The recent progress in large language models (LLMs), especially the invention of chain-of-thought prompting, has made it possible to automatically answer questions by stepwise reasoning. However, when faced with more complicated problems that require non-linear thinking, even the strongest LLMs make mistakes. To address this, we explore whether LLMs are able to recognize errors in their own step-bystep reasoning, without resorting to external resources. To this end, we propose SelfCheck, a general-purpose zero-shot verification schema for recognizing such errors. We then use the results of these checks to improve question-answering performance by conducting weighted voting on multiple solutions to the question. We test SelfCheck on three datasets--GSM8K, MathQA, and MATH--and find that it successfully recognizes errors and, in turn, increases final answer accuracies. Recent years have witnessed dramatic changes in the areas of NLP and AI brought on by significant advances in LLMs. From GPT-3 (Brown et al., 2020), PaLM (Chowdhery et al., 2022), Llama (Touvron et al., 2023) and Falcon (Almazrouei et al., 2023) to GPT-4 (OpenAI, 2023) and PaLM-2 (Google, 2023), the increasing model sizes and exploding amount of training data have empowered LLMs to achieve human-level performance on a large range of tasks, including summarization, translation, and question answering. The invention of Chain-of-Thought prompting (CoT, Wei et al. (2022)) has further enhanced LLMs' ability to solve complex problems by generating step-by-step solutions. However, the performance of even the largest LLMs is still unsatisfactory on more difficult reasoning problems. For example, GPT-4 with CoT prompting only correctly answers 42.5% of problems in the MATH dataset (Bubeck et al., 2023; Hendrycks et al., 2021), which is far below human level. Such problems require careful and extensive multi-step reasoning to solve, and LLMs are consequently prone to make mistakes: even though their error rate on individual steps may be low, the probability of generating at least one erroneous step can still be quite high, undermining the final answer. Recent works have tried to overcome this limitation by checking for errors in these step-by-step solutions (Cobbe et al., 2021; Li et al., 2022; Ling et al., 2023).

Task planning systems have been developed to help robots use human knowledge (about actions) to complete long-horizon tasks. Most of them have been developed for "closed worlds" while assuming the robot is provided with complete world knowledge. However, the real world is generally open, and the robots frequently encounter unforeseen situations that can potentially break the planner's completeness. Could we leverage the recent advances on pre-trained Large Language Models (LLMs) to enable classical planning systems to deal with novel situations? This paper introduces a novel framework, called COWP, for open-world task planning and situation handling. COWP dynamically augments the robot's action knowledge, including the preconditions and effects of actions, with task-oriented commonsense knowledge. COWP embraces the openness from LLMs, and is grounded to specific domains via action knowledge. For systematic evaluations, we collected a dataset that includes 1,085 execution-time situations. Each situation corresponds to a state instance wherein a robot is potentially unable to complete a task using a solution that normally works. Experimental results show that our approach outperforms competitive baselines from the literature in the success rate of service tasks. Additionally, we have demonstrated COWP using a mobile manipulator. Supplementary materials are available at: https://cowplanning.github.io/

Recent research has highlighted the importance of dataset size in scaling language models. However, large language models (LLMs) are notoriously token-hungry during pre-training, and high-quality text data on the web is likely to be approaching its scaling limit for LLMs. To further enhance LLMs, a straightforward approach is to repeat the pre-training data for additional epochs. In this study, we empirically investigate three key aspects under this approach. First, we explore the consequences of repeating pre-training data, revealing that the model is susceptible to overfitting, leading to multi-epoch degradation. Second, we examine the key factors contributing to multi-epoch degradation, finding that significant factors include dataset size, model parameters, and training objectives, while less influential factors consist of dataset quality and model FLOPs. Finally, we explore whether widely used regularization can alleviate multi-epoch degradation. Most regularization techniques do not yield significant improvements, except for dropout, which demonstrates remarkable effectiveness but requires careful tuning when scaling up the model size. Additionally, we discover that leveraging mixture-of-experts (MoE) enables cost-effective and efficient hyper-parameter tuning for computationally intensive dense LLMs with comparable trainable parameters, potentially impacting efficient LLM development on a broader scale.

Large language models (LLMs) have initiated a paradigm shift in transfer learning. In contrast to the classic pretraining-then-finetuning procedure, in order to use LLMs for downstream prediction tasks, one only needs to provide a few demonstrations, known as in-context examples, without adding more or updating existing model parameters. This in-context learning (ICL) capability of LLMs is intriguing, and it is not yet fully understood how pretrained LLMs acquire such capabilities. In this paper, we investigate the reason why a transformer-based language model can accomplish in-context learning after pre-training on a general language corpus by proposing one hypothesis that LLMs can simulate kernel regression with internal representations when faced with in-context examples. More concretely, we first prove that Bayesian inference on in-context prompts can be asymptotically understood as kernel regression $\hat y = \sum_i y_i K(x, x_i)/\sum_i K(x, x_i)$ as the number of in-context demonstrations grows. Then, we empirically investigate the in-context behaviors of language models. We find that during ICL, the attention and hidden features in LLMs match the behaviors of a kernel regression. Finally, our theory provides insights into multiple phenomena observed in the ICL field: why retrieving demonstrative samples similar to test samples can help, why ICL performance is sensitive to the output formats, and why ICL accuracy benefits from selecting in-distribution and representative samples.

