Aligning large language models (LLMs) with human intentions has become a critical task for safely deploying models in real-world systems. While existing alignment approaches have seen empirical success, theoretically understanding how these methods affect model behavior remains an open question. Our work provides an initial attempt to theoretically analyze the learning dynamics of human preference alignment. We formally show how the distribution of preference datasets influences the rate of model updates and provide rigorous guarantees on the training accuracy. Our theory also reveals an intricate phenomenon where the optimization is prone to prioritizing certain behaviors with higher preference distinguishability. We empirically validate our findings on contemporary LLMs and alignment tasks, reinforcing our theoretical insights and shedding light on considerations for future alignment approaches. Disclaimer: This paper contains potentially offensive text; reader discretion is advised.
相關內容
The logic of information flows (LIF) has recently been proposed as a general framework in the field of knowledge representation. In this framework, tasks of procedural nature can still be modeled in a declarative, logic-based fashion. In this paper, we focus on the task of query processing under limited access patterns, a well-studied problem in the database literature. We show that LIF is well-suited for modeling this task. Toward this goal, we introduce a variant of LIF called "forward" LIF (FLIF), in a first-order setting. FLIF takes a novel graph-navigational approach; it is an XPath-like language that nevertheless turns out to be equivalent to the "executable" fragment of first-order logic defined by Nash and Lud\"ascher. One can also classify the variables in FLIF expressions as inputs and outputs. Expressions where inputs and outputs are disjoint, referred to as io-disjoint FLIF expressions, allow a particularly transparent translation into algebraic query plans that respect the access limitations. Finally, we show that general FLIF expressions can always be put into io-disjoint form.
Hidden Markov models (HMM) have been widely used by scientists to model stochastic systems: the underlying process is a discrete Markov chain and the observations are noisy realizations of the underlying process. Determining the number of hidden states for an HMM is a model selection problem, which is yet to be satisfactorily solved, especially for the popular Gaussian HMM with heterogeneous covariance. In this paper, we propose a consistent method for determining the number of hidden states of HMM based on the marginal likelihood, which is obtained by integrating out both the parameters and hidden states. Moreover, we show that the model selection problem of HMM includes the order selection problem of finite mixture models as a special case. We give rigorous proof of the consistency of the proposed marginal likelihood method and provide an efficient computation method for practical implementation. We numerically compare the proposed method with the Bayesian information criterion (BIC), demonstrating the effectiveness of the proposed marginal likelihood method.
In recent years, the integration of large language models (LLMs) has revolutionized the field of robotics, enabling robots to communicate, understand, and reason with human-like proficiency. This paper explores the multifaceted impact of LLMs on robotics, addressing key challenges and opportunities for leveraging these models across various domains. By categorizing and analyzing LLM applications within core robotics elements -- communication, perception, planning, and control -- we aim to provide actionable insights for researchers seeking to integrate LLMs into their robotic systems. Our investigation focuses on LLMs developed post-GPT-3.5, primarily in text-based modalities while also considering multimodal approaches for perception and control. We offer comprehensive guidelines and examples for prompt engineering, facilitating beginners' access to LLM-based robotics solutions. Through tutorial-level examples and structured prompt construction, we illustrate how LLM-guided enhancements can be seamlessly integrated into robotics applications. This survey serves as a roadmap for researchers navigating the evolving landscape of LLM-driven robotics, offering a comprehensive overview and practical guidance for harnessing the power of language models in robotics development.
Large language models (LLMs) have achieved human-level text generation, emphasizing the need for effective AI-generated text detection to mitigate risks like the spread of fake news and plagiarism. Existing research has been constrained by evaluating detection methods on specific domains or particular language models. In practical scenarios, however, the detector faces texts from various domains or LLMs without knowing their sources. To this end, we build a comprehensive testbed by gathering texts from diverse human writings and texts generated by different LLMs. Empirical results show challenges in distinguishing machine-generated texts from human-authored ones across various scenarios, especially out-of-distribution. These challenges are due to the decreasing linguistic distinctions between the two sources. Despite challenges, the top-performing detector can identify 86.54% out-of-domain texts generated by a new LLM, indicating the feasibility for application scenarios. We release our resources at //github.com/yafuly/MAGE.
