Fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback) is a crucial step in training language models to robustly carry out tasks of interest. However, we lack a systematic understanding of the effects of fine-tuning, particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of suppressing model capabilities on other tasks. This degradation is especially pronounced for tasks "closest" to the fine-tuning distribution. We hypothesize that language models implicitly infer the task of the prompt corresponds, and the fine-tuning process predominantly skews this task inference towards tasks in the fine-tuning distribution. To test this hypothesis, we propose Conjugate Prompting to see if we can recover pretrained capabilities. Conjugate prompting artificially makes the task look farther from the fine-tuning distribution while requiring the same capability. We find that conjugate prompting systematically recovers some of the pretraining capabilities on our synthetic setup. We then apply conjugate prompting to real-world LLMs using the observation that fine-tuning distributions are typically heavily skewed towards English. We find that simply translating the prompts to different languages can cause the fine-tuned models to respond like their pretrained counterparts instead. This allows us to recover the in-context learning abilities lost via instruction tuning, and more concerningly, to recover harmful content generation suppressed by safety fine-tuning in chatbots like ChatGPT. The development of large language models (LLMs) typically involves two stages--pretraining (next token prediction) on vast text corpora and fine-tuning on carefully curated datasets to adapt the pretrained model to the application of interest. One fundamental concern is that fine-tuning datasets are considerably smaller and less diverse than web-scale pretraining datasets (Raffel et al., 2020; Arivazhagan et al., 2019; Gao et al., 2021), and there is always a risk that the fine-tuned model "catastrophically forgets" (McCloskey & Cohen, 1989) how to solve problems that the pretrained model could solve. Such a gap has been reported as an "alignment tax" in works such as Ouyang et al. (2022) and Bai et al. (2022), but there is no clear understanding of what these trade-offs are and how to mitigate them. Given the importance of the fine-tuning process, it is imperative to build a systematic understanding of the effects.

Abstract-- We introduce a metric for evaluating the robustness of a classifier, with particular attention to adversarial perturbations, in terms of expected functionality with respect to possible adversarial perturbations. Defining robustness in terms of an expected value is motivated by a domain general approach to robustness quantification. In support of this, it is desirable to have an approach Adversarial ML arises when some aspect of the system is to quantifying robustness that applies generally across intentionally manipulated to cause the classifier to make errors. Adversarial robustness specifically seeks to measure a model's We adopt the perspective that robustness is the ability of an performance when these perturbations are chosen selectively ML system to maintain its functionality at an acceptable level to be maximally disruptive. For example, evasion attacks add of performance when some aspect of the system is subject to human-imperceptible perturbations to a data instance to alter perturbation, consistent with the domain-general definition laid the output of a classifier, as illustrated in Figure 1.

Much research effort is being applied to the task of compressing the knowledge of self-supervised models, which are powerful, yet large and memory consuming. In this work, we show that the original method of knowledge distillation (and its more recently proposed extension, decoupled knowledge distillation) can be applied to the task of distilling HuBERT. In contrast to methods that focus on distilling internal features, this allows for more freedom in the network architecture of the compressed model. We thus propose to distill HuBERT's Transformer layers into an LSTM-based distilled model that reduces the number of parameters even below DistilHuBERT and at the same time shows improved performance in automatic speech recognition.

The growing body of political texts opens up new opportunities for rich insights into political dynamics and ideologies but also increases the workload for manual analysis. Automated speaker attribution, which detects who said what to whom in a speech event and is closely related to semantic role labeling, is an important processing step for computational text analysis. We study the potential of the large language model family Llama 2 to automate speaker attribution in German parliamentary debates from 2017-2021. We fine-tune Llama 2 with QLoRA, an efficient training strategy, and observe our approach to achieve competitive performance in the GermEval 2023 Shared Task On Speaker Attribution in German News Articles and Parliamentary Debates. Our results shed light on the capabilities of large language models in automating speaker attribution, revealing a promising avenue for computational analysis of political discourse and the development of semantic role labeling systems.

Causality and eXplainable Artificial Intelligence (XAI) have developed as separate fields in computer science, even though the underlying concepts of causation and explanation share common ancient roots. This is further enforced by the lack of review works jointly covering these two fields. In this paper, we investigate the literature to try to understand how and to what extent causality and XAI are intertwined. More precisely, we seek to uncover what kinds of relationships exist between the two concepts and how one can benefit from them, for instance, in building trust in AI systems. As a result, three main perspectives are identified. In the first one, the lack of causality is seen as one of the major limitations of current AI and XAI approaches, and the "optimal" form of explanations is investigated. The second is a pragmatic perspective and considers XAI as a tool to foster scientific exploration for causal inquiry, via the identification of pursue-worthy experimental manipulations. Finally, the third perspective supports the idea that causality is propaedeutic to XAI in three possible manners: exploiting concepts borrowed from causality to support or improve XAI, utilizing counterfactuals for explainability, and considering accessing a causal model as explaining itself. To complement our analysis, we also provide relevant software solutions used to automate causal tasks. We believe our work provides a unified view of the two fields of causality and XAI by highlighting potential domain bridges and uncovering possible limitations.