Tabular data prediction has been employed in medical applications such as patient health risk prediction. However, existing methods usually revolve around the algorithm design while overlooking the significance of data engineering. As such, previous predictors are often trained on manually curated small datasets that struggle to generalize across different tabular datasets during inference. This paper proposes to scale medical tabular data predictors (MediTab) to various tabular inputs with varying features. The method uses a data engine that leverages large language models (LLMs) to consolidate tabular samples to overcome the barrier across tables with distinct schema. It also aligns out-domain data with the target task using a "learn, annotate, and refinement" pipeline. The expanded training data then enables the pre-trained MediTab to infer for arbitrary tabular input in the domain without fine-tuning, resulting in significant improvements over supervised baselines: it reaches an average ranking of 1.57 and 1.00 on 7 patient outcome prediction datasets and 3 trial outcome prediction datasets, respectively. In addition, MediTab exhibits impressive zero-shot performances: it outperforms supervised XGBoost models by 8.9% and 17.2% on average in two prediction tasks, respectively. Tabular data are structured as tables or spreadsheets in a relational database. Each row in the table represents a data sample, while columns represent various feature variables of different types, including categorical, numerical, binary, and textual features. Most previous papers focused on the model design of tabular predictors, mainly by (1) augmenting feature interactions via neural networks (Arik & Pfister, 2021), (2) improving tabular data representation learning by self-supervised pre-training (Yin et al., 2020; Yoon et al., 2020; Bahri et al., 2022), and (3) performing cross-tabular pre-training for transfer learning (Wang & Sun, 2022b; Zhu et al., 2023). Tabular data predictor was also employed in medicine, such as patient health risk prediction (Wang & Sun, 2022b) and clinical trial outcome prediction (Fu et al., 2022). Additionally, LLMs have been shown to be able to sample synthetic and yet highly realistic tabular data as well Borisov et al. (2022); Theodorou et al. (2023).

The way users acquire information is undergoing a paradigm shift with the advent of ChatGPT. Unlike conventional search engines, ChatGPT retrieves knowledge from the model itself and generates answers for users. ChatGPT's impressive question-answering (QA) capability has attracted more than 100 million users within a short period of time but has also raised concerns regarding its reliability. In this paper, we perform the first large-scale measurement of ChatGPT's reliability in the generic QA scenario with a carefully curated set of 5,695 questions across ten datasets and eight domains. We find that ChatGPT's reliability varies across different domains, especially underperforming in law and science questions. We also demonstrate that system roles, originally designed by OpenAI to allow users to steer ChatGPT's behavior, can impact ChatGPT's reliability in an imperceptible way. We further show that ChatGPT is vulnerable to adversarial examples, and even a single character change can negatively affect its reliability in certain cases. We believe that our study provides valuable insights into ChatGPT's reliability and underscores the need for strengthening the reliability and security of large language models (LLMs).

Large language models (LLMs) have achieved impressive performance on code generation. However, for complex programming tasks, generating the correct solution in one go becomes challenging, thus some prior works have designed program repair approaches to improve code generation performance. In this work, we propose Self-Debugging, which teaches a large language model to debug its predicted program via few-shot demonstrations. In particular, we demonstrate that Self-Debugging can teach the large language model to perform rubber duck debugging; i.e., without any human feedback on the code correctness or error messages, the model is able to identify its mistakes by investigating the execution results and explaining the generated code in natural language. Self-Debugging achieves the state-of-the-art performance on several code generation benchmarks, including the Spider dataset for text-to-SQL generation, TransCoder for C++-to-Python translation, and MBPP for text-to-Python generation. On the Spider benchmark where there are no unit tests to verify the correctness of predictions, Self-Debugging with code explanation consistently improves the baseline by 2-3%, and improves the prediction accuracy on problems of the hardest level by 9%. On TransCoder and MBPP where unit tests are available, Self-Debugging improves the baseline accuracy by up to 12%. Meanwhile, by leveraging feedback messages and reusing failed predictions, Self-Debugging notably improves sample efficiency, and can match or outperform baseline models that generate more than 10x candidate programs.

SILICON Valley's feverish embrace of large language models (LLMs) shows no sign of letting up. Google is integrating its chatbot Bard into every one of its services, while OpenAI is imbuing its own offering, ChatGPT, with new senses, such as the ability to "see" and "speak", envisaging a new kind of personal assistant. But deep mysteries remain about how these tools function: what is really going on behind their shiny interfaces, which tasks are they truly good at and how might they fail? Should we really be betting the house on technology with so many unknowns?

Last month, I found myself in a particular seat. A few places to my left was Elon Musk. Down the table to my right sat Bill Gates. Across the room sat Satya Nadella, Microsoft's CEO, and not too far to his left was Eric Schmidt, the former CEO of Google. At the other end of the table sat Sam Altman, the head of OpenAI, the company responsible for ChatGPT.

Big Tech companies have been rushing to design and produce new "generative" AI products since OpenAI unveiled ChatGPT last November. But the question of how the companies would get people to use -- and pay for -- the expensive new technology has swirled around the industry for months. Google, Amazon and Apple all have millions of customers already speaking to the companies' existing voice assistants to set alarms, check the weather and make notes for them, representing a ready-made group of consumers to test out the new AI chatbots on.