Meta-Prompting: Enhancing Language Models with Task-Agnostic Scaffolding

We introduce meta-prompting, an effective scaffolding technique designed to enhance the functionality of language models (LMs). This approach transforms a single LM into a multi-faceted conductor, adept at managing and integrating multiple independent LM queries. By employing high-level instructions, meta-prompting guides the LM to break down complex tasks into smaller, more manageable subtasks. These subtasks are then handled by distinct "expert" instances of the same LM, each operating under specific, tailored instructions. Central to this process is the LM itself, in its role as the conductor, which ensures seamless communication and effective integration of the outputs from these expert models. It additionally employs its inherent critical thinking and robust verification processes to refine and authenticate the end result. This collaborative prompting approach empowers a single LM to simultaneously act as a comprehensive orchestrator and a panel of diverse experts, significantly enhancing its performance across a wide array of tasks. The zero-shot, task-agnostic nature of meta-prompting greatly simplifies user interaction by obviating the need for detailed, task-specific instructions. Furthermore, our research demonstrates the seamless integration of external tools, such as a Python interpreter, into the meta-prompting framework, thereby broadening its applicability and utility. Through rigorous experimentation with GPT-4, we establish the superiority of meta-prompting over conventional scaffolding methods: When averaged across all tasks, including the Game of 24, Checkmate-in-One, and Python Programming Puzzles, meta-prompting, augmented with a Python interpreter functionality, surpasses standard prompting by 17.1%, expert (dynamic) prompting by 17.3%, and multipersona prompting by 15.2%.

Calibrated Language Models Must Hallucinate

Recent language models generate false but plausible-sounding text with surprising frequency. Such "hallucinations" are an obstacle to the usability of language-based AI systems and can harm people who rely upon their outputs. This work shows shows that there is an inherent statistical lower-bound on the rate that pretrained language models hallucinate certain types of facts, having nothing to do with the transformer LM architecture or data quality. For "arbitrary" facts whose veracity cannot be determined from the training data, we show that hallucinations must occur at a certain rate for language models that satisfy a statistical calibration condition appropriate for generative language models. Specifically, if the maximum probability of any fact is bounded, we show that the probability of generating a hallucination is close to the fraction of facts that occur exactly once in the training data (a "Good-Turing" estimate), even assuming ideal training data without errors. One conclusion is that models pretrained to be sufficiently good predictors (i.e., calibrated) may require post-training to mitigate hallucinations on the type of arbitrary facts that tend to appear once in the training set. However, our analysis also suggests that there is no statistical reason that pretraining will lead to hallucination on facts that tend to appear more than once in the training data (like references to publications such as articles and books, whose hallucinations have been particularly notable and problematic) or on systematic facts (like arithmetic calculations). Therefore, different architectures and learning algorithms may mitigate these latter types of hallucinations.

Testing Language Model Agents Safely in the Wild

A prerequisite for safe autonomy-in-the-wild is safe testing-in-the-wild. Yet real-world autonomous tests face several unique safety challenges, both due to the possibility of causing harm during a test, as well as the risk of encountering new unsafe agent behavior through interactions with real-world and potentially malicious actors. We propose a framework for conducting safe autonomous agent tests on the open internet: agent actions are audited by a context-sensitive monitor that enforces a stringent safety boundary to stop an unsafe test, with suspect behavior ranked and logged to be examined by humans. We design a basic safety monitor (AgentMonitor) that is flexible enough to monitor existing LLM agents, and, using an adversarial simulated agent, we measure its ability to identify and stop unsafe situations. Then we apply the AgentMonitor on a battery of real-world tests of AutoGPT, and we identify several limitations and challenges that will face the creation of safe in-the-wild tests as autonomous agents grow more capable.

Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation

Several recent advances in AI systems (e.g., Tree-of-Thoughts and Program-Aided Language Models) solve problems by providing a "scaffolding" program that structures multiple calls to language models to generate better outputs. A scaffolding program is written in a programming language such as Python. In this work, we use a language-model-infused scaffolding program to improve itself. We start with a seed "improver" that improves an input program according to a given utility function by querying a language model several times and returning the best solution. We then run this seed improver to improve itself. Across a small set of downstream tasks, the resulting improved improver generates programs with significantly better performance than its seed improver. Afterward, we analyze the variety of self-improvement strategies proposed by the language model, including beam search, genetic algorithms, and simulated annealing. Since the language models themselves are not altered, this is not full recursive self-improvement. Nonetheless, it demonstrates that a modern language model, GPT-4 in our proof-of-concept experiments, is capable of writing code that can call itself to improve itself. We critically consider concerns around the development of self-improving technologies and evaluate the frequency with which the generated code bypasses a sandbox.

