Large language models (LLMs) have shown remarkable success across a wide range of natural language generation tasks, where proper prompt designs make great impacts. While existing prompting methods are normally restricted to providing correct information, in this paper, we encourage the model to deliberate by proposing a novel Deliberate then Generate (DTG) prompting framework, which consists of error detection instructions and candidates that may contain errors. DTG is a simple yet effective technique that can be applied to various text generation tasks with minimal modifications. We conduct extensive experiments on 20+ datasets across 7 text generation tasks, including summarization, translation, dialogue, and more. We show that DTG consistently outperforms existing prompting methods and achieves state-of-the-art performance on multiple text generation tasks. We also provide in-depth analyses to reveal the underlying mechanisms of DTG, which may inspire future research on prompting for LLMs.
Modern large language models (LLMs) like ChatGPT have shown remarkable performance on general language tasks but still struggle on complex reasoning tasks, which drives the research on cognitive behaviors of LLMs to explore human-like problem-solving strategies. Along this direction, one representative strategy is self-reflection, which asks an LLM to refine the solution with the feedback generated by itself iteratively. However, our study shows that such reflection-style methods suffer from the Degeneration-of-Thought (DoT) problem: once the LLM has established confidence in its solutions, it is unable to generate novel thoughts later through reflection even if its initial stance is incorrect. To address the DoT problem, we propose a Multi-Agent Debate (MAD) framework, in which multiple agents express their arguments in the state of "tit for tat" and a judge manages the debate process to obtain a final solution. Clearly, our MAD framework encourages divergent thinking in LLMs which would be helpful for tasks that require deep levels of contemplation. Experiment results on two challenging datasets, commonsense machine translation and counter-intuitive arithmetic reasoning, demonstrate the effectiveness of our MAD framework. Extensive analyses suggest that the adaptive break of debate and the modest level of "tit for tat" state are required for MAD to obtain good performance. Moreover, we find that LLMs might not be a fair judge if different LLMs are used for agents. Codes: https://github.com/Skytliang/Multi-Agents-Debate
Years have passed since the NLP community has last focused on linguistic acceptability. In this work, we revisit this topic in the context of large language models. We introduce CoLAC - Corpus of Linguistic Acceptability in Chinese, the first large-scale non-English acceptability dataset that is verified by native speakers and comes with two sets of labels. Our experiments show that even the largest InstructGPT model performs only at chance level on CoLAC, while ChatGPT's performance (48.30 MCC) is also way below supervised models (59.03 MCC) and human (65.11 MCC). Through cross-lingual transfer experiments and fine-grained linguistic analysis, we demonstrate for the first time that knowledge of linguistic acceptability can be transferred across typologically distinct languages, as well as be traced back to pre-training.
Tense inconsistency frequently occurs in machine translation. However, there are few criteria to assess the model's mastery of tense prediction from a linguistic perspective. In this paper, we present a parallel tense test set, containing French-English 552 utterances. We also introduce a corresponding benchmark, tense prediction accuracy. With the tense test set and the benchmark, researchers are able to measure the tense consistency performance of machine translation systems for the first time.
Numerous works are proposed to improve or evaluate the capabilities of Large language models (LLMs) to fulfill user instructions. However, they neglect the possibility that user inputs may inherently contain incorrect information due to users' false beliefs or malicious intents. In this way, blindly adhering to users' false content will cause deception and harm. To address this problem, we propose a challenging benchmark consisting of Inductive Instructions (INDust) to evaluate whether LLMs could resist these instructions. The INDust includes 15K instructions across three categories: Fact-Checking Instructions, Questions based on False Premises, and Creative Instructions based on False Premises. Our experiments on several strong LLMs reveal that current LLMs can be easily deceived by INDust into generating misleading and malicious statements. Hence we employ Self-Critique prompting to encourage LLMs to not only critique themselves like in previous works but also the users, which show remarkable improvement in handling inductive instructions under both zero-shot and few-shot settings.
