Methods for adapting language models (LMs) to new tasks and domains have traditionally assumed white-box access to the model, and work by modifying its parameters. However, this is incompatible with a recent trend in the field, where the highest quality models are only available as black-boxes through inference APIs. Even when the model weights are available, the computational cost of fine-tuning large LMs can be prohibitive for most practitioners. In this work, we present a lightweight method for adapting large LMs to new domains and tasks, assuming no access to their weights or intermediate activations. Our approach fine-tunes a small white-box LM and combines it with the large black-box LM at the probability level through a small network, learned on a small validation set. We validate our approach by adapting a large LM (OPT-30B) to several domains and a downstream task (machine translation), observing improved performance in all cases, of up to 9%, while using a domain expert 23x smaller.

ChatGPT, launched in November 2022, has gained widespread attention from students and educators globally, with an online report by Hu (2023) stating it as the fastest-growing consumer application in history. While discussions on the use of ChatGPT in higher education are abundant, empirical studies on its impact on collaborative interdisciplinary learning are rare. To investigate its potential, we conducted a quasi-experimental study with 130 undergraduate students (STEM and non-STEM) learning digital literacy with or without ChatGPT over two weeks. Weekly surveys were conducted on collaborative interdisciplinary problem-solving, physical and cognitive engagement, and individual reflections on ChatGPT use. Analysis of survey responses showed significant main effects of topics on collaborative interdisciplinary problem-solving and physical and cognitive engagement, a marginal interaction effect between disciplinary backgrounds and ChatGPT conditions for cognitive engagement, and a significant interaction effect for physical engagement. Sentiment analysis of student reflections suggested no significant difference between STEM and non-STEM students' opinions towards ChatGPT. Qualitative analysis of reflections generated eight positive themes, including efficiency, addressing knowledge gaps, and generating human-like responses, and eight negative themes, including generic responses, lack of innovation, and counterproductive to self-discipline and thinking. Our findings suggest that ChatGPT use needs to be optimized by considering the topics being taught and the disciplinary backgrounds of students rather than applying it uniformly. These findings have implications for both pedagogical research and practices.

The rise of big data has amplified the need for efficient, user-friendly automated machine learning (AutoML) tools. However, the intricacy of understanding domain-specific data and defining prediction tasks necessitates human intervention making the process time-consuming while preventing full automation. Instead, envision an intelligent agent capable of assisting users in conducting AutoML tasks through intuitive, natural conversations without requiring in-depth knowledge of the underlying machine learning (ML) processes. This agent's key challenge is to accurately comprehend the user's prediction goals and, consequently, formulate precise ML tasks, adjust data sets and model parameters accordingly, and articulate results effectively. In this paper, we take a pioneering step towards this ambitious goal by introducing a ChatGPT-based conversational data-science framework to act as a "personal data scientist". Precisely, we utilize Large Language Models (ChatGPT) to build a natural interface between the users and the ML models (Scikit-Learn), which in turn, allows us to approach this ambitious problem with a realistic solution. Our model pivots around four dialogue states: Data Visualization, Task Formulation, Prediction Engineering, and Result Summary and Recommendation. Each state marks a unique conversation phase, impacting the overall user-system interaction. Multiple LLM instances, serving as "micro-agents", ensure a cohesive conversation flow, granting us granular control over the conversation's progression. In summary, we developed an end-to-end system that not only proves the viability of the novel concept of conversational data science but also underscores the potency of LLMs in solving complex tasks. Interestingly, its development spotlighted several critical weaknesses in the current LLMs (ChatGPT) and highlighted substantial opportunities for improvement.

Strategies such as chain-of-thought prompting improve the performance of large language models (LLMs) on complex reasoning tasks by decomposing input examples into intermediate steps. However, it remains unclear how to apply such methods to reason over long input documents, in which both the decomposition and the output of each intermediate step are non-trivial to obtain. In this work, we propose PEARL, a prompting framework to improve reasoning over long documents, which consists of three stages: action mining, plan formulation, and plan execution. More specifically, given a question about a long document, PEARL decomposes the question into a sequence of actions (e.g., SUMMARIZE, FIND_EVENT, FIND_RELATION) and then executes them over the document to obtain the answer. Each stage of PEARL is implemented via zero-shot or few-shot prompting of LLMs (in our work, GPT-4) with minimal human input. We evaluate PEARL on a challenging subset of the QuALITY dataset, which contains questions that require complex reasoning over long narrative texts. PEARL outperforms zero-shot and chain-of-thought prompting on this dataset, and ablation experiments show that each stage of PEARL is critical to its performance. Overall, PEARL is a first step towards leveraging LLMs to reason over long documents.

