Eliciting harmful behavior from large language models (LLMs) is an important task to ensure the proper alignment and safety of the models. Often when training LLMs, ethical guidelines are followed yet alignment failures may still be uncovered through red teaming adversarial attacks. This work frames the red-teaming problem as a Markov decision process (MDP) and uses Monte Carlo tree search to find harmful behaviors of black-box, closed-source LLMs. We optimize token-level prompt suffixes towards targeted harmful behaviors on white-box LLMs and include a naturalistic loss term, log-perplexity, to generate more natural language attacks for better interpretability. The proposed algorithm, Kov, trains on white-box LLMs to optimize the adversarial attacks and periodically evaluates responses from the black-box LLM to guide the search towards more harmful black-box behaviors. In our preliminary study, results indicate that we can jailbreak black-box models, such as GPT-3.5, in only 10 queries, yet fail on GPT-4$-$which may indicate that newer models are more robust to token-level attacks. All work to reproduce these results is open sourced (https://github.com/sisl/Kov.jl).

Text-to-image (TTI) systems, particularly those utilizing open-source frameworks, have become increasingly prevalent in the production of Artificial Intelligence (AI)-generated visuals. While existing literature has explored various problematic aspects of TTI technologies, such as bias in generated content, intellectual property concerns, and the reinforcement of harmful stereotypes, open-source TTI frameworks have not yet been systematically examined from a cultural perspective. This study addresses this gap by analyzing the CivitAI platform, a leading open-source platform dedicated to TTI AI. We introduce the Civiverse prompt dataset, encompassing millions of images and related metadata. We focus on prompt analysis, specifically examining the semantic characteristics of text prompts, as it is crucial for addressing societal issues related to generative technologies. This analysis provides insights into user intentions, preferences, and behaviors, which in turn shape the outputs of these models. Our findings reveal a predominant preference for generating explicit content, along with a focus on homogenization of semantic content. These insights underscore the need for further research into the perpetuation of misogyny, harmful stereotypes, and the uniformity of visual culture within these models.

Large Language Models (LLMs) exhibit potentially harmful biases that reinforce culturally inherent stereotypes, cloud moral judgments, or amplify positive evaluations of majority groups. Previous explanations mainly attributed bias in LLMs to human annotators and the selection of training data. Consequently, they have typically been addressed with bottom-up approaches such as reinforcement learning or debiasing corpora. However, these methods only treat the effects of LLM biases by indirectly influencing the model architecture, but do not address the underlying causes in the computational process. Here, we propose metacognitive myopia as a cognitive-ecological framework that can account for a conglomerate of established and emerging LLM biases and provide a lever to address problems in powerful but vulnerable tools. Our theoretical framework posits that a lack of the two components of metacognition, monitoring and control, causes five symptoms of metacognitive myopia in LLMs: integration of invalid tokens and embeddings, susceptibility to redundant information, neglect of base rates in conditional computation, decision rules based on frequency, and inappropriate higher-order statistical inference for nested data structures. As a result, LLMs produce erroneous output that reaches into the daily high-stakes decisions of humans. By introducing metacognitive regulatory processes into LLMs, engineers and scientists can develop precise remedies for the underlying causes of these biases. Our theory sheds new light on flawed human-machine interactions and raises ethical concerns regarding the increasing, imprudent implementation of LLMs in organizational structures.

Episodic structures are inherently interpretable and adaptable to evolving large-scale key events. However, state-of-the-art automatic event detection methods overlook event episodes and, therefore, struggle with these crucial characteristics. This paper introduces a novel task, episode detection, aimed at identifying episodes from a news corpus containing key event articles. An episode describes a cohesive cluster of core entities (e.g., "protesters", "police") performing actions at a specific time and location. Furthermore, an episode is a significant part of a larger group of episodes under a particular key event. Automatically detecting episodes is challenging because, unlike key events and atomic actions, we cannot rely on explicit mentions of times and locations to distinguish between episodes or use semantic similarity to merge inconsistent episode co-references. To address these challenges, we introduce EpiMine, an unsupervised episode detection framework that (1) automatically identifies the most salient, key-event-relevant terms and segments, (2) determines candidate episodes in an article based on natural episodic partitions estimated through shifts in discriminative term combinations, and (3) refines and forms final episode clusters using large language model-based reasoning on the candidate episodes. We construct three diverse, real-world event datasets annotated at the episode level. EpiMine outperforms all baselines on these datasets by an average 59.2% increase across all metrics.

