Alex Wang

Alex Wang, Georg Meinhardt, Jacob Katz et al.

Financial-research answers are decision-relevant only when another analyst can audit how they were produced: which source was chosen, which period and accounting definition were used, which assumptions were made, and how the calculation was performed. Existing finance benchmarks largely evaluate isolated subskills or final answers, leaving the auditable derivation itself under-measured. We introduce BigFinanceBench, a 928-item expert-authored benchmark of open-ended financial-research tasks in which each item pairs a ground-truth reference answer with a point-weighted rubric that decomposes the derivation into independently checkable steps. BigFinanceBench is workflow-grounded in that it evaluates the full derivation rather than only the final output. Across 36,241 rubric points, the benchmark supports partial-credit evaluation and localization of failures across the analyst workflow. Evaluating ten current frontier and open-weight agents, we find substantial headroom: the best system reaches only 58.8% rubric score, final-answer accuracy is a useful but lossy proxy for derivation quality, and model capability varies non-uniformly across financial workflows.

Sebastian Gehrmann, Abhik Bhattacharjee, Abinaya Mahendiran et al. · amazon-science, cmu

Evaluation in machine learning is usually informed by past choices, for example which datasets or metrics to use. This standardization enables the comparison on equal footing using leaderboards, but the evaluation choices become sub-optimal as better alternatives arise. This problem is especially pertinent in natural language generation which requires ever-improving suites of datasets, metrics, and human evaluation to make definitive claims. To make following best model evaluation practices easier, we introduce GEMv2. The new version of the Generation, Evaluation, and Metrics Benchmark introduces a modular infrastructure for dataset, model, and metric developers to benefit from each others work. GEMv2 supports 40 documented datasets in 51 languages. Models for all datasets can be evaluated online and our interactive data card creation and rendering tools make it easier to add new datasets to the living benchmark.

Nicholas Gray, Megan Moraes, Jiang Bian et al.

Real-time machine learning object detection algorithms are often found within autonomous vehicle technology and depend on quality datasets. It is essential that these algorithms work correctly in everyday conditions as well as under strong sun glare. Reports indicate glare is one of the two most prominent environment-related reasons for crashes. However, existing datasets, such as the Laboratory for Intelligent & Safe Automobiles Traffic Sign (LISA) Dataset and the German Traffic Sign Recognition Benchmark, do not reflect the existence of sun glare at all. This paper presents the GLARE (GLARE is available at: https://github.com/NicholasCG/GLARE_Dataset ) traffic sign dataset: a collection of images with U.S-based traffic signs under heavy visual interference by sunlight. GLARE contains 2,157 images of traffic signs with sun glare, pulled from 33 videos of dashcam footage of roads in the United States. It provides an essential enrichment to the widely used LISA Traffic Sign dataset. Our experimental study shows that although several state-of-the-art baseline architectures have demonstrated good performance on traffic sign detection in conditions without sun glare in the past, they performed poorly when tested against GLARE (e.g., average mAP0.5:0.95 of 19.4). We also notice that current architectures have better detection when trained on images of traffic signs in sun glare performance (e.g., average mAP0.5:0.95 of 39.6), and perform best when trained on a mixture of conditions (e.g., average mAP0.5:0.95 of 42.3).

Jiashuo Yu, Yao Yao, Boyu Chen et al.

We address the challenge of generating high-fidelity, long-form soundtracks that remain coherent across scene transitions. Existing AI music systems are mainly designed for short, isolated clips and lack mechanisms to ensure narrative continuity. We present JenBridge, a modular and interpretable framework for adaptive long-form video soundtracking that ensures both high-fidelity audio generation and transition naturalness. The core architecture is a Transformer-based generative model trained with a flow-matching objective, following a two-stage paradigm: pretraining on large-scale text-audio corpora to establish robust musical priors, then adapting to the video domain with dual text-visual conditioning for precise cross-modal alignment. Crucially, to achieve long-form coherence across diverse scene changes, JenBridge incorporates a novel adaptive transition mechanism. This system features a versatile toolkit of transition styles, including a generative transition method, and uniquely employs a Large Language Model (LLM) Agent that acts as a director to select the most appropriate transition for each narrative shift intelligently. To rigorously assess this task, we propose the LVS Benchmark, a new benchmark that includes a curated dataset and novel evaluation metrics focusing on holistic and transition-aware assessment. Extensive experiments on the proposed benchmark demonstrate that JenBridge significantly outperforms existing methods in both objective and subjective metrics, particularly in terms of transition naturalness and overall narrative coherence. JenBridge represents a significant step towards fully automated, professional-quality video soundtracking.

