NeurIPS 2023 Papers

In this work we explore recent advances in instruction-tuning language models on a range of open instruction-following datasets. Despite recent claims that open models can be on par with state-of-the-art proprietary models, these claims are often accompanied by limited evaluation, making it difficult to compare models across the board and determine the utility of various resources. We provide a large set of instruction-tuned models from 6.7B to 65B parameters in size, trained on 12 instruction datasets ranging from manually curated (e.g., OpenAssistant) to synthetic and distilled (e.g., Alpaca) and systematically evaluate them on their factual knowledge, reasoning, multilinguality, coding, safety, and open-ended instruction following abilities through a collection of automatic, model-based, and human-based metrics. We further introduce Tülu, our best performing instruction-tuned model suite finetuned on a combination of high-quality open resources. Our experiments show that different instruction-tuning datasets can uncover or enhance specific skills, while no single dataset (or combination) provides the best performance across all evaluations. Interestingly, we find that model and human preference-based evaluations fail to reflect differences in model capabilities exposed by benchmark-based evaluations, suggesting the need for the type of systemic evaluation performed in this work. Our evaluations show that the best model in any given evaluation reaches on average 87% of ChatGPT performance, and 73% of GPT-4 performance, suggesting that further investment in building better base models and instruction-tuning data is required to close the gap. We release our instruction-tuned models, including a fully finetuned 65B Tülu, along with our code, data, and evaluation framework to facilitate future research.

In many unpaired image domain translation problems, e.g., style transfer or super-resolution, it is important to keep the translated image similar to its respective input image. We propose the extremal transport (ET) which is a mathematical formalization of the theoretically best possible unpaired translation between a pair of domains w.r.t. the given similarity function. Inspired by the recent advances in neural optimal transport (OT), we propose a scalable algorithm to approximate ET maps as a limit of partial OT maps. We test our algorithm on toy examples and on the unpaired image-to-image translation task. The code is publicly available at https://github.com/milenagazdieva/ExtremalNeuralOptimalTransport

In many search applications related to passage retrieval, text entailment, and subgraph search, the query and each 'document' is a set of elements, with a document being relevant if it contains the query. These elements are not represented by atomic IDs, but by embedded representations, thereby extending set containment to *soft* set containment. Recent applications address soft set containment by encoding sets into fixed-size vectors and checking for elementwise *vector* *dominance*. This 0/1 property can be relaxed to an asymmetric *hinge* *distance* for scoring and ranking candidate documents. Here we focus on data-sensitive, trainable indices for fast retrieval of relevant documents. Existing LSH methods are designed for mostly symmetric or few simple asymmetric distance functions, which are not suitable for hinge distance. Instead, we transform hinge distance into a proposed *dominance* *similarity* measure, to which we then apply a Fourier transform, thereby expressing dominance similarity as an expectation of inner products of functions in the frequency domain. Next, we approximate the expectation with an importance-sampled estimate. The overall consequence is that now we can use a traditional LSH, but in the frequency domain. To ensure that the LSH uses hash bits efficiently, we learn hash functions that are sensitive to both corpus and query distributions, mapped to the frequency domain. Our experiments show that the proposed asymmetric dominance similarity is critical to the targeted applications, and that our LSH, which we call FourierHashNet, provides a better query time vs. retrieval quality trade-off, compared to several baselines. Both the Fourier transform and the trainable hash codes contribute to performance gains.

The mixture proportions of pretraining data domains (e.g., Wikipedia, books, web text) greatly affect language model (LM) performance. In this paper, we propose Domain Reweighting with Minimax Optimization (DoReMi), which first trains a small proxy model using group distributionally robust optimization (Group DRO) over domains to produce domain weights (mixture proportions) without knowledge of downstream tasks. We then resample a dataset with these domain weights and train a larger, full-sized model. In our experiments, we use DoReMi on a 280M-parameter proxy model to set the domain weights for training an 8B-parameter model (30x larger) more efficiently. On The Pile, DoReMi improves perplexity across all domains, even when it downweights a domain. DoReMi improves average few-shot downstream accuracy by 6.5% points over a baseline model trained using The Pile's default domain weights and reaches the baseline accuracy with 2.6x fewer training steps. On the GLaM dataset, DoReMi, which has no knowledge of downstream tasks, even matches the performance of using domain weights tuned on downstream tasks.

