Large language models (large LMs) are susceptible to producing text that contains hallucinated content. An important instance of this problem is self-contradiction, where the LM generates two contradictory sentences within the same context. In this work, we present a comprehensive investigation into self-contradiction for various instruction-tuned LMs, covering evaluation, detection, and mitigation. Our analysis reveals the prevalence of self-contradictions when LMs generate text for open-domain topics, e.g., in 17.7% of all sentences produced by ChatGPT. Self-contradiction also complements retrieval-based methods, as a large portion of them (e.g., 35.8% for ChatGPT) cannot be verified using Wikipedia. We then propose a novel prompting-based framework designed to effectively detect and mitigate self-contradictions. Our detector achieves high accuracy, e.g., around 80% F1 score when prompting ChatGPT. The mitigation algorithm iteratively refines the generated text to remove contradictory information while preserving text fluency and informativeness. Importantly, our entire framework is applicable to black-box LMs and does not require external grounded knowledge. Our approach is practically effective and has been released as a push-button tool to benefit the public, available at https://chatprotect.ai/.

Language models (LMs) are becoming the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models. First, we taxonomize the vast space of potential scenarios (i.e. use cases) and metrics (i.e. desiderata) that are of interest for LMs. Then we select a broad subset based on coverage and feasibility, noting what's missing or underrepresented (e.g. question answering for neglected English dialects, metrics for trustworthiness). Second, we adopt a multi-metric approach: We measure 7 metrics (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency) for each of 16 core scenarios when possible (87.5% of the time). This ensures metrics beyond accuracy don't fall to the wayside, and that trade-offs are clearly exposed. We also perform 7 targeted evaluations, based on 26 targeted scenarios, to analyze specific aspects (e.g. reasoning, disinformation). Third, we conduct a large-scale evaluation of 30 prominent language models (spanning open, limited-access, and closed models) on all 42 scenarios, 21 of which were not previously used in mainstream LM evaluation. Prior to HELM, models on average were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common. We improve this to 96.0%: now all 30 models have been densely benchmarked on the same core scenarios and metrics under standardized conditions. Our evaluation surfaces 25 top-level findings. For full transparency, we release all raw model prompts and completions publicly for further analysis, as well as a general modular toolkit. We intend for HELM to be a living benchmark for the community, continuously updated with new scenarios, metrics, and models.

In-Context Learning (ICL) is a new paradigm for natural language processing (NLP), where a large language model (LLM) observes a small number of demonstrations and a test instance as its input, and directly makes predictions without updating model parameters. Previous studies have revealed that ICL is sensitive to the selection and the ordering of demonstrations. However, there are few studies regarding the impact of the demonstration number on the ICL performance within a limited input length of LLM, because it is commonly believed that the number of demonstrations is positively correlated with model performance. In this paper, we found this conclusion does not always hold true. Through pilot experiments, we discover that increasing the number of demonstrations does not necessarily lead to improved performance. Building upon this insight, we propose a Dynamic Demonstrations Controller (D$^2$Controller), which can improve the ICL performance by adjusting the number of demonstrations dynamically. The experimental results show that D$^2$Controller yields a 5.4% relative improvement on eight different sizes of LLMs across ten datasets. Moreover, we also extend our method to previous ICL models and achieve competitive results.

In today's digital age, images have emerged as powerful tools for politicians to engage with their voters on social media platforms. Visual content possesses a unique emotional appeal that often leads to increased user engagement. However, research on visual communication remains relatively limited, particularly in the Global South. This study aims to bridge this gap by employing a combination of computational methods and qualitative approach to investigate the visual communication strategies employed in a dataset of 11,263 Instagram posts by 19 Brazilian presidential candidates in 2018 and 2022 national elections. Through two studies, we observed consistent patterns across these candidates on their use of visual political communication. Notably, we identify a prevalence of celebratory and positively toned images. They also exhibit a strong sense of personalization, portraying candidates connected with their voters on a more emotional level. We note a substantial presence of screenshots from news websites and other social media platforms. Furthermore, text-edited images with portrayals emerge as a prominent feature. In light of these results, we engage in a discussion regarding the implications for the broader field of visual political communication. This article serves as a testament to the pivotal role that Instagram has played in shaping the narrative of two fiercely polarized Brazilian elections, casting a revealing light on the ever-evolving dynamics of visual political communication in the digital age. Finally, we propose avenues for future research in the realm of visual political communication. Introduction In the ever-evolving arena of election campaigns, candidates rely heavily on the media as their megaphone to amplify their messages to the masses. Over the years, the landscape of political communication has undergone a profound transformation. This transformation has been driven by the rise of online social media platforms, which have emerged as indispensable tools for candidates in their quest to gauge public sentiment and rally support from the electorate (Boulianne & Olof Larsson, 2023; Farkas & Bene, 2021). The significance of this transformation has been further accentuated by the global ascent of populist leaders, spanning diverse nations, who have wholeheartedly embraced social media as their primary mode of communication (Bernardi & Costa, 2020; Novoselova, 2020).

Implicit communication plays such a crucial role during social exchanges that it must be considered for a good experience in human-robot interaction. This work addresses implicit communication associated with the detection of physical properties, transport, and manipulation of objects. We propose an ecological approach to infer object characteristics from subtle modulations of the natural kinematics occurring during human object manipulation. Similarly, we take inspiration from human strategies to shape robot movements to be communicative of the object properties while pursuing the action goals. In a realistic HRI scenario, participants handed over cups - filled with water or empty - to a robotic manipulator that sorted them. We implemented an online classifier to differentiate careful/not careful human movements, associated with the cups' content. We compared our proposed "expressive" controller, which modulates the movements according to the cup filling, against a neutral motion controller. Results show that human kinematics is adjusted during the task, as a function of the cup content, even in reach-to-grasp motion. Moreover, the carefulness during the handover of full cups can be reliably inferred online, well before action completion. Finally, although questionnaires did not reveal explicit preferences from participants, the expressive robot condition improved task efficiency.