Alex Wang, Richard Yuanzhe Pang, Angelica Chen et al.

Summarization datasets are often assembled either by scraping naturally occurring public-domain summaries -- which are nearly always in difficult-to-work-with technical domains -- or by using approximate heuristics to extract them from everyday text -- which frequently yields unfaithful summaries. In this work, we turn to a slower but more straightforward approach to developing summarization benchmark data: We hire highly-qualified contractors to read stories and write original summaries from scratch. To amortize reading time, we collect five summaries per document, with the first giving an overview and the subsequent four addressing specific questions. We use this protocol to collect SQuALITY, a dataset of question-focused summaries built on the same public-domain short stories as the multiple-choice dataset QuALITY (Pang et al., 2021). Experiments with state-of-the-art summarization systems show that our dataset is challenging and that existing automatic evaluation metrics are weak indicators of quality.

James Wang, Primo Pu, Zephyr Fung et al.

The acquisition of high-quality, action-aligned demonstration data remains a fundamental bottleneck in scaling foundation models for dexterous robot manipulation. Although robot-free human demonstrations (e.g., the UMI paradigm) offer a scalable alternative to traditional teleoperation, current systems are constrained by sub-optimal hardware ergonomics, open-loop workflows, and a lack of systematic data-mixing strategies. To address these limitations, we present XRZero-G0, a hardware-software co-designed system for embodied data collection and policy learning. The system features an ergonomic, virtual reality interface equipped with a top-view camera and dual specialized grippers to directly improve collection efficiency. To ensure dataset reliability, we propose a closed-loop collection, inspection, training, and evaluation pipeline for non-proprioceptive data. This workflow achieves an 85% data validity rate and establishes a transparent mechanism for quality control. Furthermore, we investigate the empirical scaling behaviors and optimal mixing ratios of robot-free data. Extensive experiments indicate that combining a minimal volume of real-robot data with large-scale robot-free data (e.g., a 10:1 ratio) achieves performance comparable to exclusively real-robot datasets, while reducing acquisition costs by a factor of twenty. Utilizing XRZero-G0, we construct a 2,000-hour robot-free dataset that enables zero-shot cross-embodiment transfer to a target physical robot, demonstrating a highly scalable methodology for generalized real-world manipulation.Our project repository: https://github.com/X-Square-Robot/XRZero-G0

Julian Michael, Ari Holtzman, Alicia Parrish et al.

We present the results of the NLP Community Metasurvey. Run from May to June 2022, the survey elicited opinions on controversial issues, including industry influence in the field, concerns about AGI, and ethics. Our results put concrete numbers to several controversies: For example, respondents are split almost exactly in half on questions about the importance of artificial general intelligence, whether language models understand language, and the necessity of linguistic structure and inductive bias for solving NLP problems. In addition, the survey posed meta-questions, asking respondents to predict the distribution of survey responses. This allows us not only to gain insight on the spectrum of beliefs held by NLP researchers, but also to uncover false sociological beliefs where the community's predictions don't match reality. We find such mismatches on a wide range of issues. Among other results, the community greatly overestimates its own belief in the usefulness of benchmarks and the potential for scaling to solve real-world problems, while underestimating its own belief in the importance of linguistic structure, inductive bias, and interdisciplinary science.

Alex Wang, Haotong Liang, Austin McDannald et al.

Autonomous physical science is revolutionizing materials science. In these systems, machine learning controls experiment design, execution, and analysis in a closed loop. Active learning, the machine learning field of optimal experiment design, selects each subsequent experiment to maximize knowledge toward the user goal. Autonomous system performance can be further improved with implementation of scientific machine learning, also known as inductive bias-engineered artificial intelligence, which folds prior knowledge of physical laws (e.g., Gibbs phase rule) into the algorithm. As the number, diversity, and uses for active learning strategies grow, there is an associated growing necessity for real-world reference datasets to benchmark strategies. We present a reference dataset and demonstrate its use to benchmark active learning strategies in the form of various acquisition functions. Active learning strategies are used to rapidly identify materials with optimal physical properties within a ternary materials system. The data is from an actual Fe-Co-Ni thin-film library and includes previously acquired experimental data for materials compositions, X-ray diffraction patterns, and two functional properties of magnetic coercivity and the Kerr rotation. Popular active learning methods along with a recent scientific active learning method are benchmarked for their materials optimization performance. We discuss the relationship between algorithm performance, materials search space complexity, and the incorporation of prior knowledge.