Active learning (AL) methods have been proven to be an effective way to reduce the labeling effort by intelligently selecting valuable instances for annotation. Despite their great success with in-distribution (ID) scenarios, AL methods suffer from performance degradation in many real-world applications because out-of-distribution (OOD) instances are always inevitably contained in unlabeled data, which may lead to inefficient sampling. Therefore, several attempts have been explored open-set AL by strategically selecting pure ID instances while filtering OOD instances. However, concentrating solely on selecting pseudo-ID instances may cause the training constraint of the ID classifier and OOD detector. To address this issue, we propose a simple yet effective sampling scheme, Progressive Active Learning (PAL), which employs a progressive sampling mechanism to leverage the active selection of valuable OOD instances. The proposed PAL measures unlabeled instances by synergistically evaluating instances' informativeness and representativeness, and thus it can balance the pseudo-ID and pseudo-OOD instances in each round to enhance both the capacity of the ID classifier and the OOD detector. %Meanwhile, PAL measures unlabeled instances by synergistically evaluating instances' informativeness and representativeness, which can more effectively estimate the values of instances. Extensive experiments on various open-set AL scenarios demonstrate the effectiveness of the proposed PAL, compared with the state-of-the-art methods. The code is available at \url{https://github.com/njustkmg/PAL}.

Many methods of semantic image segmentation have borrowed the success of open compound domain adaptation. They minimize the style gap between the images of source and target domains, more easily predicting the accurate pseudo annotations for target domain's images that train segmentation network. The existing methods globally adapt the scene style of the images, whereas the object styles of different categories or instances are adapted improperly. This paper proposes the Object Style Compensation, where we construct the Object-Level Discrepancy Memory with multiple sets of discrepancy features. The discrepancy features in a set capture the style changes of the same category's object instances adapted from target to source domains. We learn the discrepancy features from the images of source and target domains, storing the discrepancy features in memory. With this memory, we select appropriate discrepancy features for compensating the style information of the object instances of various categories, adapting the object styles to a unified style of source domain. Our method enables a more accurate computation of the pseudo annotations for target domain's images, thus yielding state-of-the-art results on different datasets.

Human Intelligence (HI) excels at combining basic skills to solve complex tasks. This capability is vital for Artificial Intelligence (AI) and should be embedded in comprehensive AI Agents, enabling them to harness expert models for complex task-solving towards Artificial General Intelligence (AGI). Large Language Models (LLMs) show promising learning and reasoning abilities, and can effectively use external models, tools, plugins, or APIs to tackle complex problems. In this work, we introduce OpenAGI, an open-source AGI research and development platform designed for solving multi-step, real-world tasks. Specifically, OpenAGI uses a dual strategy, integrating standard benchmark tasks for benchmarking and evaluation, and open-ended tasks including more expandable models, tools, plugins, or APIs for creative problem-solving. Tasks are presented as natural language queries to the LLM, which then selects and executes appropriate models. We also propose a Reinforcement Learning from Task Feedback (RLTF) mechanism that uses task results to improve the LLM's task-solving ability, which creates a self-improving AI feedback loop. While we acknowledge that AGI is a broad and multifaceted research challenge with no singularly defined solution path, the integration of LLMs with domain-specific expert models, inspired by mirroring the blend of general and specialized intelligence in humans, offers a promising approach towards AGI. We are open-sourcing the OpenAGI project's code, dataset, benchmarks, evaluation methods, and the UI demo to foster community involvement in AGI advancement: https://github.com/agiresearch/OpenAGI.

Question answering (QA) in the field of healthcare has received much attention due to significant advancements in natural language processing. However, existing healthcare QA datasets primarily focus on medical images, clinical notes, or structured electronic health record tables. This leaves the vast potential of combining electrocardiogram (ECG) data with these systems largely untapped. To address this gap, we present ECG-QA, the first QA dataset specifically designed for ECG analysis. The dataset comprises a total of 70 question templates that cover a wide range of clinically relevant ECG topics, each validated by an ECG expert to ensure their clinical utility. As a result, our dataset includes diverse ECG interpretation questions, including those that require a comparative analysis of two different ECGs. In addition, we have conducted numerous experiments to provide valuable insights for future research directions. We believe that ECG-QA will serve as a valuable resource for the development of intelligent QA systems capable of assisting clinicians in ECG interpretations.

