Vision-language models (VLMs) [Radford et al., 2021, Li et al., 2022a, Jia et al., 2021, Li et al., 2022c] pre-trained on web-scale image-text pairs have exhibited robust zero-shot prediction capabilities, which have recently attracted increasing attention from the research community. As an example, Contrastive Language-Image Pretraining (CLIP) [Radford et al., 2021] leverages a contrastive objective to obtain a modality-agnostic embedding space in which the paired images and texts are pulled closer and unpaired images and texts are pushed apart. Subsequently, CLIP can perform zero-shot visual prediction by matching the embeddings of test images and prompt-based textual descriptions (e.g., "a photo of a [CLASS]" and "this is a picture of a [CLASS]"), merely requiring the names of all the semantic classes in downstream tasks. Apart from the extensive research dedicated to the pre-training stage, numerous studies [Zhou et al., 2022b, Zhang et al., 2022b, Bahng et al., 2022] have concentrated on adapting VLMs to specific downstream tasks by using task-specific labeled data. This fine-tuning paradigm empowers VLMs to bridge both data and task gaps, leading to improved performance in recognition tasks. In addition to multi-class classification, these pioneering strategies have also been harnessed in a spectrum of computer vision tasks, including ordinal regression [Li et al., 2022b], point cloud understanding [Zhang et al., 2022a], and dense prediction [Rao et al., 2022]. When considering fine-tuning setups, most efforts have primarily revolved around fully supervised and few-shot supervised learning scenarios. To pursue annotation efficiency and scalability, several recent studies [Huang et al., 2022, Shu et al., 2022, Tanwisuth et al., 2023] have delved into the realm of unsupervised fine-tuning for VLMs, remarkably achieving performance on par with few-shot supervised

Recent innovation in large language models (LLMs), and their myriad use-cases have rapidly driven up the compute capacity demand for datacenter GPUs. Several cloud providers and other enterprises have made substantial plans of growth in their datacenters to support these new workloads. One of the key bottleneck resources in datacenters is power, and given the increasing model sizes of LLMs, they are becoming increasingly power intensive. In this paper, we show that there is a significant opportunity to oversubscribe power in LLM clusters. Power oversubscription improves the power efficiency of these datacenters, allowing more deployable servers per datacenter, and reduces the deployment time, since building new datacenters is slow. We extensively characterize the power consumption patterns of a variety of LLMs and their configurations. We identify the differences between the inference and training power consumption patterns. Based on our analysis of these LLMs, we claim that the average and peak power utilization in LLM clusters for inference should not be very high. Our deductions align with the data from production LLM clusters, revealing that inference workloads offer substantial headroom for power oversubscription. However, the stringent set of telemetry and controls that GPUs offer in a virtualized environment, makes it challenging to have a reliable and robust power oversubscription mechanism. We propose POLCA, our framework for power oversubscription that is robust, reliable, and readily deployable for GPU clusters. Using open-source models to replicate the power patterns observed in production, we simulate POLCA and demonstrate that we can deploy 30% more servers in the same GPU cluster for inference, with minimal performance loss

Spurred by the recent rapid increase in the development and distribution of large language models (LLMs) across industry and academia, much recent work has drawn attention to safety- and security-related threats and vulnerabilities of LLMs, including in the context of potentially criminal activities. Specifically, it has been shown that LLMs can be misused for fraud, impersonation, and the generation of malware; while other authors have considered the more general problem of AI alignment. It is important that developers and practitioners alike are aware of security-related problems with such models. In this paper, we provide an overview of existing - predominantly scientific - efforts on identifying and mitigating threats and vulnerabilities arising from LLMs. We present a taxonomy describing the relationship between threats caused by the generative capabilities of LLMs, prevention measures intended to address such threats, and vulnerabilities arising from imperfect prevention measures. With our work, we hope to raise awareness of the limitations of LLMs in light of such security concerns, among both experienced developers and novel users of such technologies.

