Pretrained language models are commonly aligned with human preferences and downstream tasks via reinforcement finetuning (RFT), which entails maximizing a (possibly learned) reward function using policy gradient algorithms. This work highlights a fundamental optimization obstacle in RFT: we prove that the expected gradient for an input vanishes when its reward standard deviation under the model is small, even if the expected reward is far from optimal. Through experiments on an RFT benchmark and controlled environments, as well as a theoretical analysis, we then demonstrate that vanishing gradients due to small reward standard deviation are prevalent and detrimental, leading to extremely slow reward maximization. Lastly, we explore ways to overcome vanishing gradients in RFT. We find the common practice of an initial supervised finetuning (SFT) phase to be the most promising candidate, which sheds light on its importance in an RFT pipeline. Moreover, we show that a relatively small number of SFT optimization steps on as few as 1% of the input samples can suffice, indicating that the initial SFT phase need not be expensive in terms of compute and data labeling efforts. Overall, our results emphasize that being mindful for inputs whose expected gradient vanishes, as measured by the reward standard deviation, is crucial for successful execution of RFT.
相關內容
Neural language models have become powerful tools for learning complex representations of entities in natural language processing tasks. However, their interpretability remains a significant challenge, particularly in domains like computational biology where trust in model predictions is crucial. In this work, we aim to enhance the interpretability of protein language models, specifically the state-of-the-art ESM model, by identifying and characterizing knowledge neurons - components that express understanding of key information. After fine-tuning the ESM model for the task of enzyme sequence classification, we compare two knowledge neuron selection methods that preserve a subset of neurons from the original model. The two methods, activation-based and integrated gradient-based selection, consistently outperform a random baseline. In particular, these methods show that there is a high density of knowledge neurons in the key vector prediction networks of self-attention modules. Given that key vectors specialize in understanding different features of input sequences, these knowledge neurons could capture knowledge of different enzyme sequence motifs. In the future, the types of knowledge captured by each neuron could be characterized.
ChatGPT is currently the most popular large language model (LLM), with over 100 million users, making a significant impact on people's lives. However, due to the presence of jailbreak vulnerabilities, ChatGPT might have negative effects on people's lives, potentially even facilitating criminal activities. Testing whether ChatGPT can cause jailbreak is crucial because it can enhance ChatGPT's security, reliability, and social responsibility. Inspired by previous research revealing the varied performance of LLMs in different language translations, we suspected that wrapping prompts in multiple languages might lead to ChatGPT jailbreak. To investigate this, we designed a study with a fuzzing testing approach to analyzing ChatGPT's cross-linguistic proficiency. Our study includes three strategies by automatically posing different formats of malicious questions to ChatGPT: (1) each malicious question involving only one language, (2) multilingual malicious questions, (3) specifying that ChatGPT responds in a language different from the prompts. In addition, we also combine our strategies by utilizing prompt injection templates to wrap the three aforementioned types of questions. We examined a total of 7,892 Q&A data points, discovering that multilingual wrapping can indeed lead to ChatGPT's jailbreak, with different wrapping methods having varying effects on jailbreak probability. Prompt injection can amplify the probability of jailbreak caused by multilingual wrapping. This work provides insights for OpenAI developers to enhance ChatGPT's support for language diversity and inclusion.
Deep neural networks are becoming increasingly pervasive in academia and industry, matching and surpassing human performance on a wide variety of fields and related tasks. However, just as humans, even the largest artificial neural networks make mistakes, and once-correct predictions can become invalid as the world progresses in time. Augmenting datasets with samples that account for mistakes or up-to-date information has become a common workaround in practical applications. However, the well-known phenomenon of catastrophic forgetting poses a challenge in achieving precise changes in the implicitly memorized knowledge of neural network parameters, often requiring a full model re-training to achieve desired behaviors. That is expensive, unreliable, and incompatible with the current trend of large self-supervised pre-training, making it necessary to find more efficient and effective methods for adapting neural network models to changing data. To address this need, knowledge editing is emerging as a novel area of research that aims to enable reliable, data-efficient, and fast changes to a pre-trained target model, without affecting model behaviors on previously learned tasks. In this survey, we provide a brief review of this recent artificial intelligence field of research. We first introduce the problem of editing neural networks, formalize it in a common framework and differentiate it from more notorious branches of research such as continuous learning. Next, we provide a review of the most relevant knowledge editing approaches and datasets proposed so far, grouping works under four different families: regularization techniques, meta-learning, direct model editing, and architectural strategies. Finally, we outline some intersections with other fields of research and potential directions for future works.
Given large language models' (LLMs) increasing integration into workplace software, it is important to examine how potential biases they introduce can impact workers. Stylistic biases in the language suggested by LLMs may cause feelings of alienation and result in increased labor for individuals or groups whose style does not match. We examine how such writer-style bias impacts inclusion, control, and ownership over the work when co-writing with LLMs. In an online experiment, participants wrote hypothetical job promotion requests using either hesitant or self-assured auto-complete suggestions from an LLM and reported their subsequent perceptions of inclusion, control, and ownership. We found that the style of the AI model did not impact perceived inclusion. However, individuals with higher perceived inclusion did perceive greater agency and ownership, an effect more strongly impacting participants of minoritized genders. Feelings of inclusion mitigated a loss of control and agency when accepting more AI suggestions.