In the era of generative artificial intelligence (AI), the fusion of large language models (LLMs) offers unprecedented opportunities for innovation in the field of modern education. We embark on an exploration of prompted LLMs within the context of educational and assessment applications to uncover their potential. Through a series of carefully crafted research questions, we investigate the effectiveness of prompt-based techniques in generating open-ended questions from school-level textbooks, assess their efficiency in generating open-ended questions from undergraduate-level technical textbooks, and explore the feasibility of employing a chain-of-thought inspired multi-stage prompting approach for language-agnostic multiple-choice question (MCQ) generation. Additionally, we evaluate the ability of prompted LLMs for language learning, exemplified through a case study in the low-resource Indian language Bengali, to explain Bengali grammatical errors. We also evaluate the potential of prompted LLMs to assess human resource (HR) spoken interview transcripts. By juxtaposing the capabilities of LLMs with those of human experts across various educational tasks and domains, our aim is to shed light on the potential and limitations of LLMs in reshaping educational practices.
In the rapidly advancing field of artificial intelligence, the concept of Red-Teaming or Jailbreaking large language models (LLMs) has emerged as a crucial area of study. This approach is especially significant in terms of assessing and enhancing the safety and robustness of these models. This paper investigates the intricate consequences of such modifications through model editing, uncovering a complex relationship between enhancing model accuracy and preserving its ethical integrity. Our in-depth analysis reveals a striking paradox: while injecting accurate information is crucial for model reliability, it can paradoxically destabilize the model's foundational framework, resulting in unpredictable and potentially unsafe behaviors. Additionally, we propose a benchmark dataset NicheHazardQA to investigate this unsafe behavior both within the same and cross topical domain. This aspect of our research sheds light on how the edits, impact the model's safety metrics and guardrails. Our findings show that model editing serves as a cost-effective tool for topical red-teaming by methodically applying targeted edits and evaluating the resultant model behavior.
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.
Large language models (LLMs) have become crucial for many generative downstream tasks, leading to an inevitable trend and significant challenge to deploy them efficiently on resource-constrained devices. Structured pruning is a widely used method to address this challenge. However, when dealing with the complex structure of the multiple decoder layers, general methods often employ common estimation approaches for pruning. These approaches lead to a decline in accuracy for specific downstream tasks. In this paper, we introduce a simple yet efficient method that adaptively models the importance of each substructure. Meanwhile, it can adaptively fuse coarse-grained and finegrained estimations based on the results from complex and multilayer structures. All aspects of our design seamlessly integrate into the endto-end pruning framework. Our experimental results, compared with state-of-the-art methods on mainstream datasets, demonstrate average accuracy improvements of 1.1%, 1.02%, 2.0%, and 1.2% for LLaMa-7B,Vicuna-7B, Baichuan-7B, and Bloom-7b1, respectively.
Large language models (LLMs) are transforming human-computer interaction and conceptions of artificial intelligence (AI) with their impressive capacities for conversing and reasoning in natural language. There is growing interest in whether LLMs have theory of mind (ToM); the ability to reason about the mental and emotional states of others that is core to human social intelligence. As LLMs are integrated into the fabric of our personal, professional and social lives and given greater agency to make decisions with real-world consequences, there is a critical need to understand how they can be aligned with human values. ToM seems to be a promising direction of inquiry in this regard. Following the literature on the role and impacts of human ToM, this paper identifies key areas in which LLM ToM will show up in human:LLM interactions at individual and group levels, and what opportunities and risks for alignment are raised in each. On the individual level, the paper considers how LLM ToM might manifest in goal specification, conversational adaptation, empathy and anthropomorphism. On the group level, it considers how LLM ToM might facilitate collective alignment, cooperation or competition, and moral judgement-making. The paper lays out a broad spectrum of potential implications and suggests the most pressing areas for future research.
Large language models (LLMs) are now being used in a wide variety of contexts, including as creativity support tools (CSTs) intended to help their users come up with new ideas. But do LLMs actually support user creativity? We hypothesized that the use of an LLM as a CST might make the LLM's users feel more creative, and even broaden the range of ideas suggested by each individual user, but also homogenize the ideas suggested by different users. We conducted a 36-participant comparative user study and found, in accordance with the homogenization hypothesis, that different users tended to produce less semantically distinct ideas with ChatGPT than with an alternative CST. Additionally, ChatGPT users generated a greater number of more detailed ideas, but felt less responsible for the ideas they generated. We discuss potential implications of these findings for users, designers, and developers of LLM-based CSTs.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191