Peike Li, Boyu Chen, Yao Yao et al.

Music generation has attracted growing interest with the advancement of deep generative models. However, generating music conditioned on textual descriptions, known as text-to-music, remains challenging due to the complexity of musical structures and high sampling rate requirements. Despite the task's significance, prevailing generative models exhibit limitations in music quality, computational efficiency, and generalization. This paper introduces JEN-1, a universal high-fidelity model for text-to-music generation. JEN-1 is a diffusion model incorporating both autoregressive and non-autoregressive training. Through in-context learning, JEN-1 performs various generation tasks including text-guided music generation, music inpainting, and continuation. Evaluations demonstrate JEN-1's superior performance over state-of-the-art methods in text-music alignment and music quality while maintaining computational efficiency. Our demos are available at https://jenmusic.ai/audio-demos

Logan Saar, Haotong Liang, Alex Wang et al.

The next generation of physical science involves robot scientists - autonomous physical science systems capable of experimental design, execution, and analysis in a closed loop. Such systems have shown real-world success for scientific exploration and discovery, including the first discovery of a best-in-class material. To build and use these systems, the next generation workforce requires expertise in diverse areas including ML, control systems, measurement science, materials synthesis, decision theory, among others. However, education is lagging. Educators need a low-cost, easy-to-use platform to teach the required skills. Industry can also use such a platform for developing and evaluating autonomous physical science methodologies. We present the next generation in science education, a kit for building a low-cost autonomous scientist. The kit was used during two courses at the University of Maryland to teach undergraduate and graduate students autonomous physical science. We discuss its use in the course and its greater capability to teach the dual tasks of autonomous model exploration, optimization, and determination, with an example of autonomous experimental "discovery" of the Henderson-Hasselbalch equation.

Shivalika Singh, Yiyang Nan, Alex Wang et al.

Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. Providers like Google and OpenAI have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. We show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on the arena distribution, based on our conservative estimates. Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field

Vincent Fan, Yujie Qian, Alex Wang et al.

Information extraction from chemistry literature is vital for constructing up-to-date reaction databases for data-driven chemistry. Complete extraction requires combining information across text, tables, and figures, whereas prior work has mainly investigated extracting reactions from single modalities. In this paper, we present OpenChemIE to address this complex challenge and enable the extraction of reaction data at the document level. OpenChemIE approaches the problem in two steps: extracting relevant information from individual modalities and then integrating the results to obtain a final list of reactions. For the first step, we employ specialized neural models that each address a specific task for chemistry information extraction, such as parsing molecules or reactions from text or figures. We then integrate the information from these modules using chemistry-informed algorithms, allowing for the extraction of fine-grained reaction data from reaction condition and substrate scope investigations. Our machine learning models attain state-of-the-art performance when evaluated individually, and we meticulously annotate a challenging dataset of reaction schemes with R-groups to evaluate our pipeline as a whole, achieving an F1 score of 69.5%. Additionally, the reaction extraction results of \ours attain an accuracy score of 64.3% when directly compared against the Reaxys chemical database. We provide OpenChemIE freely to the public as an open-source package, as well as through a web interface.

Yao Yao, Peike Li, Boyu Chen et al.

With rapid advances in generative artificial intelligence, the text-to-music synthesis task has emerged as a promising direction for music generation. Nevertheless, achieving precise control over multi-track generation remains an open challenge. While existing models excel in directly generating multi-track mix, their limitations become evident when it comes to composing individual tracks and integrating them in a controllable manner. This departure from the typical workflows of professional composers hinders the ability to refine details in specific tracks. To address this gap, we propose JEN-1 Composer, a unified framework designed to efficiently model marginal, conditional, and joint distributions over multi-track music using a single model. Building upon an audio latent diffusion model, JEN-1 Composer extends the versatility of multi-track music generation. We introduce a progressive curriculum training strategy, which gradually escalates the difficulty of training tasks while ensuring the model's generalization ability and facilitating smooth transitions between different scenarios. During inference, users can iteratively generate and select music tracks, thus incrementally composing entire musical pieces in accordance with the Human-AI co-composition workflow. Our approach demonstrates state-of-the-art performance in controllable and high-fidelity multi-track music synthesis, marking a significant advancement in interactive AI-assisted music creation. Our demo pages are available at www.jenmusic.ai/research.

