This paper explores a simple method for improving the zero-shot learning abilities of language models. We show that instruction tuning -- finetuning language models on a collection of tasks described via instructions -- substantially boosts zero-shot performance on unseen tasks. 
We take a 137B parameter pretrained language model and instruction-tune it on over 60 NLP tasks verbalized via natural language instruction templates. We evaluate this instruction-tuned model, which we call FLAN, on unseen task types. FLAN substantially improves the performance of its unmodified counterpart and surpasses zero-shot 175B GPT-3 on 19 of 25 tasks that we evaluate. FLAN even outperforms few-shot GPT-3 by a large margin on ANLI, RTE, BoolQ, AI2-ARC, OpenbookQA, and StoryCloze. Ablation studies reveal that number of tasks and model scale are key components to the success of instruction tuning.

Finetuned Language Models Are Zero-Shot Learners

Although large language models (LLMs) have demonstrated impressive potential on simple tasks, their breadth of scope, lack of transparency, and insufficient controllability can make them less effective when assisting humans on more complex tasks. In response, we introduce the concept of Chaining LLM steps together, where the output of one step becomes the input for the next, thus aggregating the gains per step. We first define a set of LLM primitive operations useful for Chain construction, then present an interactive system where users can modify these Chains, along with their intermediate results, in a modular way. In a 20-person user study, we found that Chaining not only improved the quality of task outcomes, but also significantly enhanced system transparency, controllability, and sense of collaboration. Additionally, we saw that users developed new ways of interacting with LLMs through Chains: they leveraged sub-tasks to calibrate model expectations, compared and contrasted alternative strategies by observing parallel downstream effects, and debugged unexpected model outputs by "unit-testing" sub-components of a Chain. In two case studies, we further explore how LLM Chains may be used in future applications.

AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts

Abstract Large language models (LLMs) have been transformative. They are pretrained foundational models that are self-supervised and can be adapted with fine-tuning to a wide range of natural language tasks, each of which previously would have required a separate network model. This is one step closer to the extraordinary versatility of human language. GPT-3 and, more recently, LaMDA, both of them LLMs, can carry on dialogs with humans on many topics after minimal priming with a few examples. However, there has been a wide range of reactions and debate on whether these LLMs understand what they are saying or exhibit signs of intelligence. This high variance is exhibited in three interviews with LLMs reaching wildly different conclusions. A new possibility was uncovered that could explain this divergence. What appears to be intelligence in LLMs may in fact be a mirror that reflects the intelligence of the interviewer, a remarkable twist that could be considered a reverse Turing test. If so, then by studying interviews, we may be learning more about the intelligence and beliefs of the interviewer than the intelligence of the LLMs. As LLMs become more capable, they may transform the way we interact with machines and how they interact with each other. Increasingly, LLMs are being coupled with sensorimotor devices. LLMs can talk the talk, but can they walk the walk? A road map for achieving artificial general autonomy is outlined with seven major improvements inspired by brain systems and how LLMs could in turn be used to uncover new insights into brain function.

https://direct.mit.edu/neco/article-pdf/35/3/309/2071839/neco_a_01563.pdf

Large Language Models and the Reverse Turing Test

We describe a method to generate 3D architected materials based on mathematically parameterized human readable word input, offering a direct materialization of language. Our method uses a combination of a vector quantized generative adversarial network and contrastive language-image pre-training neural networks to generate images, which are translated into 3D architectures that are then 3D printed using fused deposition modeling into materials with varying rigidity. The novel materials are further analyzed in a metallic realization as an aluminum-based nano-architecture, using molecular dynamics modeling and thereby providing mechanistic insights into the physical behavior of the material under extreme compressive loading. This work offers a novel way to design, understand, and manufacture 3D architected materials designed from mathematically parameterized language input. Our work features, at its core, a generally applicable algorithm that transforms any 2D image data into hierarchical fully tileable, periodic architected materials. This method can have broader applications beyond language-based materials design and can render other avenues for the analysis and manufacturing of architected materials, including microstructure gradients through parametric modeling. As an emerging field, language-based design approaches can have a profound impact on end-to-end design environments and drive a new understanding of physical phenomena that intersect directly with human language and creativity. It may also be used to exploit information mined from diverse and complex databases and data sources.

https://aip.scitation.org/doi/pdf/10.1063/5.0082338

Generative design, manufacturing, and molecular modeling of 3D architected materials based on natural language input

Large pre-trained models such as CLIP offer consistent accuracy across a range of data distributions when performing zero-shot inference (i.e., without fine-tuning on a specific dataset). Although existing fine-tuning approaches substantially improve accuracy in-distribution, they also reduce out-of-distribution robustness. We address this tension by introducing a simple and effective method for improving robustness: ensembling the weights of the zero-shot and fine-tuned models. Compared to standard fine-tuning, the resulting weight-space ensembles provide large accuracy improvements out-of-distribution, while matching or improving in-distribution accuracy. On ImageNet and five derived distribution shifts, weight-space ensembles improve out-of-distribution accuracy by 2 to 10 percentage points while increasing in-distribution accuracy by nearly 1 percentage point relative to standard fine-tuning. These improvements come at no additional computational cost during fine-tuning or inference.

/pdf/robust-fine-tuning-of-zero-shot-models-e9bi3z5o0t.pdf

Robust fine-tuning of zero-shot models

AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.

On the Opportunities and Risks of Foundation Models.

1 THE MODEL IS THE MESSAGE A large language model is not simply a tool for text generation—it is a new medium for content, shaping textual artifacts through algorithms and data. Marshall McLuhan famously introduced the idea that “the medium is the message:” the form of content instigates or accelerates cultural shifts [15]. McLuhan presented TV as a medium that revolutionized the political system in the West by creating a new type of national leader who was more of a “showman” than a politician [16]. In this way, LLMs are a new medium that support the creation of machine-generated text. To understand the influence of the form, we need to first understand the process by which text generation occurs. LLMs and, by proxy, the text they generate, are a direct product of data (collection, curation, annotation, etc.), algorithms (encoding and decoding), and computational resources [8, 11, 17]. These components mold different political and technical aspects of the medium from scale to diversity of output to accessibility. Of particular concern to us is how the centralization of LLM development—a product of economic forces around data and compute—shapes writers’ and readers’ control over and experiences of language production and consumption. In the following section, we present historical examples of language control, abuse, and change that situate LLMs within a history of other media mediums and provide a foundation for understanding how LLMs can have deep influence over language.

Isabelle Levent

Papers

On the Opportunities and Risks of Foundation Models.

The Model is the Message