Large language models (LLMs) have the potential to transform our lives and work through the content they generate, known as AI-Generated Content (AIGC). To harness this transformation, we need to understand the limitations of LLMs. Here, we investigate the bias of AIGC produced by seven representative LLMs, including ChatGPT and LLaMA. We collect news articles from The New York Times and Reuters, both known for their dedication to provide unbiased news. We then apply each examined LLM to generate news content with headlines of these news articles as prompts, and evaluate the gender and racial biases of the AIGC produced by the LLM by comparing the AIGC and the original news articles. We further analyze the gender bias of each LLM under biased prompts by adding gender-biased messages to prompts constructed from these news headlines. Our study reveals that the AIGC produced by each examined LLM demonstrates substantial gender and racial biases. Moreover, the AIGC generated by each LLM exhibits notable discrimination against females and individuals of the Black race. Among the LLMs, the AIGC generated by ChatGPT demonstrates the lowest level of bias, and ChatGPT is the sole model capable of declining content generation when provided with biased prompts.

Sentiments inherently drive politics. How we receive and process information plays an essential role in political decision-making, shaping our judgment with strategic consequences both on the level of legislators and the masses. If sentiment plays such an important role in politics, how can we study and measure it systematically? The paper presents a new dataset of sentiment-annotated sentences, which are used in a series of experiments focused on training a robust sentiment classifier for parliamentary proceedings. The paper also introduces the first domain-specific LLM for political science applications additionally pre-trained on 1.72 billion domain-specific words from proceedings of 27 European parliaments. We present experiments demonstrating how the additional pre-training of LLM on parliamentary data can significantly improve the model downstream performance on the domain-specific tasks, in our case, sentiment detection in parliamentary proceedings. We further show that multilingual models perform very well on unseen languages and that additional data from other languages significantly improves the target parliament's results. The paper makes an important contribution to multiple domains of social sciences and bridges them with computer science and computational linguistics. Lastly, it sets up a more robust approach to sentiment analysis of political texts in general, which allows scholars to study political sentiment from a comparative perspective using standardized tools and techniques.

Identification of cracks is essential to assess the structural integrity of concrete infrastructure. However, robust crack segmentation remains a challenging task for computer vision systems due to the diverse appearance of concrete surfaces, variable lighting and weather conditions, and the overlapping of different defects. In particular recent data-driven methods struggle with the limited availability of data, the fine-grained and time-consuming nature of crack annotation, and face subsequent difficulty in generalizing to out-of-distribution samples. In this work, we move past these challenges in a two-fold way. We introduce a high-fidelity crack graphics simulator based on fractals and a corresponding fully-annotated crack dataset. We then complement the latter with a system that learns generalizable representations from simulation, by leveraging both a pointwise mutual information estimate along with adaptive instance normalization as inductive biases. Finally, we empirically highlight how different design choices are symbiotic in bridging the simulation to real gap, and ultimately demonstrate that our introduced system can effectively handle real-world crack segmentation.

Understanding the framing of political issues is of paramount importance as it significantly shapes how individuals perceive, interpret, and engage with these matters. While prior research has independently explored framing within news media and by social media users, there remains a notable gap in our comprehension of the disparities in framing political issues between these two distinct groups. To address this gap, we conduct a comprehensive investigation, focusing on the nuanced distinctions both qualitatively and quantitatively in the framing of social media and traditional media outlets concerning a series of American Supreme Court rulings on affirmative action, student loans, and abortion rights. Our findings reveal that, while some overlap in framing exists between social media and traditional media outlets, substantial differences emerge both across various topics and within specific framing categories. Compared to traditional news media, social media platforms tend to present more polarized stances across all framing categories. Further, we observe significant polarization in the news media's treatment (i.e., Left vs. Right leaning media) of affirmative action and abortion rights, whereas the topic of student loans tends to exhibit a greater degree of consensus. The disparities in framing between traditional and social media platforms carry significant implications for the formation of public opinion, policy decision-making, and the broader political landscape.

Transformers have significantly advanced the field of natural language processing, but comprehending their internal mechanisms remains a challenge. In this paper, we introduce a novel geometric perspective that elucidates the inner mechanisms of transformer operations. Our primary contribution is illustrating how layer normalization confines the latent features to a hyper-sphere, subsequently enabling attention to mold the semantic representation of words on this surface. This geometric viewpoint seamlessly connects established properties such as iterative refinement and contextual embeddings. We validate our insights by probing a pre-trained 124M parameter GPT-2 model. Our findings reveal clear query-key attention patterns in early layers and build upon prior observations regarding the subject-specific nature of attention heads at deeper layers. Harnessing these geometric insights, we present an intuitive understanding of transformers, depicting them as processes that model the trajectory of word particles along the hyper-sphere.