We propose a method to determine whether a given article was entirely written by a generative language model versus an alternative situation in which the article includes some significant edits by a different author, possibly a human. Our process involves many perplexity tests for the origin of individual sentences or other text atoms, combining these multiple tests using Higher Criticism (HC). As a by-product, the method identifies parts suspected to be edited. The method is motivated by the convergence of the log-perplexity to the cross-entropy rate and by a statistical model for edited text saying that sentences are mostly generated by the language model, except perhaps for a few sentences that might have originated via a different mechanism. We demonstrate the effectiveness of our method using real data and analyze the factors affecting its success. This analysis raises several interesting open challenges whose resolution may improve the method's effectiveness.

Large Language Models (LLMs) present strong general capabilities, and a current compelling challenge is stimulating their specialized capabilities, such as machine translation, through low-cost instruction tuning. The standard instruction-following data is sequentially organized as the concatenation of an instruction, an input, and a response. As the attention mechanism of LLMs has limitations on local focus, LLMs tend to focus more on the words or sentences nearby at each position. This leads to a high risk of instruction forgetting during decoding. To alleviate the above issues, We propose SWIE (Segment-Weighted Instruction Embedding) and an instruction-following dataset OVERMISS. SWIE improves the model instruction understanding by adding a global instruction representation on the following input and response representations. OVERMISS improves model faithfulness by comparing over-translation and miss-translation results with the correct translation. We apply our methods to two main-stream open-source LLMs, BLOOM and LLaMA. The experimental results demonstrate significant improvements in translation performance with SWIE based on BLOOMZ-3b, particularly in zero-shot and long text translations due to reduced instruction forgetting risk. Additionally, OVERMISS outperforms the baseline in translation performance (e.g. an increase in BLEU scores from 0.69 to 3.12 and an average improvement of 0.48 percentage comet scores for LLaMA-7b) with further enhancements seen in models combining OVERMISS and SWIE (e.g. the BLUE scores increase up to 0.56 from English to German across three different backbones), and both exhibit improvements in the faithfulness metric based on word alignment.

Recent researches indicate that Pre-trained Large Language Models (LLMs) possess cognitive constructs similar to those observed in humans, prompting researchers to investigate the cognitive aspects of LLMs. This paper focuses on explicit and implicit social bias, a distinctive two-level cognitive construct in psychology. It posits that individuals' explicit social bias, which is their conscious expression of bias in the statements, may differ from their implicit social bias, which represents their unconscious bias. We propose a two-stage approach and discover a parallel phenomenon in LLMs known as "re-judge inconsistency" in social bias. In the initial stage, the LLM is tasked with automatically completing statements, potentially incorporating implicit social bias. However, in the subsequent stage, the same LLM re-judges the biased statement generated by itself but contradicts it. We propose that this re-judge inconsistency can be similar to the inconsistency between human's unaware implicit social bias and their aware explicit social bias. Experimental investigations on ChatGPT and GPT-4 concerning common gender biases examined in psychology corroborate the highly stable nature of the re-judge inconsistency. This finding may suggest that diverse cognitive constructs emerge as LLMs' capabilities strengthen. Consequently, leveraging psychological theories can provide enhanced insights into the underlying mechanisms governing the expressions of explicit and implicit constructs in LLMs.

The integration of retrieved passages and large language models (LLMs), such as ChatGPTs, has significantly contributed to improving open-domain question answering. However, there is still a lack of exploration regarding the optimal approach for incorporating retrieved passages into the answer generation process. This paper aims to fill this gap by investigating different methods of combining retrieved passages with LLMs to enhance answer generation. We begin by examining the limitations of a commonly-used concatenation approach. Surprisingly, this approach often results in generating "unknown" outputs, even when the correct document is among the top-k retrieved passages. To address this issue, we explore four alternative strategies for integrating the retrieved passages with the LLMs. These strategies include two single-round methods that utilize chain-of-thought reasoning and two multi-round strategies that incorporate feedback loops. Through comprehensive analyses and experiments, we provide insightful observations on how to effectively leverage retrieved passages to enhance the answer generation capability of LLMs.

