we introduce novel systems that leverage abstract concepts and common-sense knowledge acquired from Large Language Models and Natural Language Inference models during previous stages of training on multiple data sources. By doing so, we aim to develop versatile and robust methods for detecting moral values in real-world scenarios. Our approach uses the GPT 3.5 model as a zero-shot ready-made unsupervised multi-label classifier for moral values detection, eliminating the need for explicit training on labeled data. We compare it with a smaller NLI-based zero-shot model. The results show that the NLI approach achieves competitive results compared to the Davinci model. Furthermore, we conduct an in-depth investigation of the performance of supervised systems in the context of cross-domain multi-label moral value detection. This involves training supervised models on different domains to explore their effectiveness in handling data from different sources and comparing their performance with the unsupervised methods. Our contributions encompass a thorough analysis of both supervised and unsupervised methodologies for cross-domain value detection. We introduce the Davinci model as a state-of-the-art zero-shot unsupervised moral values classifier, pushing the boundaries of moral value detection without the need for explicit training on labeled data. Additionally, we perform a comparative evaluation of our approach with the supervised models, shedding light on their respective strengths and weaknesses.

A redescription is a pattern that characterises roughly the same entities in two different ways, and redescription mining is the task of automatically extracting redescriptions from the input dataset, given user-defined constraints. Redescription mining has found applications in various fields of science, such as ecometrics. Ecometrics aims to identify and model the functional relationships between traits of organisms and their environments [5, 7]. For instance, the teeth of large plant-eating mammals are adapted to the food that is available in their environment, which in turn depends on the climatic conditions, potentially allowing one to reason about the climate in the past based on the fossil record. To apply redescription mining in this context, the entities in the dataset represent localities, with two sets of attributes recording respectively the distribution of dental traits among species and the climatic conditions at each locality [11, 19].

Sparse autoencoders provide a promising unsupervised approach for extracting interpretable features from a language model by reconstructing activations from a sparse bottleneck layer. Since language models learn many concepts, autoencoders need to be very large to recover all relevant features. However, studying the properties of autoencoder scaling is difficult due to the need to balance reconstruction and sparsity objectives and the presence of dead latents. We propose using k-sparse autoencoders [Makhzani and Frey, 2013] to directly control sparsity, simplifying tuning and improving the reconstruction-sparsity frontier. Additionally, we find modifications that result in few dead latents, even at the largest scales we tried. Using these techniques, we find clean scaling laws with respect to autoencoder size and sparsity. We also introduce several new metrics for evaluating feature quality based on the recovery of hypothesized features, the explainability of activation patterns, and the sparsity of downstream effects. These metrics all generally improve with autoencoder size. To demonstrate the scalability of our approach, we train a 16 million latent autoencoder on GPT-4 activations for 40 billion tokens. We release training code and autoencoders for open-source models, as well as a visualizer.

Given a dataset of individuals each described by a covariate vector, a treatment, and an observed outcome on the treatment, the goal of the individual treatment effect (ITE) estimation task is to predict outcome changes resulting from a change in treatment. A fundamental challenge is that in the observational data, a covariate's outcome is observed only under one treatment, whereas we need to infer the difference in outcomes under two different treatments. Several existing approaches address this issue through training with inferred pseudo-outcomes, but their success relies on the quality of these pseudo-outcomes. We propose PairNet, a novel ITE estimation training strategy that minimizes losses over pairs of examples based on their factual observed outcomes. Theoretical analysis for binary treatments reveals that PairNet is a consistent estimator of ITE risk, and achieves smaller generalization error than baseline models. Empirical comparison with thirteen existing methods across eight benchmarks, covering both discrete and continuous treatments, shows that PairNet achieves significantly lower ITE error compared to the baselines. Also, it is model-agnostic and easy to implement.

Growing concerns about the AI alignment problem have emerged in recent years, with previous work focusing mainly on (1) qualitative descriptions of the alignment problem; (2) attempting to align AI actions with human interests by focusing on value specification and learning; and/or (3) focusing on a single agent or on humanity as a singular unit. Recent work in sociotechnical AI alignment has made some progress in defining alignment inclusively, but the field as a whole still lacks a systematic understanding of how to specify, describe, and analyze misalignment among entities, which may include individual humans, AI agents, and complex compositional entities such as corporations, nation-states, and so forth. Previous work on controversy in computational social science offers a mathematical model of contention among populations (of humans). In this paper, we adapt this contention model to the alignment problem, and show how misalignment can vary depending on the population of agents (human or otherwise) being observed, the domain in question, and the agents' probability-weighted preferences between possible outcomes. Our model departs from value specification approaches and focuses instead on the morass of complex, interlocking, sometimes contradictory goals that agents may have in practice. We apply our model by analyzing several case studies ranging from social media moderation to autonomous vehicle behavior. By applying our model with appropriately representative value data, AI engineers can ensure that their systems learn values maximally aligned with diverse human interests.