When solving challenging problems, language models (LMs) are able to identify relevant information from long and complicated contexts. To study how LMs solve retrieval tasks in diverse situations, we introduce ORION, a collection of structured retrieval tasks spanning six domains, from text understanding to coding. Each task in ORION can be represented abstractly by a request (e.g. a question) that retrieves an attribute (e.g. the character name) from a context (e.g. a story). We apply causal analysis on 18 open-source language models with sizes ranging from 125 million to 70 billion parameters. We find that LMs internally decompose retrieval tasks in a modular way: middle layers at the last token position process the request, while late layers retrieve the correct entity from the context. After causally enforcing this decomposition, models are still able to solve the original task, preserving 70% of the original correct token probability in 98 of the 106 studied model-task pairs. We connect our macroscopic decomposition with a microscopic description by performing a fine-grained case study of a question-answering task on Pythia-2.8b. Building on our high-level understanding, we demonstrate a proof of concept application for scalable internal oversight of LMs to mitigate prompt-injection while requiring human supervision on only a single input. Our solution improves accuracy drastically (from 15.5% to 97.5% on Pythia-12b). This work presents evidence of a universal emergent modular processing of tasks across varied domains and models and is a pioneering effort in applying interpretability for scalable internal oversight of LMs.
Although large language models (LLMs) are impressive in solving various tasks, they can quickly be outdated after deployment. Maintaining their up-to-date status is a pressing concern in the current era. This paper provides a comprehensive review of recent advances in aligning LLMs with the ever-changing world knowledge without re-training from scratch. We categorize research works systemically and provide in-depth comparisons and discussion. We also discuss existing challenges and highlight future directions to facilitate research in this field. We release the paper list at //github.com/hyintell/awesome-refreshing-llms
Graph neural networks (GNNs) have been demonstrated to be a powerful algorithmic model in broad application fields for their effectiveness in learning over graphs. To scale GNN training up for large-scale and ever-growing graphs, the most promising solution is distributed training which distributes the workload of training across multiple computing nodes. However, the workflows, computational patterns, communication patterns, and optimization techniques of distributed GNN training remain preliminarily understood. In this paper, we provide a comprehensive survey of distributed GNN training by investigating various optimization techniques used in distributed GNN training. First, distributed GNN training is classified into several categories according to their workflows. In addition, their computational patterns and communication patterns, as well as the optimization techniques proposed by recent work are introduced. Second, the software frameworks and hardware platforms of distributed GNN training are also introduced for a deeper understanding. Third, distributed GNN training is compared with distributed training of deep neural networks, emphasizing the uniqueness of distributed GNN training. Finally, interesting issues and opportunities in this field are discussed.
Deep neural networks have revolutionized many machine learning tasks in power systems, ranging from pattern recognition to signal processing. The data in these tasks is typically represented in Euclidean domains. Nevertheless, there is an increasing number of applications in power systems, where data are collected from non-Euclidean domains and represented as the graph-structured data with high dimensional features and interdependency among nodes. The complexity of graph-structured data has brought significant challenges to the existing deep neural networks defined in Euclidean domains. Recently, many studies on extending deep neural networks for graph-structured data in power systems have emerged. In this paper, a comprehensive overview of graph neural networks (GNNs) in power systems is proposed. Specifically, several classical paradigms of GNNs structures (e.g., graph convolutional networks, graph recurrent neural networks, graph attention networks, graph generative networks, spatial-temporal graph convolutional networks, and hybrid forms of GNNs) are summarized, and key applications in power systems such as fault diagnosis, power prediction, power flow calculation, and data generation are reviewed in detail. Furthermore, main issues and some research trends about the applications of GNNs in power systems are discussed.
Recent developments in image classification and natural language processing, coupled with the rapid growth in social media usage, have enabled fundamental advances in detecting breaking events around the world in real-time. Emergency response is one such area that stands to gain from these advances. By processing billions of texts and images a minute, events can be automatically detected to enable emergency response workers to better assess rapidly evolving situations and deploy resources accordingly. To date, most event detection techniques in this area have focused on image-only or text-only approaches, limiting detection performance and impacting the quality of information delivered to crisis response teams. In this paper, we present a new multimodal fusion method that leverages both images and texts as input. In particular, we introduce a cross-attention module that can filter uninformative and misleading components from weak modalities on a sample by sample basis. In addition, we employ a multimodal graph-based approach to stochastically transition between embeddings of different multimodal pairs during training to better regularize the learning process as well as dealing with limited training data by constructing new matched pairs from different samples. We show that our method outperforms the unimodal approaches and strong multimodal baselines by a large margin on three crisis-related tasks.
Neural machine translation (NMT) is a deep learning based approach for machine translation, which yields the state-of-the-art translation performance in scenarios where large-scale parallel corpora are available. Although the high-quality and domain-specific translation is crucial in the real world, domain-specific corpora are usually scarce or nonexistent, and thus vanilla NMT performs poorly in such scenarios. Domain adaptation that leverages both out-of-domain parallel corpora as well as monolingual corpora for in-domain translation, is very important for domain-specific translation. In this paper, we give a comprehensive survey of the state-of-the-art domain adaptation techniques for NMT.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191