Jacob Devlin

Hyung Won Chung, Le Hou, Shayne Longpre et al. · cmu, deepmind

Finetuning language models on a collection of datasets phrased as instructions has been shown to improve model performance and generalization to unseen tasks. In this paper we explore instruction finetuning with a particular focus on (1) scaling the number of tasks, (2) scaling the model size, and (3) finetuning on chain-of-thought data. We find that instruction finetuning with the above aspects dramatically improves performance on a variety of model classes (PaLM, T5, U-PaLM), prompting setups (zero-shot, few-shot, CoT), and evaluation benchmarks (MMLU, BBH, TyDiQA, MGSM, open-ended generation). For instance, Flan-PaLM 540B instruction-finetuned on 1.8K tasks outperforms PALM 540B by a large margin (+9.4% on average). Flan-PaLM 540B achieves state-of-the-art performance on several benchmarks, such as 75.2% on five-shot MMLU. We also publicly release Flan-T5 checkpoints, which achieve strong few-shot performance even compared to much larger models, such as PaLM 62B. Overall, instruction finetuning is a general method for improving the performance and usability of pretrained language models.

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin et al. · deepmind, stanford

Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning, which drastically reduces the number of task-specific training examples needed to adapt the model to a particular application. To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model PaLM. We trained PaLM on 6144 TPU v4 chips using Pathways, a new ML system which enables highly efficient training across multiple TPU Pods. We demonstrate continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. On a number of these tasks, PaLM 540B achieves breakthrough performance, outperforming the finetuned state-of-the-art on a suite of multi-step reasoning tasks, and outperforming average human performance on the recently released BIG-bench benchmark. A significant number of BIG-bench tasks showed discontinuous improvements from model scale, meaning that performance steeply increased as we scaled to our largest model. PaLM also has strong capabilities in multilingual tasks and source code generation, which we demonstrate on a wide array of benchmarks. We additionally provide a comprehensive analysis on bias and toxicity, and study the extent of training data memorization with respect to model scale. Finally, we discuss the ethical considerations related to large language models and discuss potential mitigation strategies.

Reiner Pope, Sholto Douglas, Aakanksha Chowdhery et al.

We study the problem of efficient generative inference for Transformer models, in one of its most challenging settings: large deep models, with tight latency targets and long sequence lengths. Better understanding of the engineering tradeoffs for inference for large Transformer-based models is important as use cases of these models are growing rapidly throughout application areas. We develop a simple analytical model for inference efficiency to select the best multi-dimensional partitioning techniques optimized for TPU v4 slices based on the application requirements. We combine these with a suite of low-level optimizations to achieve a new Pareto frontier on the latency and model FLOPS utilization (MFU) tradeoffs on 500B+ parameter models that outperforms the FasterTransformer suite of benchmarks. We further show that with appropriate partitioning, the lower memory requirements of multiquery attention (i.e. multiple query heads share single key/value head) enables scaling up to 32x larger context lengths. Finally, we achieve a low-batch-size latency of 29ms per token during generation (using int8 weight quantization) and a 76% MFU during large-batch-size processing of input tokens, while supporting a long 2048-token context length on the PaLM 540B parameter model.

Zifeng Wang, Zizhao Zhang, Jacob Devlin et al.

Zero-shot transfer learning for document understanding is a crucial yet under-investigated scenario to help reduce the high cost involved in annotating document entities. We present a novel query-based framework, QueryForm, that extracts entity values from form-like documents in a zero-shot fashion. QueryForm contains a dual prompting mechanism that composes both the document schema and a specific entity type into a query, which is used to prompt a Transformer model to perform a single entity extraction task. Furthermore, we propose to leverage large-scale query-entity pairs generated from form-like webpages with weak HTML annotations to pre-train QueryForm. By unifying pre-training and fine-tuning into the same query-based framework, QueryForm enables models to learn from structured documents containing various entities and layouts, leading to better generalization to target document types without the need for target-specific training data. QueryForm sets new state-of-the-art average F1 score on both the XFUND (+4.6%~10.1%) and the Payment (+3.2%~9.5%) zero-shot benchmark, with a smaller model size and no additional image input.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Lajanugen Logeswaran, Ming-Wei Chang, Kenton Lee et al.

We present the zero-shot entity linking task, where mentions must be linked to unseen entities without in-domain labeled data. The goal is to enable robust transfer to highly specialized domains, and so no metadata or alias tables are assumed. In this setting, entities are only identified by text descriptions, and models must rely strictly on language understanding to resolve the new entities. First, we show that strong reading comprehension models pre-trained on large unlabeled data can be used to generalize to unseen entities. Second, we propose a simple and effective adaptive pre-training strategy, which we term domain-adaptive pre-training (DAP), to address the domain shift problem associated with linking unseen entities in a new domain. We present experiments on a new dataset that we construct for this task and show that DAP improves over strong pre-training baselines, including BERT. The data and code are available at https://github.com/lajanugen/zeshel.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Rohan Anil, Andrew M. Dai, Orhan Firat et al.