Large Language Models (LLMs) have achieved remarkable success, demonstrating powerful instruction-following capabilities across diverse tasks. Instruction fine-tuning is critical in enabling LLMs to align with user intentions and effectively follow instructions. In this work, we investigate how instruction fine-tuning modifies pre-trained models, focusing on two perspectives: instruction recognition and knowledge evolution. To study the behavior shift of LLMs, we employ a suite of local and global explanation methods, including a gradient-based approach for input-output attribution and techniques for interpreting patterns and concepts in self-attention and feed-forward layers. Our findings reveal three significant impacts of instruction fine-tuning: 1) It empowers LLMs to better recognize the instruction parts from user prompts, thereby facilitating high-quality response generation and addressing the ``lost-in-the-middle'' issue observed in pre-trained models; 2) It aligns the knowledge stored in feed-forward layers with user-oriented tasks, exhibiting minimal shifts across linguistic levels. 3) It facilitates the learning of word-word relations with instruction verbs through the self-attention mechanism, particularly in the lower and middle layers, indicating enhanced recognition of instruction words. These insights contribute to a deeper understanding of the behavior shifts in LLMs after instruction fine-tuning and lay the groundwork for future research aimed at interpreting and optimizing LLMs for various applications. We will release our code and data soon.

Generative data augmentation (GDA) has emerged as a promising technique to alleviate data scarcity in machine learning applications. This thesis presents a comprehensive survey and unified framework of the GDA landscape. We first provide an overview of GDA, discussing its motivation, taxonomy, and key distinctions from synthetic data generation. We then systematically analyze the critical aspects of GDA - selection of generative models, techniques to utilize them, data selection methodologies, validation approaches, and diverse applications. Our proposed unified framework categorizes the extensive GDA literature, revealing gaps such as the lack of universal benchmarks. The thesis summarises promising research directions, including , effective data selection, theoretical development for large-scale models' application in GDA and establishing a benchmark for GDA. By laying a structured foundation, this thesis aims to nurture more cohesive development and accelerate progress in the vital arena of generative data augmentation.

In role-playing games a Game Master (GM) is the player in charge of the game, who must design the challenges the players face and narrate the outcomes of their actions. In this work we discuss some challenges to model GMs from an Interactive Storytelling and Natural Language Processing perspective. Following those challenges we propose three test categories to evaluate such dialogue systems, and we use them to test ChatGPT, Bard and OpenAssistant as out-of-the-box GMs.

While VideoQA Transformer models demonstrate competitive performance on standard benchmarks, the reasons behind their success are not fully understood. Do these models jointly capture and leverage the rich multimodal structures and dynamics from video and text? Or are they merely exploiting shortcuts to achieve high scores? Hence, we design QUAG (QUadrant AveraGe), a lightweight and non-parametric probe, to critically analyze multimodal representations. QUAG facilitates combined dataset-model study by systematic ablation of model's coupled multimodal understanding during inference. Surprisingly, it demonstrates that the models manage to maintain high performance even under multimodal impairment. We extend QUAG to design "QUAG-attention", a simplistic and lessexpressive replacement of self-attention. We find that the models with QUAGattention achieve similar performance with significantly less mulops without any finetuning. These findings indicate that the current VideoQA benchmarks and metrics do not penalize models that find shortcuts and discount joint multimodal understanding. Motivated by this, we propose the CLAVI (Counterfactual in LAnguage and VIdeo), a diagnostic dataset for coupled multimodal understanding in VideoQA. CLAVI consists of temporal questions and videos that are augmented to curate balanced counterfactuals in language and video domains. We evaluate models on CLAVI and find that all models achieve high performance on multimodal shortcut instances, but most of them have very poor performance on the counterfactual instances that necessitate joint multimodal understanding. Overall, with the multimodal representation analysis using QUAG and diagnostic analysis using CLAVI, we show that many VideoQA models are incapable of learning multimodal representations and that their success on standard datasets is an illusion of joint multimodal understanding. Multimodal learning with videos and language is challenging, despite the shared sequential nature of these modalities, due to their distinct underlying structures. That is, videos exhibit spatio-temporal dynamics in the pixel space, whereas language representation is composed of the syntax and semantics of word sequences. Hence, tasks like Video Question Answering (VideoQA) (Zhong et al., 2022) are difficult as they necessitate the model to acquire accurate representations of both the modalities and establish meaningful connections between them. Transformers have demonstrated exceptional performance on VideoQA benchmarks (Zhong et al., 2022).

Concerned with user data privacy, this paper presents a new federated learning (FL) method that trains machine learning models on edge devices without accessing sensitive data. Traditional FL methods, although privacy-protective, fail to manage model heterogeneity and incur high communication costs due to their reliance on aggregation methods. To address this limitation, we propose a resource-aware FL method that aggregates local knowledge from edge models and distills it into robust global knowledge through knowledge distillation. This method allows efficient multi-model knowledge fusion and the deployment of resource-aware models while preserving model heterogeneity. Our method improves communication cost and performance in heterogeneous data and models compared to existing FL algorithms. Notably, it reduces the communication cost of ResNet-32 by up to 50\% and VGG-11 by up to 10$\times$ while delivering superior performance.