Yada Pruksachatkun, Phil Yeres, Haokun Liu et al.

We introduce jiant, an open source toolkit for conducting multitask and transfer learning experiments on English NLU tasks. jiant enables modular and configuration-driven experimentation with state-of-the-art models and implements a broad set of tasks for probing, transfer learning, and multitask training experiments. jiant implements over 50 NLU tasks, including all GLUE and SuperGLUE benchmark tasks. We demonstrate that jiant reproduces published performance on a variety of tasks and models, including BERT and RoBERTa. jiant is available at https://jiant.info.

Team Cohere, Aakanksha, Arash Ahmadian et al. · mila

In this report we describe the development of Command A, a powerful large language model purpose-built to excel at real-world enterprise use cases. Command A is an agent-optimised and multilingual-capable model, with support for 23 languages of global business, and a novel hybrid architecture balancing efficiency with top of the range performance. It offers best-in-class Retrieval Augmented Generation (RAG) capabilities with grounding and tool use to automate sophisticated business processes. These abilities are achieved through a decentralised training approach, including self-refinement algorithms and model merging techniques. We also include results for Command R7B which shares capability and architectural similarities to Command A. Weights for both models have been released for research purposes. This technical report details our original training pipeline and presents an extensive evaluation of our models across a suite of enterprise-relevant tasks and public benchmarks, demonstrating excellent performance and efficiency.

Boyu Chen, Peike Li, Yao Yao et al.

Large models for text-to-music generation have achieved significant progress, facilitating the creation of high-quality and varied musical compositions from provided text prompts. However, input text prompts may not precisely capture user requirements, particularly when the objective is to generate music that embodies a specific concept derived from a designated reference collection. In this paper, we propose a novel method for customized text-to-music generation, which can capture the concept from a two-minute reference music and generate a new piece of music conforming to the concept. We achieve this by fine-tuning a pretrained text-to-music model using the reference music. However, directly fine-tuning all parameters leads to overfitting issues. To address this problem, we propose a Pivotal Parameters Tuning method that enables the model to assimilate the new concept while preserving its original generative capabilities. Additionally, we identify a potential concept conflict when introducing multiple concepts into the pretrained model. We present a concept enhancement strategy to distinguish multiple concepts, enabling the fine-tuned model to generate music incorporating either individual or multiple concepts simultaneously. Since we are the first to work on the customized music generation task, we also introduce a new dataset and evaluation protocol for the new task. Our proposed Jen1-DreamStyler outperforms several baselines in both qualitative and quantitative evaluations. Demos will be available at https://www.jenmusic.ai/research#DreamStyler.

Adib Hasan, Ileana Rugina, Alex Wang

This paper investigates the impact of model compression on the way Large Language Models (LLMs) process prompts, particularly concerning jailbreak resistance. We show that moderate WANDA pruning can enhance resistance to jailbreaking attacks without fine-tuning, while maintaining performance on standard benchmarks. To systematically evaluate this safety enhancement, we introduce a dataset of 225 harmful tasks across five categories. Our analysis of LLaMA-2 Chat, Vicuna 1.3, and Mistral Instruct v0.2 reveals that pruning benefits correlate with initial model safety levels. We interpret these results by examining changes in attention patterns and perplexity shifts, demonstrating that pruned models exhibit sharper attention and increased sensitivity to artificial jailbreak constructs. We extend our evaluation to the AdvBench harmful behavior tasks and the GCG attack method. We find that LLaMA-2 is much safer on AdvBench prompts than on our dataset when evaluated with manual jailbreak attempts, and that pruning is effective against both automated attacks and manual jailbreaking on Advbench.

Max Marion, Ahmet Üstün, Luiza Pozzobon et al.

Large volumes of text data have contributed significantly to the development of large language models (LLMs) in recent years. This data is typically acquired by scraping the internet, leading to pretraining datasets comprised of noisy web text. To date, efforts to prune these datasets down to a higher quality subset have relied on hand-crafted heuristics encoded as rule-based filters. In this work, we take a wider view and explore scalable estimates of data quality that can be used to systematically measure the quality of pretraining data. We perform a rigorous comparison at scale of the simple data quality estimator of perplexity, as well as more sophisticated and computationally intensive estimates of the Error L2-Norm and memorization. These metrics are used to rank and prune pretraining corpora, and we subsequently compare LLMs trained on these pruned datasets. Surprisingly, we find that the simple technique of perplexity outperforms our more computationally expensive scoring methods. We improve over our no-pruning baseline while training on as little as 30% of the original training dataset. Our work sets the foundation for unexplored strategies in automatically curating high quality corpora and suggests the majority of pretraining data can be removed while retaining performance.