Directly learning from examples of random difficulty levels is often challenging for both humans and machine learning models. A more effective strategy involves exposing learners to examples in a progressive order, from easy to difficult. Curriculum Learning (CL) has been proposed to implement this strategy in machine learning model training. However, two key challenges persist in CL framework design: defining the difficulty of training data and determining the appropriate amount of data to input at each training step. This paper presents a Psychology-based Unified Dynamic Framework for Curriculum Learning (PUDF), drawing inspiration from psychometrics. We quantify the difficulty of training data by applying Item Response Theory (IRT) to responses from Artificial Crowds (AC). This theory-driven IRT-AC approach leads to global (i.e., model-independent) and interpretable difficulty values. Leveraging IRT, we propose a Dynamic Data Selection via Model Ability Estimation (DDS-MAE) strategy to schedule the appropriate amount of data during model training. Since our difficulty labeling and model ability estimation are based on a consistent theory, namely IRT, their values are comparable within the same scope, potentially leading to a faster convergence compared to the other CL methods. Experimental results demonstrate that fine-tuning pre-trained language models with PUDF enhances their performance on the GLUE benchmark. Moreover, PUDF surpasses other state-of-the-art (SOTA) CL methods on the GLUE benchmark. We further explore the components of PUDF, namely the difficulty measurer (IRT-AC) and the training scheduler (DDS-MAE) qualitatively and quantitatively. Lastly, we conduct an ablation study to clarify which components of PUDF contribute to faster convergence and higher accuracy.

The task of deciding whether two documents are written by the same author is challenging for both machines and humans. This task is even more challenging when the two documents are written about different topics (e.g. baseball vs. politics) or in different genres (e.g. a blog post vs. an academic article). For machines, the problem is complicated by the relative lack of real-world training examples that cross the topic boundary and the vanishing scarcity of cross-genre data. We propose targeted methods for training data selection and a novel learning curriculum that are designed to discourage a model's reliance on topic information for authorship attribution and correspondingly force it to incorporate information more robustly indicative of style no matter the topic. These refinements yield a 62.7% relative improvement in average cross-genre authorship attribution, as well as 16.6% in the per-genre condition.

Guitar tablatures enrich the structure of traditional music notation by assigning each note to a string and fret of a guitar in a particular tuning, indicating precisely where to play the note on the instrument. The problem of generating tablature from a symbolic music representation involves inferring this string and fret assignment per note across an entire composition or performance. On the guitar, multiple string-fret assignments are possible for most pitches, which leads to a large combinatorial space that prevents exhaustive search approaches. Most modern methods use constraint-based dynamic programming to minimize some cost function (e.g.\ hand position movement). In this work, we introduce a novel deep learning solution to symbolic guitar tablature estimation. We train an encoder-decoder Transformer model in a masked language modeling paradigm to assign notes to strings. The model is first pre-trained on DadaGP, a dataset of over 25K tablatures, and then fine-tuned on a curated set of professionally transcribed guitar performances. Given the subjective nature of assessing tablature quality, we conduct a user study amongst guitarists, wherein we ask participants to rate the playability of multiple versions of tablature for the same four-bar excerpt. The results indicate our system significantly outperforms competing algorithms.

Technology does not exist in a vacuum; technological development, media representation, public perception, and governmental regulation cyclically influence each other to produce the collective understanding of a technology's capabilities, utilities, and risks. When these capabilities are overestimated, there is an enhanced risk of subjecting the public to dangerous or harmful technology, artificially restricting research and development directions, and enabling misguided or detrimental policy. The dangers of technological hype are particularly relevant in the rapidly evolving space of AI. Centering the research community as a key player in the development and proliferation of hype, we examine the origins and risks of AI hype to the research community and society more broadly and propose a set of measures that researchers, regulators, and the public can take to mitigate these risks and reduce the prevalence of unfounded claims about the technology.

Retrieval-augmented generation (RAG) methods encounter difficulties when addressing complex questions like multi-hop queries. While iterative retrieval methods improve performance by gathering additional information, current approaches often rely on multiple calls of large language models (LLMs). In this paper, we introduce EfficientRAG, an efficient retriever for multi-hop question answering. EfficientRAG iteratively generates new queries without the need for LLM calls at each iteration and filters out irrelevant information. Experimental results demonstrate that EfficientRAG surpasses existing RAG methods on three open-domain multi-hop question-answering datasets.

This paper focuses on detecting propagandistic spans and persuasion techniques in Arabic text from tweets and news paragraphs. Each entry in the dataset contains a text sample and corresponding labels that indicate the start and end positions of propaganda techniques within the text. Tokens falling within a labeled span were assigned "B" (Begin) or "I" (Inside), "O", corresponding to the specific propaganda technique. Using attention masks, we created uniform lengths for each span and assigned BIO tags to each token based on the provided labels. Then, we used AraBERT-base pre-trained model for Arabic text tokenization and embeddings with a token classification layer to identify propaganda techniques. Our training process involves a two-phase fine-tuning approach. First, we train only the classification layer for a few epochs, followed by full model fine-tuning, updating all parameters. This methodology allows the model to adapt to the specific characteristics of the propaganda detection task while leveraging the knowledge captured by the pre-trained AraBERT model. Our approach achieved an F1 score of 0.2774, securing the 3rd position in the leaderboard of Task 1.