Vision-language models (VLMs) are impactful in part because they can be applied to a variety of visual understanding tasks in a zero-shot fashion, without any fine-tuning. We study $\textit{generative VLMs}$ that are trained for next-word generation given an image. We explore their zero-shot performance on the illustrative task of image-text retrieval across 8 popular vision-language benchmarks. Our first observation is that they can be repurposed for discriminative tasks (such as image-text retrieval) by simply computing the match score of generating a particular text string given an image. We call this probabilistic score the $\textit{Visual Generative Pre-Training Score}$ (VisualGPTScore). While the VisualGPTScore produces near-perfect accuracy on some retrieval benchmarks, it yields poor accuracy on others. We analyze this behavior through a probabilistic lens, pointing out that some benchmarks inadvertently capture unnatural language distributions by creating adversarial but unlikely text captions. In fact, we demonstrate that even a "blind" language model that ignores any image evidence can sometimes outperform all prior art, reminiscent of similar challenges faced by the visual-question answering (VQA) community many years ago. We derive a probabilistic post-processing scheme that controls for the amount of linguistic bias in generative VLMs at test time without having to retrain or fine-tune the model. We show that the VisualGPTScore, when appropriately debiased, is a strong zero-shot baseline for vision-language understanding, oftentimes producing state-of-the-art accuracy.
相關內容
Classification algorithms using Transformer architectures can be affected by the sequence length learning problem whenever observations from different classes have a different length distribution. This problem causes models to use sequence length as a predictive feature instead of relying on important textual information. Although most public datasets are not affected by this problem, privately owned corpora for fields such as medicine and insurance may carry this data bias. The exploitation of this sequence length feature poses challenges throughout the value chain as these machine learning models can be used in critical applications. In this paper, we empirically expose this problem and present approaches to minimize its impacts.
The commercialization of large language models (LLMs) has led to the common practice of high-level API-only access to proprietary models. In this work, we show that even with a conservative assumption about the model architecture, it is possible to learn a surprisingly large amount of non-public information about an API-protected LLM from a relatively small number of API queries (e.g., costing under $1,000 for OpenAI's gpt-3.5-turbo). Our findings are centered on one key observation: most modern LLMs suffer from a softmax bottleneck, which restricts the model outputs to a linear subspace of the full output space. We show that this lends itself to a model image or a model signature which unlocks several capabilities with affordable cost: efficiently discovering the LLM's hidden size, obtaining full-vocabulary outputs, detecting and disambiguating different model updates, identifying the source LLM given a single full LLM output, and even estimating the output layer parameters. Our empirical investigations show the effectiveness of our methods, which allow us to estimate the embedding size of OpenAI's gpt-3.5-turbo to be about 4,096. Lastly, we discuss ways that LLM providers can guard against these attacks, as well as how these capabilities can be viewed as a feature (rather than a bug) by allowing for greater transparency and accountability.
Large language models (LLMs) are complex artificial intelligence systems capable of understanding, generating and translating human language. They learn language patterns by analyzing large amounts of text data, allowing them to perform writing, conversation, summarizing and other language tasks. When LLMs process and generate large amounts of data, there is a risk of leaking sensitive information, which may threaten data privacy. This paper concentrates on elucidating the data privacy concerns associated with LLMs to foster a comprehensive understanding. Specifically, a thorough investigation is undertaken to delineate the spectrum of data privacy threats, encompassing both passive privacy leakage and active privacy attacks within LLMs. Subsequently, we conduct an assessment of the privacy protection mechanisms employed by LLMs at various stages, followed by a detailed examination of their efficacy and constraints. Finally, the discourse extends to delineate the challenges encountered and outline prospective directions for advancement in the realm of LLM privacy protection.
Establishing whether language models can use contextual information in a human-plausible way is important to ensure their trustworthiness in real-world settings. However, the questions of when and which parts of the context affect model generations are typically tackled separately, with current plausibility evaluations being practically limited to a handful of artificial benchmarks. To address this, we introduce Plausibility Evaluation of Context Reliance (PECoRe), an end-to-end interpretability framework designed to quantify context usage in language models' generations. Our approach leverages model internals to (i) contrastively identify context-sensitive target tokens in generated texts and (ii) link them to contextual cues justifying their prediction. We use \pecore to quantify the plausibility of context-aware machine translation models, comparing model rationales with human annotations across several discourse-level phenomena. Finally, we apply our method to unannotated model translations to identify context-mediated predictions and highlight instances of (im)plausible context usage throughout generation.
The presence of toxic and gender-identity derogatory language in open-source software (OSS) communities has recently become a focal point for researchers. Such comments not only lead to frustration and disengagement among developers but may also influence their leave from the OSS projects. Despite ample evidence suggesting that diverse teams enhance productivity, the existence of toxic or gender identity discriminatory communications poses a significant threat to the participation of individuals from marginalized groups and, as such, may act as a barrier to fostering diversity and inclusion in OSS projects. However, there is a notable lack of research dedicated to exploring the association between gender-based toxic and derogatory language with a perceptible diversity of open-source software teams. Consequently, this study aims to investigate how such content influences the gender, ethnicity, and tenure diversity of open-source software development teams. To achieve this, we extract data from active GitHub projects, assess various project characteristics, and identify instances of toxic and gender-discriminatory language within issue/pull request comments. Using these attributes, we construct a regression model to explore how they associate with the perceptible diversity of those projects.