Thomas Scialom, Paul-Alexis Dray, Patrick Gallinari et al.

Summarization evaluation remains an open research problem: current metrics such as ROUGE are known to be limited and to correlate poorly with human judgments. To alleviate this issue, recent work has proposed evaluation metrics which rely on question answering models to assess whether a summary contains all the relevant information in its source document. Though promising, the proposed approaches have so far failed to correlate better than ROUGE with human judgments. In this paper, we extend previous approaches and propose a unified framework, named QuestEval. In contrast to established metrics such as ROUGE or BERTScore, QuestEval does not require any ground-truth reference. Nonetheless, QuestEval substantially improves the correlation with human judgments over four evaluation dimensions (consistency, coherence, fluency, and relevance), as shown in the extensive experiments we report.

Alex Wang, Kyunghyun Cho, Mike Lewis

Practical applications of abstractive summarization models are limited by frequent factual inconsistencies with respect to their input. Existing automatic evaluation metrics for summarization are largely insensitive to such errors. We propose an automatic evaluation protocol called QAGS (pronounced "kags") that is designed to identify factual inconsistencies in a generated summary. QAGS is based on the intuition that if we ask questions about a summary and its source, we will receive similar answers if the summary is factually consistent with the source. To evaluate QAGS, we collect human judgments of factual consistency on model-generated summaries for the CNN/DailyMail (Hermann et al., 2015) and XSUM (Narayan et al., 2018) summarization datasets. QAGS has substantially higher correlations with these judgments than other automatic evaluation metrics. Also, QAGS offers a natural form of interpretability: The answers and questions generated while computing QAGS indicate which tokens of a summary are inconsistent and why. We believe QAGS is a promising tool in automatically generating usable and factually consistent text.

Elman Mansimov, Alex Wang, Sean Welleck et al.

Undirected neural sequence models such as BERT (Devlin et al., 2019) have received renewed interest due to their success on discriminative natural language understanding tasks such as question-answering and natural language inference. The problem of generating sequences directly from these models has received relatively little attention, in part because generating from undirected models departs significantly from conventional monotonic generation in directed sequence models. We investigate this problem by proposing a generalized model of sequence generation that unifies decoding in directed and undirected models. The proposed framework models the process of generation rather than the resulting sequence, and under this framework, we derive various neural sequence models as special cases, such as autoregressive, semi-autoregressive, and refinement-based non-autoregressive models. This unification enables us to adapt decoding algorithms originally developed for directed sequence models to undirected sequence models. We demonstrate this by evaluating various handcrafted and learned decoding strategies on a BERT-like machine translation model (Lample & Conneau, 2019). The proposed approach achieves constant-time translation results on par with linear-time translation results from the same undirected sequence model, while both are competitive with the state-of-the-art on WMT'14 English-German translation.

Ian Tenney, Patrick Xia, Berlin Chen et al.

Contextualized representation models such as ELMo (Peters et al., 2018a) and BERT (Devlin et al., 2018) have recently achieved state-of-the-art results on a diverse array of downstream NLP tasks. Building on recent token-level probing work, we introduce a novel edge probing task design and construct a broad suite of sub-sentence tasks derived from the traditional structured NLP pipeline. We probe word-level contextual representations from four recent models and investigate how they encode sentence structure across a range of syntactic, semantic, local, and long-range phenomena. We find that existing models trained on language modeling and translation produce strong representations for syntactic phenomena, but only offer comparably small improvements on semantic tasks over a non-contextual baseline.

Alex Wang, Yada Pruksachatkun, Nikita Nangia et al.

In the last year, new models and methods for pretraining and transfer learning have driven striking performance improvements across a range of language understanding tasks. The GLUE benchmark, introduced a little over one year ago, offers a single-number metric that summarizes progress on a diverse set of such tasks, but performance on the benchmark has recently surpassed the level of non-expert humans, suggesting limited headroom for further research. In this paper we present SuperGLUE, a new benchmark styled after GLUE with a new set of more difficult language understanding tasks, a software toolkit, and a public leaderboard. SuperGLUE is available at super.gluebenchmark.com.

Najoung Kim, Roma Patel, Adam Poliak et al.