Automatic summarization generates concise summaries that contain key ideas of source documents. As the most mainstream datasets for the news sub-domain, CNN/DailyMail and BBC XSum have been widely used for performance benchmarking. However, the reference summaries of those datasets turn out to be noisy, mainly in terms of factual hallucination and information redundancy. To address this challenge, we first annotate new expert-writing Element-aware test sets following the "Lasswell Communication Model" proposed by Lasswell (1948), allowing reference summaries to focus on more fine-grained news elements objectively and comprehensively. Utilizing the new test sets, we observe the surprising zero-shot summary ability of LLMs, which addresses the issue of the inconsistent results between human preference and automatic evaluation metrics of LLMs' zero-shot summaries in prior work. Further, we propose a Summary Chain-of-Thought (SumCoT) technique to elicit LLMs to generate summaries step by step, which helps them integrate more fine-grained details of source documents into the final summaries that correlate with the human writing mindset. Experimental results show our method outperforms state-of-the-art fine-tuned PLMs and zero-shot LLMs by +4.33/+4.77 in ROUGE-L on the two datasets, respectively. Dataset and code are publicly available at https://github.com/Alsace08/SumCoT.
Nearest Neighbor Machine Translation ($k$NN-MT) has achieved great success on domain adaptation tasks by integrating pre-trained Neural Machine Translation (NMT) models with domain-specific token-level retrieval. However, the reasons underlying its success have not been thoroughly investigated. In this paper, we provide a comprehensive analysis of $k$NN-MT through theoretical and empirical studies. Initially, we offer a theoretical interpretation of the working mechanism of $k$NN-MT as an efficient technique to implicitly execute gradient descent on the output projection layer of NMT, indicating that it is a specific case of model fine-tuning. Subsequently, we conduct multi-domain experiments and word-level analysis to examine the differences in performance between $k$NN-MT and entire-model fine-tuning. Our findings suggest that: (1) Incorporating $k$NN-MT with adapters yields comparable translation performance to fine-tuning on in-domain test sets, while achieving better performance on out-of-domain test sets; (2) Fine-tuning significantly outperforms $k$NN-MT on the recall of low-frequency domain-specific words, but this gap could be bridged by optimizing the context representations with additional adapter layers.
Sparsity of a learning solution is a desirable feature in machine learning. Certain reproducing kernel Banach spaces (RKBSs) are appropriate hypothesis spaces for sparse learning methods. The goal of this paper is to understand what kind of RKBSs can promote sparsity for learning solutions. We consider two typical learning models in an RKBS: the minimum norm interpolation (MNI) problem and the regularization problem. We first establish an explicit representer theorem for solutions of these problems, which represents the extreme points of the solution set by a linear combination of the extreme points of the subdifferential set, of the norm function, which is data-dependent. We then propose sufficient conditions on the RKBS that can transform the explicit representation of the solutions to a sparse kernel representation having fewer terms than the number of the observed data. Under the proposed sufficient conditions, we investigate the role of the regularization parameter on sparsity of the regularized solutions. We further show that two specific RKBSs: the sequence space $\ell_1(\mathbb{N})$ and the measure space can have sparse representer theorems for both MNI and regularization models.
The way and content in which users ask questions can provide insight into their current status, including their personality, emotions, and psychology. Instead of directly prompting the large language models (LLMs), we explore how chain-of-thought prompting helps in this scenario to perform reasoning and planning according to user status, aiming to provide a more personalized and engaging experience for the user query. To this end, we first construct a benchmark of 6 dialogue or question-answering datasets in both English and Chinese, covering 3 different aspects of user status (\textit{including} \textit{personality}, \textit{emotion}, and \textit{psychology}). Then we prompt the LLMs to generate the response regarding the user status as intermediate reasoning processing. We propose a novel demonstration selection strategy using the semantic similarity of intermediate reasoning instead of test queries. To evaluate the effectiveness and robustness of our approach, we conduct extensive experiments with 7 LLMs under zero-shot and one-shot settings. The experimental results show that our approach consistently outperforms standard prompting in terms of both \textit{helpfulness} and \textit{acceptness} across all datasets, regardless of the LLMs used. The code and dataset can be found at \url{https://github.com/ruleGreen/Dialogue\_CoT.git}.
Knowledge distillation has attracted a great deal of interest recently to compress pre-trained language models. However, existing knowledge distillation methods suffer from two limitations. First, the student model simply imitates the teacher's behavior while ignoring the underlying reasoning. Second, these methods usually focus on the transfer of sophisticated model-specific knowledge but overlook data-specific knowledge. In this paper, we present a novel attribution-driven knowledge distillation approach, which explores the token-level rationale behind the teacher model based on Integrated Gradients (IG) and transfers attribution knowledge to the student model. To enhance the knowledge transfer of model reasoning and generalization, we further explore multi-view attribution distillation on all potential decisions of the teacher. Comprehensive experiments are conducted with BERT on the GLUE benchmark. The experimental results demonstrate the superior performance of our approach to several state-of-the-art methods.