Although large language models (LLMs) have demonstrated impressive text generation capabilities, they are easily misled by the untruthful context provided by users or knowledge argumentation tools, thereby producing hallucinations. To alleviate the LLMs from being misled by untruthful information and take advantage of knowledge argumentation, we propose Truth-Aware Context Selection (TACS), a lightweight method to shield untruthful context from the inputs. TACS begins by performing truth detection on the input context, leveraging the parameterized knowledge within the LLM. Subsequently, it constructs a corresponding attention mask based on the truthfulness of each position, selecting the truthful context and discarding the untruthful context. Additionally, we introduce a new evaluation metric, Disturbance Adaption Rate, to further study the LLMs' ability to accept truthful information and resist untruthful information. Experimental results show that TACS can effectively filter information in context and significantly improve the overall quality of LLMs' responses when presented with misleading information.
Despite the rapid development of large language models (LLMs) for the Korean language, there remains an obvious lack of benchmark datasets that test the requisite Korean cultural and linguistic knowledge. Because many existing Korean benchmark datasets are derived from the English counterparts through translation, they often overlook the different cultural contexts. For the few benchmark datasets that are sourced from Korean data capturing cultural knowledge, only narrow tasks such as bias and hate speech detection are offered. To address this gap, we introduce a benchmark of Cultural and Linguistic Intelligence in Korean (CLIcK), a dataset comprising 1,995 QA pairs. CLIcK sources its data from official Korean exams and textbooks, partitioning the questions into eleven categories under the two main categories of language and culture. For each instance in CLIcK, we provide fine-grained annotation of which cultural and linguistic knowledge is required to answer the question correctly. Using CLIcK, we test 13 language models to assess their performance. Our evaluation uncovers insights into their performances across the categories, as well as the diverse factors affecting their comprehension. CLIcK offers the first large-scale comprehensive Korean-centric analysis of LLMs' proficiency in Korean culture and language.
Large language models (LLMs) such as ChatGPT have shown remarkable capabilities in code generation. Despite the great achievement, they rely on enormous training data to acquire a broad spectrum of open-domain knowledge. Besides, their evaluation revolves around open-domain benchmarks like HumanEval, which primarily consist of programming contests. Therefore, it is hard to fully characterize the intricacies and challenges associated with particular domains (e.g., web, game, and math). In this paper, we conduct an in-depth study of the LLMs in domain-specific code generation. Our results demonstrate that LLMs exhibit sub-optimal performance in generating domain-specific code, due to their limited proficiency in utilizing domain-specific libraries. We further observe that incorporating API knowledge as prompts can empower LLMs to generate more professional code. Based on these findings, we further investigate how to efficiently incorporate API knowledge into the code generation process. We experiment with three strategies for incorporating domain knowledge, namely, external knowledge inquirer, chain-of-thought prompting, and chain-of-thought fine-tuning. We refer to these strategies as a new code generation approach called DomCoder. Experimental results show that all strategies of DomCoder lead to improvement in the effectiveness of domain-specific code generation under certain settings. The results also show that there is still ample room for further improvement, based on which we suggest possible future works.
Large language models (LLMs) have made significant progress in NLP. However, their ability to memorize, represent, and leverage commonsense knowledge has been a well-known pain point. In this paper, we specifically focus on ChatGPT, a widely used and easily accessible LLM, and ask the following questions: (1) Can ChatGPT effectively answer commonsense questions? (2) Is ChatGPT aware of the underlying commonsense knowledge for answering a specific question? (3) Is ChatGPT knowledgeable in commonsense? (4) Can ChatGPT effectively leverage commonsense for answering questions? We conduct a series of experiments on 11 datasets to evaluate ChatGPT's commonsense abilities, including answering commonsense questions, identifying necessary knowledge, generating knowledge descriptions, and using knowledge descriptions to answer questions again. Experimental results show that: (1) ChatGPT can achieve good QA accuracies in commonsense tasks, while still struggling with certain domains of datasets. (2) ChatGPT is knowledgeable, and can accurately generate most of the commonsense knowledge using knowledge prompts. (3) Despite its knowledge, ChatGPT is an inexperienced commonsense problem solver, which cannot precisely identify the needed commonsense for answering a specific question. These findings raise the need to explore improved mechanisms for effectively incorporating commonsense into LLMs like ChatGPT, such as better instruction following and commonsense guidance.
Large language models (LLMs) exhibit superior performance on various natural language tasks, but they are susceptible to issues stemming from outdated data and domain-specific limitations. In order to address these challenges, researchers have pursued two primary strategies, knowledge editing and retrieval augmentation, to enhance LLMs by incorporating external information from different aspects. Nevertheless, there is still a notable absence of a comprehensive survey. In this paper, we propose a review to discuss the trends in integration of knowledge and large language models, including taxonomy of methods, benchmarks, and applications. In addition, we conduct an in-depth analysis of different methods and point out potential research directions in the future. We hope this survey offers the community quick access and a comprehensive overview of this research area, with the intention of inspiring future research endeavors.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191