We introduce PaLM 2, a new state-of-the-art language model that has better multilingual and reasoning capabilities and is more compute-efficient than its predecessor PaLM. PaLM 2 is a Transformer-based model trained using a mixture of objectives. Through extensive evaluations on English and multilingual language, and reasoning tasks, we demonstrate that PaLM 2 has significantly improved quality on downstream tasks across different model sizes, while simultaneously exhibiting faster and more efficient inference compared to PaLM. This improved efficiency enables broader deployment while also allowing the model to respond faster, for a more natural pace of interaction. PaLM 2 demonstrates robust reasoning capabilities exemplified by large improvements over PaLM on BIG-Bench and other reasoning tasks. PaLM 2 exhibits stable performance on a suite of responsible AI evaluations, and enables inference-time control over toxicity without additional overhead or impact on other capabilities. Overall, PaLM 2 achieves state-of-the-art performance across a diverse set of tasks and capabilities. When discussing the PaLM 2 family, it is important to distinguish between pre-trained models (of various sizes), fine-tuned variants of these models, and the user-facing products that use these models. In particular, user-facing products typically include additional pre- and post-processing steps. Additionally, the underlying models may evolve over time. Therefore, one should not expect the performance of user-facing products to exactly match the results reported in this report.

Chris Alberti, Daniel Andor, Emily Pitler et al.

We introduce a novel method of generating synthetic question answering corpora by combining models of question generation and answer extraction, and by filtering the results to ensure roundtrip consistency. By pretraining on the resulting corpora we obtain significant improvements on SQuAD2 and NQ, establishing a new state-of-the-art on the latter. Our synthetic data generation models, for both question generation and answer extraction, can be fully reproduced by finetuning a publicly available BERT model on the extractive subsets of SQuAD2 and NQ. We also describe a more powerful variant that does full sequence-to-sequence pretraining for question generation, obtaining exact match and F1 at less than 0.1% and 0.4% from human performance on SQuAD2.

Ming-Wei Chang, Kristina Toutanova, Kenton Lee et al.

Hierarchical neural architectures are often used to capture long-distance dependencies and have been applied to many document-level tasks such as summarization, document segmentation, and sentiment analysis. However, effective usage of such a large context can be difficult to learn, especially in the case where there is limited labeled data available. Building on the recent success of language model pretraining methods for learning flat representations of text, we propose algorithms for pre-training hierarchical document representations from unlabeled data. Unlike prior work, which has focused on pre-training contextual token representations or context-independent {sentence/paragraph} representations, our hierarchical document representations include fixed-length sentence/paragraph representations which integrate contextual information from the entire documents. Experiments on document segmentation, document-level question answering, and extractive document summarization demonstrate the effectiveness of the proposed pre-training algorithms.

Jacob Devlin, Ming-Wei Chang, Kenton Lee et al.

We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).

Rudy Bunel, Matthew Hausknecht, Jacob Devlin et al.

Program synthesis is the task of automatically generating a program consistent with a specification. Recent years have seen proposal of a number of neural approaches for program synthesis, many of which adopt a sequence generation paradigm similar to neural machine translation, in which sequence-to-sequence models are trained to maximize the likelihood of known reference programs. While achieving impressive results, this strategy has two key limitations. First, it ignores Program Aliasing: the fact that many different programs may satisfy a given specification (especially with incomplete specifications such as a few input-output examples). By maximizing the likelihood of only a single reference program, it penalizes many semantically correct programs, which can adversely affect the synthesizer performance. Second, this strategy overlooks the fact that programs have a strict syntax that can be efficiently checked. To address the first limitation, we perform reinforcement learning on top of a supervised model with an objective that explicitly maximizes the likelihood of generating semantically correct programs. For addressing the second limitation, we introduce a training procedure that directly maximizes the probability of generating syntactically correct programs that fulfill the specification. We show that our contributions lead to improved accuracy of the models, especially in cases where the training data is limited.

Jiatao Gu, Hany Hassan, Jacob Devlin et al.

In this paper, we propose a new universal machine translation approach focusing on languages with a limited amount of parallel data. Our proposed approach utilizes a transfer-learning approach to share lexical and sentence level representations across multiple source languages into one target language. The lexical part is shared through a Universal Lexical Representation to support multilingual word-level sharing. The sentence-level sharing is represented by a model of experts from all source languages that share the source encoders with all other languages. This enables the low-resource language to utilize the lexical and sentence representations of the higher resource languages. Our approach is able to achieve 23 BLEU on Romanian-English WMT2016 using a tiny parallel corpus of 6k sentences, compared to the 18 BLEU of strong baseline system which uses multilingual training and back-translation. Furthermore, we show that the proposed approach can achieve almost 20 BLEU on the same dataset through fine-tuning a pre-trained multi-lingual system in a zero-shot setting.