Textbooks Are All You Need

We introduce phi-1, a new large language model for code, with significantly smaller size than competing models: phi-1 is a Transformer-based model with 1.3B parameters, trained for 4 days on 8 A100s, using a selection of ``textbook quality" data from the web (6B tokens) and synthetically generated textbooks and exercises with GPT-3.5 (1B tokens). Despite this small scale, phi-1 attains pass@1 accuracy 50.6% on HumanEval and 55.5% on MBPP. It also displays surprising emergent properties compared to phi-1-base, our model before our finetuning stage on a dataset of coding exercises, and phi-1-small, a smaller model with 350M parameters trained with the same pipeline as phi-1 that still achieves 45% on HumanEval.

Do Language Models Know When They're Hallucinating References?

Current state-of-the-art language models (LMs) are notorious for generating text with "hallucinations," a primary example being book and paper references that lack any solid basis in their training data. However, we find that many of these fabrications can be identified using the same LM, using only black-box queries without consulting any external resources. Consistency checks done with direct queries about whether the generated reference title is real (inspired by Kadavath et al. 2022, Lin et al. 2022, Manakul et al. 2023) are compared to consistency checks with indirect queries which ask for ancillary details such as the authors of the work. These consistency checks are found to be partially reliable indicators of whether or not the reference is a hallucination. In particular, we find that LMs in the GPT-series will hallucinate differing authors of hallucinated references when queried in independent sessions, while it will consistently identify authors of real references. This suggests that the hallucination may be more a result of generation techniques than the underlying representation.

Loss minimization yields multicalibration for large neural networks

Multicalibration is a notion of fairness that aims to provide accurate predictions across a large set of groups. Multicalibration is known to be a different goal than loss minimization, even for simple predictors such as linear functions. In this note, we show that for (almost all) large neural network sizes, optimally minimizing squared error leads to multicalibration. Our results are about representational aspects of neural networks, and not about algorithmic or sample complexity considerations. Previous such results were known only for predictors that were nearly Bayes-optimal and were therefore representation independent. We emphasize that our results do not apply to specific algorithms for optimizing neural networks, such as SGD, and they should not be interpreted as "fairness comes for free from optimizing neural networks".

A Theory of Unsupervised Translation Motivated by Understanding Animal Communication

Recent years have seen breakthroughs in neural language models that capture nuances of language, culture, and knowledge. Neural networks are capable of translating between languages -- in some cases even between two languages where there is little or no access to parallel translations, in what is known as Unsupervised Machine Translation (UMT). Given this progress, it is intriguing to ask whether machine learning tools can ultimately enable understanding animal communication, particularly that of highly intelligent animals. Our work is motivated by an ambitious interdisciplinary initiative, Project CETI, which is collecting a large corpus of sperm whale communications for machine analysis. We propose a theoretical framework for analyzing UMT when no parallel data are available and when it cannot be assumed that the source and target corpora address related subject domains or posses similar linguistic structure. The framework requires access to a prior probability distribution that should assign non-zero probability to possible translations. We instantiate our framework with two models of language. Our analysis suggests that accuracy of translation depends on the complexity of the source language and the amount of ``common ground'' between the source language and target prior. We also prove upper bounds on the amount of data required from the source language in the unsupervised setting as a function of the amount of data required in a hypothetical supervised setting. Surprisingly, our bounds suggest that the amount of source data required for unsupervised translation is comparable to the supervised setting. For one of the language models which we analyze we also prove a nearly matching lower bound. Our analysis is purely information-theoretic and as such can inform how much source data needs to be collected, but does not yield a computationally efficient procedure.

Recurrent Convolutional Neural Networks Learn Succinct Learning Algorithms

Neural Networks (NNs) struggle to efficiently learn certain problems, such as parity problems, even when there are simple learning algorithms for those problems. Can NNs discover learning algorithms on their own? We exhibit a NN architecture that, in polynomial time, learns as well as any efficient learning algorithm describable by a constant-sized learning algorithm. For example, on parity problems, the NN learns as well as row reduction, an efficient algorithm that can be succinctly described. Our architecture combines both recurrent weight-sharing between layers and convolutional weight-sharing to reduce the number of parameters down to a constant, even though the network itself may have trillions of nodes. While in practice the constants in our analysis are too large to be directly meaningful, our work suggests that the synergy of Recurrent and Convolutional NNs (RCNNs) may be more powerful than either alone.

Using Large Language Models to Simulate Multiple Humans

We propose a method for using a large language model, such as GPT-3, to simulate responses of different humans in a given context. We test our method by attempting to reproduce well-established economic, psycholinguistic, and social experiments. The method requires prompt templates for each experiment. Simulations are run by varying the (hypothetical) subject details, such as name, and analyzing the text generated by the language model. To validate our methodology, we use GPT-3 to simulate the Ultimatum Game, garden path sentences, risk aversion, and the Milgram Shock experiments. In order to address concerns of exposure to these studies in training data, we also evaluate simulations on novel variants of these studies. We show that it is possible to simulate responses of different people and that their responses are consistent with prior human studies from the literature. Across all studies, the distributions generated by larger language models better align with prior experimental results, suggesting a trend that future language models may be used for even more faithful simulations of human responses. Our use of a language model for simulation is contrasted with anthropomorphic views of a language model as having its own behavior.