Transformer models bring propelling advances in various NLP tasks, thus inducing lots of interpretability research on the learned representations of the models. However, we raise a fundamental question regarding the reliability of the representations. Specifically, we investigate whether transformers learn essentially isomorphic representation spaces, or those that are sensitive to the random seeds in their pretraining process. In this work, we formulate the Bijection Hypothesis, which suggests the use of bijective methods to align different models' representation spaces. We propose a model based on invertible neural networks, BERT-INN, to learn the bijection more effectively than other existing bijective methods such as the canonical correlation analysis (CCA). We show the advantage of BERT-INN both theoretically and through extensive experiments, and apply it to align the reproduced BERT embeddings to draw insights that are meaningful to the interpretability research. Our code is at https://github.com/twinkle0331/BERT-similarity.

The answering quality of an aligned large language model (LLM) can be drastically improved if treated with proper crafting of prompts. In this paper, we propose ExpertPrompting to elicit the potential of LLMs to answer as distinguished experts. We first utilize In-Context Learning to automatically synthesize detailed and customized descriptions of the expert identity for each specific instruction, and then ask LLMs to provide answer conditioned on such agent background. Based on this augmented prompting strategy, we produce a new set of instruction-following data using GPT-3.5, and train a competitive open-source chat assistant called ExpertLLaMA. We employ GPT4-based evaluation to show that 1) the expert data is of significantly higher quality than vanilla answers, and 2) ExpertLLaMA outperforms existing open-source opponents and achieves 96\% of the original ChatGPT's capability. All data and the ExpertLLaMA model will be made publicly available at \url{https://github.com/OFA-Sys/ExpertLLaMA}.

Large language model (LLM)-based decision-making agents have shown the ability to generalize across multiple tasks. However, their performance relies on massive data and compute. We argue that this inefficiency stems from the forgetting phenomenon, in which a model memorizes its behaviors in parameters throughout training. As a result, training on a new task may deteriorate the model's performance on previous tasks. In contrast to LLMs' implicit memory mechanism, the human brain utilizes distributed memory storage, which helps manage and organize multiple skills efficiently, mitigating the forgetting phenomenon. Thus inspired, we propose an internal working memory module to store, blend, and retrieve information for different downstream tasks. Evaluation results show that the proposed method improves training efficiency and generalization in both Atari games and meta-world object manipulation tasks. Moreover, we demonstrate that memory fine-tuning further enhances the adaptability of the proposed architecture.

ChatGPT has stimulated the research boom in the field of large language models. In this paper, we assess the capabilities of ChatGPT from four perspectives including Performance, Evaluation Criteria, Robustness and Error Types. Specifically, we first evaluate ChatGPT's performance on 17 datasets with 14 IE sub-tasks under the zero-shot, few-shot and chain-of-thought scenarios, and find a huge performance gap between ChatGPT and SOTA results. Next, we rethink this gap and propose a soft-matching strategy for evaluation to more accurately reflect ChatGPT's performance. Then, we analyze the robustness of ChatGPT on 14 IE sub-tasks, and find that: 1) ChatGPT rarely outputs invalid responses; 2) Irrelevant context and long-tail target types greatly affect ChatGPT's performance; 3) ChatGPT cannot understand well the subject-object relationships in RE task. Finally, we analyze the errors of ChatGPT, and find that "unannotated spans" is the most dominant error type. This raises concerns about the quality of annotated data, and indicates the possibility of annotating data with ChatGPT. The data and code are released at Github site.

Abductive reasoning aims to find plausible explanations for an event. This style of reasoning is critical for commonsense tasks where there are often multiple plausible explanations. Existing approaches for abductive reasoning in natural language processing (NLP) often rely on manually generated annotations for supervision; however, such annotations can be subjective and biased. Instead of using direct supervision, this work proposes an approach for abductive commonsense reasoning that exploits the fact that only a subset of explanations is correct for a given context. The method uses posterior regularization to enforce a mutual exclusion constraint, encouraging the model to learn the distinction between fluent explanations and plausible ones. We evaluate our approach on a diverse set of abductive reasoning datasets; experimental results show that our approach outperforms or is comparable to directly applying pretrained language models in a zero-shot manner and other knowledge-augmented zero-shot methods.

Large Language Models (LLMs) have exhibited remarkable performance across various natural language processing (NLP) tasks. However, fine-tuning these models often necessitates substantial supervision, which can be expensive and time-consuming to obtain. This paper introduces a novel unsupervised method called LanguageModel Self-Improvement by Reinforcement Learning Contemplation (SIRLC) that improves LLMs without reliance on external labels. Our approach is grounded in the observation that it is simpler for language models to assess text quality than to generate text. Building on this insight, SIRLC assigns LLMs dual roles as both student and teacher. As a student, the LLM generates answers to unlabeled questions, while as a teacher, it evaluates the generated text and assigns scores accordingly. The model parameters are updated using reinforcement learning to maximize the evaluation score. We demonstrate that SIRLC can be applied to various NLP tasks, such as reasoning problems, text generation, and machine translation. Our experiments show that SIRLC effectively improves LLM performance without external supervision, resulting in a 5.6% increase in answering accuracy for reasoning tasks and a rise in BERTScore from 0.82 to 0.86 for translation tasks. Furthermore, SIRLC can be applied to models of different sizes, showcasing its broad applicability.