Jacob Devlin, Jonathan Uesato, Rishabh Singh et al.

We study the problem of semantic code repair, which can be broadly defined as automatically fixing non-syntactic bugs in source code. The majority of past work in semantic code repair assumed access to unit tests against which candidate repairs could be validated. In contrast, the goal here is to develop a strong statistical model to accurately predict both bug locations and exact fixes without access to information about the intended correct behavior of the program. Achieving such a goal requires a robust contextual repair model, which we train on a large corpus of real-world source code that has been augmented with synthetically injected bugs. Our framework adopts a two-stage approach where first a large set of repair candidates are generated by rule-based processors, and then these candidates are scored by a statistical model using a novel neural network architecture which we refer to as Share, Specialize, and Compete. Specifically, the architecture (1) generates a shared encoding of the source code using an RNN over the abstract syntax tree, (2) scores each candidate repair using specialized network modules, and (3) then normalizes these scores together so they can compete against one another in comparable probability space. We evaluate our model on a real-world test set gathered from GitHub containing four common categories of bugs. Our model is able to predict the exact correct repair 41\% of the time with a single guess, compared to 13\% accuracy for an attentional sequence-to-sequence model.

Jacob Devlin, Rudy Bunel, Rishabh Singh et al.

Most recently proposed methods for Neural Program Induction work under the assumption of having a large set of input/output (I/O) examples for learning any underlying input-output mapping. This paper aims to address the problem of data and computation efficiency of program induction by leveraging information from related tasks. Specifically, we propose two approaches for cross-task knowledge transfer to improve program induction in limited-data scenarios. In our first proposal, portfolio adaptation, a set of induction models is pretrained on a set of related tasks, and the best model is adapted towards the new task using transfer learning. In our second approach, meta program induction, a $k$-shot learning approach is used to make a model generalize to new tasks without additional training. To test the efficacy of our methods, we constructed a new benchmark of programs written in the Karel programming language. Using an extensive experimental evaluation on the Karel benchmark, we demonstrate that our proposals dramatically outperform the baseline induction method that does not use knowledge transfer. We also analyze the relative performance of the two approaches and study conditions in which they perform best. In particular, meta induction outperforms all existing approaches under extreme data sparsity (when a very small number of examples are available), i.e., fewer than ten. As the number of available I/O examples increase (i.e. a thousand or more), portfolio adapted program induction becomes the best approach. For intermediate data sizes, we demonstrate that the combined method of adapted meta program induction has the strongest performance.

Jacob Devlin

Attentional sequence-to-sequence models have become the new standard for machine translation, but one challenge of such models is a significant increase in training and decoding cost compared to phrase-based systems. Here, we focus on efficient decoding, with a goal of achieving accuracy close the state-of-the-art in neural machine translation (NMT), while achieving CPU decoding speed/throughput close to that of a phrasal decoder. We approach this problem from two angles: First, we describe several techniques for speeding up an NMT beam search decoder, which obtain a 4.4x speedup over a very efficient baseline decoder without changing the decoder output. Second, we propose a simple but powerful network architecture which uses an RNN (GRU/LSTM) layer at bottom, followed by a series of stacked fully-connected layers applied at every timestep. This architecture achieves similar accuracy to a deep recurrent model, at a small fraction of the training and decoding cost. By combining these techniques, our best system achieves a very competitive accuracy of 38.3 BLEU on WMT English-French NewsTest2014, while decoding at 100 words/sec on single-threaded CPU. We believe this is the best published accuracy/speed trade-off of an NMT system.

Jacob Devlin, Jonathan Uesato, Surya Bhupatiraju et al.

The problem of automatically generating a computer program from some specification has been studied since the early days of AI. Recently, two competing approaches for automatic program learning have received significant attention: (1) neural program synthesis, where a neural network is conditioned on input/output (I/O) examples and learns to generate a program, and (2) neural program induction, where a neural network generates new outputs directly using a latent program representation. Here, for the first time, we directly compare both approaches on a large-scale, real-world learning task. We additionally contrast to rule-based program synthesis, which uses hand-crafted semantics to guide the program generation. Our neural models use a modified attention RNN to allow encoding of variable-sized sets of I/O pairs. Our best synthesis model achieves 92% accuracy on a real-world test set, compared to the 34% accuracy of the previous best neural synthesis approach. The synthesis model also outperforms a comparable induction model on this task, but we more importantly demonstrate that the strength of each approach is highly dependent on the evaluation metric and end-user application. Finally, we show that we can train our neural models to remain very robust to the type of noise expected in real-world data (e.g., typos), while a highly-engineered rule-based system fails entirely.