Event Extraction (EE) is an essential information extraction task that aims to extract event-related information from unstructured texts. The paradigm of this task has shifted from conventional classification-based methods to more contemporary question-answering-based (QA-based) approaches. However, in QA-based EE, the quality of the questions dramatically affects the extraction accuracy, and how to generate high-quality questions for QA-based EE remains a challenge. In this work, to tackle this challenge, we suggest four criteria to evaluate the quality of a question and propose a reinforcement learning method, RLQG, for QA-based EE that can generate generalizable, high-quality, and context-dependent questions and provides clear guidance to QA models. The extensive experiments conducted on ACE and RAMS datasets have strongly validated our approach's effectiveness, which also demonstrates its robustness in scenarios with limited training data. The corresponding code of RLQG is released for further research.

Graph neural network (GNN) models play a pivotal role in numerous tasks involving graph-related data analysis. Despite their efficacy, similar to other deep learning models, GNNs are susceptible to adversarial attacks. Even minor perturbations in graph data can induce substantial alterations in model predictions. While existing research has explored various adversarial defense techniques for GNNs, the challenge of defending against adversarial attacks on real-world scale graph data remains largely unresolved. On one hand, methods reliant on graph purification and preprocessing tend to excessively emphasize local graph information, leading to sub-optimal defensive outcomes. On the other hand, approaches rooted in graph structure learning entail significant time overheads, rendering them impractical for large-scale graphs. In this paper, we propose a new defense method named Talos, which enhances the global, rather than local, homophily of graphs as a defense. Experiments show that the proposed approach notably outperforms state-of-the-art defense approaches, while imposing little computational overhead.

Large language models (LLMs) are notorious for hallucinating, i.e., producing erroneous claims in their output. Such hallucinations can be dangerous, as occasional factual inaccuracies in the generated text might be obscured by the rest of the output being generally factually correct, making it extremely hard for the users to spot them. Current services that leverage LLMs usually do not provide any means for detecting unreliable generations. Here, we aim to bridge this gap. In particular, we propose a novel fact-checking and hallucination detection pipeline based on token-level uncertainty quantification. Uncertainty scores leverage information encapsulated in the output of a neural network or its layers to detect unreliable predictions, and we show that they can be used to fact-check the atomic claims in the LLM output. Moreover, we present a novel token-level uncertainty quantification method that removes the impact of uncertainty about what claim to generate on the current step and what surface form to use. Our method Claim Conditioned Probability (CCP) measures only the uncertainty of a particular claim value expressed by the model. Experiments on the task of biography generation demonstrate strong improvements for CCP compared to the baselines for seven LLMs and four languages. Human evaluation reveals that the fact-checking pipeline based on uncertainty quantification is competitive with a fact-checking tool that leverages external knowledge.

LLMs have marked a revolutonary shift, yet they falter when faced with compositional reasoning tasks. Our research embarks on a quest to uncover the root causes of compositional reasoning failures of LLMs, uncovering that most of them stem from the improperly generated or leveraged implicit reasoning results. Inspired by our empirical findings, we resort to Logit Lens and an intervention experiment to dissect the inner hidden states of LLMs. This deep dive reveals that implicit reasoning results indeed surface within middle layers and play a causative role in shaping the final explicit reasoning results. Our exploration further locates multi-head self-attention (MHSA) modules within these layers, which emerge as the linchpins in accurate generation and leveraing of implicit reasoning results. Grounded on the above findings, we develop CREME, a lightweight method to patch errors in compositional reasoning via editing the located MHSA modules. Our empirical evidence stands testament to CREME's effectiveness, paving the way for autonomously and continuously enhancing compositional reasoning capabilities in language models.

Large Language Model (LLM) services and models often come with legal rules on who can use them and how they must use them. Assessing the compliance of the released LLMs is crucial, as these rules protect the interests of the LLM contributor and prevent misuse. In this context, we describe the novel fingerprinting problem of Black-box Identity Verification (BBIV). The goal is to determine whether a third-party application uses a certain LLM through its chat function. We propose a method called Targeted Random Adversarial Prompt (TRAP) that identifies the specific LLM in use. We repurpose adversarial suffixes, originally proposed for jailbreaking, to get a pre-defined answer from the target LLM, while other models give random answers. TRAP detects the target LLMs with over 95% true positive rate at under 0.2% false positive rate even after a single interaction. TRAP remains effective even if the LLM has minor changes that do not significantly alter the original function.