We introduce a set of nine challenge tasks that test for the understanding of function words. These tasks are created by structurally mutating sentences from existing datasets to target the comprehension of specific types of function words (e.g., prepositions, wh-words). Using these probing tasks, we explore the effects of various pretraining objectives for sentence encoders (e.g., language modeling, CCG supertagging and natural language inference (NLI)) on the learned representations. Our results show that pretraining on language modeling performs the best on average across our probing tasks, supporting its widespread use for pretraining state-of-the-art NLP models, and CCG supertagging and NLI pretraining perform comparably. Overall, no pretraining objective dominates across the board, and our function word probing tasks highlight several intuitive differences between pretraining objectives, e.g., that NLI helps the comprehension of negation.

Chandler May, Alex Wang, Shikha Bordia et al.

The Word Embedding Association Test shows that GloVe and word2vec word embeddings exhibit human-like implicit biases based on gender, race, and other social constructs (Caliskan et al., 2017). Meanwhile, research on learning reusable text representations has begun to explore sentence-level texts, with some sentence encoders seeing enthusiastic adoption. Accordingly, we extend the Word Embedding Association Test to measure bias in sentence encoders. We then test several sentence encoders, including state-of-the-art methods such as ELMo and BERT, for the social biases studied in prior work and two important biases that are difficult or impossible to test at the word level. We observe mixed results including suspicious patterns of sensitivity that suggest the test's assumptions may not hold in general. We conclude by proposing directions for future work on measuring bias in sentence encoders.

Alex Wang, Kyunghyun Cho

We show that BERT (Devlin et al., 2018) is a Markov random field language model. This formulation gives way to a natural procedure to sample sentences from BERT. We generate from BERT and find that it can produce high-quality, fluent generations. Compared to the generations of a traditional left-to-right language model, BERT generates sentences that are more diverse but of slightly worse quality.

Alex Wang, Jan Hula, Patrick Xia et al.

Natural language understanding has recently seen a surge of progress with the use of sentence encoders like ELMo (Peters et al., 2018a) and BERT (Devlin et al., 2019) which are pretrained on variants of language modeling. We conduct the first large-scale systematic study of candidate pretraining tasks, comparing 19 different tasks both as alternatives and complements to language modeling. Our primary results support the use language modeling, especially when combined with pretraining on additional labeled-data tasks. However, our results are mixed across pretraining tasks and show some concerning trends: In ELMo's pretrain-then-freeze paradigm, random baselines are worryingly strong and results vary strikingly across target tasks. In addition, fine-tuning BERT on an intermediate task often negatively impacts downstream transfer. In a more positive trend, we see modest gains from multitask training, suggesting the development of more sophisticated multitask and transfer learning techniques as an avenue for further research.

Alex Wang, Amanpreet Singh, Julian Michael et al.

For natural language understanding (NLU) technology to be maximally useful, both practically and as a scientific object of study, it must be general: it must be able to process language in a way that is not exclusively tailored to any one specific task or dataset. In pursuit of this objective, we introduce the General Language Understanding Evaluation benchmark (GLUE), a tool for evaluating and analyzing the performance of models across a diverse range of existing NLU tasks. GLUE is model-agnostic, but it incentivizes sharing knowledge across tasks because certain tasks have very limited training data. We further provide a hand-crafted diagnostic test suite that enables detailed linguistic analysis of NLU models. We evaluate baselines based on current methods for multi-task and transfer learning and find that they do not immediately give substantial improvements over the aggregate performance of training a separate model per task, indicating room for improvement in developing general and robust NLU systems.

Abhratanu Dutta, Aravindan Vijayaraghavan, Alex Wang

The Euclidean k-means problem is arguably the most widely-studied clustering problem in machine learning. While the k-means objective is NP-hard in the worst-case, practitioners have enjoyed remarkable success in applying heuristics like Lloyd's algorithm for this problem. To address this disconnect, we study the following question: what properties of real-world instances will enable us to design efficient algorithms and prove guarantees for finding the optimal clustering? We consider a natural notion called additive perturbation stability that we believe captures many practical instances. Stable instances have unique optimal k-means solutions that do not change even when each point is perturbed a little (in Euclidean distance). This captures the property that the k-means optimal solution should be tolerant to measurement errors and uncertainty in the points. We design efficient algorithms that provably recover the optimal clustering for instances that are additive perturbation stable. When the instance has some additional separation, we show an efficient algorithm with provable guarantees that is also robust to outliers. We complement these results by studying the amount of stability in real datasets and demonstrating that our algorithm performs well on these benchmark datasets.