Ting-Hao, Huang, Francis Ferraro et al.

We introduce the first dataset for sequential vision-to-language, and explore how this data may be used for the task of visual storytelling. The first release of this dataset, SIND v.1, includes 81,743 unique photos in 20,211 sequences, aligned to both descriptive (caption) and story language. We establish several strong baselines for the storytelling task, and motivate an automatic metric to benchmark progress. Modelling concrete description as well as figurative and social language, as provided in this dataset and the storytelling task, has the potential to move artificial intelligence from basic understandings of typical visual scenes towards more and more human-like understanding of grounded event structure and subjective expression.

Nasrin Mostafazadeh, Ishan Misra, Jacob Devlin et al.

There has been an explosion of work in the vision & language community during the past few years from image captioning to video transcription, and answering questions about images. These tasks have focused on literal descriptions of the image. To move beyond the literal, we choose to explore how questions about an image are often directed at commonsense inference and the abstract events evoked by objects in the image. In this paper, we introduce the novel task of Visual Question Generation (VQG), where the system is tasked with asking a natural and engaging question when shown an image. We provide three datasets which cover a variety of images from object-centric to event-centric, with considerably more abstract training data than provided to state-of-the-art captioning systems thus far. We train and test several generative and retrieval models to tackle the task of VQG. Evaluation results show that while such models ask reasonable questions for a variety of images, there is still a wide gap with human performance which motivates further work on connecting images with commonsense knowledge and pragmatics. Our proposed task offers a new challenge to the community which we hope furthers interest in exploring deeper connections between vision & language.

Junyoung Chung, Jacob Devlin, Hany Hassan Awadalla

In this paper, we explore different neural network architectures that can predict if a speaker of a given utterance is asking a question or making a statement. We com- pare the outcomes of regularization methods that are popularly used to train deep neural networks and study how different context functions can affect the classification performance. We also compare the efficacy of gated activation functions that are favorably used in recurrent neural networks and study how to combine multimodal inputs. We evaluate our models on two multimodal datasets: MSR-Skype and CALLHOME.

Francis Ferraro, Nasrin Mostafazadeh, Ting-Hao et al.

Integrating vision and language has long been a dream in work on artificial intelligence (AI). In the past two years, we have witnessed an explosion of work that brings together vision and language from images to videos and beyond. The available corpora have played a crucial role in advancing this area of research. In this paper, we propose a set of quality metrics for evaluating and analyzing the vision & language datasets and categorize them accordingly. Our analyses show that the most recent datasets have been using more complex language and more abstract concepts, however, there are different strengths and weaknesses in each.

Hendra Setiawan, Zhongqiang Huang, Jacob Devlin et al.

We present a three-pronged approach to improving Statistical Machine Translation (SMT), building on recent success in the application of neural networks to SMT. First, we propose new features based on neural networks to model various non-local translation phenomena. Second, we augment the architecture of the neural network with tensor layers that capture important higher-order interaction among the network units. Third, we apply multitask learning to estimate the neural network parameters jointly. Each of our proposed methods results in significant improvements that are complementary. The overall improvement is +2.7 and +1.8 BLEU points for Arabic-English and Chinese-English translation over a state-of-the-art system that already includes neural network features.

Jacob Devlin, Saurabh Gupta, Ross Girshick et al.

We explore a variety of nearest neighbor baseline approaches for image captioning. These approaches find a set of nearest neighbor images in the training set from which a caption may be borrowed for the query image. We select a caption for the query image by finding the caption that best represents the "consensus" of the set of candidate captions gathered from the nearest neighbor images. When measured by automatic evaluation metrics on the MS COCO caption evaluation server, these approaches perform as well as many recent approaches that generate novel captions. However, human studies show that a method that generates novel captions is still preferred over the nearest neighbor approach.

Jacob Devlin, Hao Cheng, Hao Fang et al.

Two recent approaches have achieved state-of-the-art results in image captioning. The first uses a pipelined process where a set of candidate words is generated by a convolutional neural network (CNN) trained on images, and then a maximum entropy (ME) language model is used to arrange these words into a coherent sentence. The second uses the penultimate activation layer of the CNN as input to a recurrent neural network (RNN) that then generates the caption sequence. In this paper, we compare the merits of these different language modeling approaches for the first time by using the same state-of-the-art CNN as input. We examine issues in the different approaches, including linguistic irregularities, caption repetition, and data set overlap. By combining key aspects of the ME and RNN methods, we achieve a new record performance over previously published results on the benchmark COCO dataset. However, the gains we see in BLEU do not translate to human judgments.