Variational Information Pursuit (V-IP) is a framework for making interpretable predictions by design by sequentially selecting a short chain of task-relevant, userdefined and interpretable queries about the data that are most informative for the task. While using queries related with semantic concepts allows for built-in interpretability in predictive models, applying V-IP to any task requires data samples with concept-labeling by domain experts, limiting the application of V-IP to smallscale tasks where manual data annotation is feasible. In this work, we extend the V-IP framework with Foundational Models (FMs) to address this limitation. More specifically, we use a two-step process, by first leveraging Large Language Models (LLMs) to generate a sufficiently large candidate set of task-relevant interpretable concepts, then using multimodal models to annotate each data sample by semantic similarity with each concept in the generated concept set. While other interpretableby-design frameworks such as Concept Bottleneck Models (CBMs) require an additional step of removing repetitive and non-discriminative concepts to have good interpretability and test performance, we mathematically and empirically justify that, with a sufficiently informative and task-relevant query (concept) set, the proposed FM+V-IP method does not require any type of concept filtering. In addition, we show that FM+V-IP with LLM generated concepts can achieve better test performance than V-IP with human annotated concepts, demonstrating the effectiveness of LLMs at generating efficient query sets. Finally, when compared to other interpretable-by-design frameworks such as CBMs, FM+V-IP can achieve competitive test performance using fewer number of concepts/queries in both cases with filtered or unfiltered concept sets.

Large language models (LLMs), such as ChatGPT, have emerged with astonishing capabilities approaching artificial general intelligence. While providing convenience for various societal needs, LLMs have also lowered the cost of generating harmful content. Consequently, LLM developers have deployed semantic-level defenses to recognize and reject prompts that may lead to inappropriate content. Unfortunately, these defenses are not foolproof, and some attackers have crafted "jailbreak" prompts that temporarily hypnotize the LLM into forgetting content defense rules and answering any improper questions. To date, there is no clear explanation of the principles behind these semantic-level attacks and defenses in both industry and academia. This paper investigates the LLM jailbreak problem and proposes an automatic jailbreak method for the first time. We propose the concept of a semantic firewall and provide three technical implementation approaches. Inspired by the attack that penetrates traditional firewalls through reverse tunnels, we introduce a "self-deception" attack that can bypass the semantic firewall by inducing LLM to generate prompts that facilitate jailbreak. We generated a total of 2,520 attack payloads in six languages (English, Russian, French, Spanish, Chinese, and Arabic) across seven virtual scenarios, targeting the three most common types of violations: violence, hate, and pornography. The experiment was conducted on two models, namely the GPT-3.5-Turbo and GPT-4. The success rates on the two models were 86.2% and 67%, while the failure rates were 4.7% and 2.2%, respectively. This highlighted the effectiveness of the proposed attack method. All experimental code and raw data will be released as open-source to inspire future research. We believe that manipulating AI behavior through carefully crafted prompts will become an important research direction in the future.

Recently, large-scale pre-trained vision-language models (e.g. CLIP and ALIGN) have demonstrated remarkable effectiveness in acquiring transferable visual representations. To leverage the valuable knowledge encoded within these models for downstream tasks, several fine-tuning approaches, including prompt tuning methods and adapter-based methods, have been developed to adapt vision-language models effectively with supervision. However, these methods rely on the availability of annotated samples, which can be labor-intensive and time-consuming to acquire, thus limiting scalability. To address this issue, in this work, we design an unsupervised fine-tuning approach for vision-language models called Unsupervised Prototype Adapter (UP-Adapter). Specifically, for the unannotated target datasets, we leverage the text-image aligning capability of CLIP to automatically select the most confident samples for each class. Utilizing these selected samples, we generate class prototypes, which serve as the initialization for the learnable prototype model. After fine-tuning, the prototype model prediction is combined with the original CLIP's prediction by a residual connection to perform downstream recognition tasks. Our extensive experimental results on image recognition and domain generalization show that the proposed unsupervised method outperforms 8-shot CoOp, 8-shot Tip-Adapter, and also the state-of-the-art UPL method by large margins.