This study focuses on the evaluation of the Open Question Answering (Open-QA) task, which can directly estimate the factuality of large language models (LLMs). Current automatic evaluation methods have shown limitations, indicating that human evaluation still remains the most reliable approach. We introduce a new task, QA Evaluation (QA-Eval) and the corresponding dataset EVOUNA, designed to assess the accuracy of AI-generated answers in relation to standard answers within Open-QA. Our evaluation of these methods utilizes human-annotated results to measure their performance. Specifically, the work investigates methods that show high correlation with human evaluations, deeming them more reliable. We also discuss the pitfalls of current methods and methods to improve LLM-based evaluators. We believe this new QA-Eval task and corresponding dataset EVOUNA will facilitate the development of more effective automatic evaluation tools and prove valuable for future research in this area. All resources are available at https://github.com/wangcunxiang/QA-Eval and it is under the Apache-2.0 License.

Machine learning for sign languages is bottlenecked by data. In this paper, we present YouTube-ASL, a large-scale, open-domain corpus of American Sign Language (ASL) videos and accompanying English captions drawn from YouTube. With ~1000 hours of videos and >2500 unique signers, YouTube-ASL is ~3x as large and has ~10x as many unique signers as the largest prior ASL dataset. We train baseline models for ASL to English translation on YouTube-ASL and evaluate them on How2Sign, where we achieve a new fine-tuned state of the art of 12.397 BLEU and, for the first time, nontrivial zero-shot results.

The generalization capability of machine learning systems degenerates notably when the test distribution drifts from the training distribution. Recently, Domain Generalization (DG) has been gaining momentum in enabling machine learning models to generalize to unseen domains. However, most DG methods assume that training and test data share an identical label space, ignoring the potential unseen categories in many real-world applications. In this paper, we delve into a more general but difficult problem termed Open Test-Time DG (OTDG), where both domain shift and open class may occur on the unseen test data. We propose Compaction and Disambiguation (CODA), a novel two-stage framework for learning compact representations and adapting to open classes in the wild. To meaningfully regularize the model's decision boundary, CODA introduces virtual unknown classes and optimizes a new training objective to insert unknowns into the latent space by compacting the embedding space of source known classes. To adapt target samples to the source model, we then disambiguate the decision boundaries between known and unknown classes with a test-time training objective, mitigating the adaptivity gap and catastrophic forgetting challenges. Experiments reveal that CODA can significantly outperform the previous best method on standard DG datasets and harmonize the classification accuracy between known and unknown classes.

We introduce RealTime QA, a dynamic question answering (QA) platform that announces questions and evaluates systems on a regular basis (weekly in this version). RealTime QA inquires about the current world, and QA systems need to answer questions about novel events or information. It therefore challenges static, conventional assumptions in open-domain QA datasets and pursues instantaneous applications. We build strong baseline models upon large pretrained language models, including GPT-3 and T5. Our benchmark is an ongoing effort, and this paper presents real-time evaluation results over the past year. Our experimental results show that GPT-3 can often properly update its generation results, based on newly-retrieved documents, highlighting the importance of up-to-date information retrieval. Nonetheless, we find that GPT-3 tends to return outdated answers when retrieved documents do not provide sufficient information to find an answer. This suggests an important avenue for future research: can an open-domain QA system identify such unanswerable cases and communicate with the user or even the retrieval module to modify the retrieval results? We hope that RealTime QA will spur progress in instantaneous applications of question answering and beyond.

While semantic segmentation has seen tremendous improvements in the past, there are still significant labeling efforts necessary and the problem of limited generalization to classes that have not been present during training. To address this problem, zero-shot semantic segmentation makes use of large self-supervised vision-language models, allowing zero-shot transfer to unseen classes. In this work, we build a benchmark for Multi-domain Evaluation of Zero-Shot Semantic Segmentation (MESS), which allows a holistic analysis of performance across a wide range of domain-specific datasets such as medicine, engineering, earth monitoring, biology, and agriculture. To do this, we reviewed 120 datasets, developed a taxonomy, and classified the datasets according to the developed taxonomy. We select a representative subset consisting of 22 datasets and propose it as the MESS benchmark. We evaluate eight recently published models on the proposed MESS benchmark and analyze characteristics for the performance of zero-shot transfer models. The toolkit is available at https://github.com/blumenstiel/MESS.