Douwe Kiela

BigScience Workshop, Teven Le Scao, Angela Fan et al. · allen-ai, berkeley

Large language models (LLMs) have been shown to be able to perform new tasks based on a few demonstrations or natural language instructions. While these capabilities have led to widespread adoption, most LLMs are developed by resource-rich organizations and are frequently kept from the public. As a step towards democratizing this powerful technology, we present BLOOM, a 176B-parameter open-access language model designed and built thanks to a collaboration of hundreds of researchers. BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). We find that BLOOM achieves competitive performance on a wide variety of benchmarks, with stronger results after undergoing multitask prompted finetuning. To facilitate future research and applications using LLMs, we publicly release our models and code under the Responsible AI License.

Niklas Muennighoff, Luca Soldaini, Dirk Groeneveld et al. · allen-ai

We introduce OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE). OLMoE-1B-7B has 7 billion (B) parameters but uses only 1B per input token. We pretrain it on 5 trillion tokens and further adapt it to create OLMoE-1B-7B-Instruct. Our models outperform all available models with similar active parameters, even surpassing larger ones like Llama2-13B-Chat and DeepSeekMoE-16B. We present various experiments on MoE training, analyze routing in our model showing high specialization, and open-source all aspects of our work: model weights, training data, code, and logs.

Leandro von Werra, Lewis Tunstall, Abhishek Thakur et al. · salesforce

Evaluation is a key part of machine learning (ML), yet there is a lack of support and tooling to enable its informed and systematic practice. We introduce Evaluate and Evaluation on the Hub --a set of tools to facilitate the evaluation of models and datasets in ML. Evaluate is a library to support best practices for measurements, metrics, and comparisons of data and models. Its goal is to support reproducibility of evaluation, centralize and document the evaluation process, and broaden evaluation to cover more facets of model performance. It includes over 50 efficient canonical implementations for a variety of domains and scenarios, interactive documentation, and the ability to easily share implementations and outcomes. The library is available at https://github.com/huggingface/evaluate. In addition, we introduce Evaluation on the Hub, a platform that enables the large-scale evaluation of over 75,000 models and 11,000 datasets on the Hugging Face Hub, for free, at the click of a button. Evaluation on the Hub is available at https://huggingface.co/autoevaluate.

Hugo Laurençon, Lucile Saulnier, Léo Tronchon et al. · stanford

Large multimodal models trained on natural documents, which interleave images and text, outperform models trained on image-text pairs on various multimodal benchmarks. However, the datasets used to train these models have not been released, and the collection process has not been fully specified. We introduce the OBELICS dataset, an open web-scale filtered dataset of interleaved image-text documents comprising 141 million web pages extracted from Common Crawl, 353 million associated images, and 115 billion text tokens. We describe the dataset creation process, present comprehensive filtering rules, and provide an analysis of the dataset's content. To show the viability of OBELICS, we train vision and language models of 9 and 80 billion parameters named IDEFICS, and obtain competitive performance on different multimodal benchmarks. We release our dataset, models and code.

Tristan Thrush, Kushal Tirumala, Anmol Gupta et al. · meta-ai

We introduce Dynatask: an open source system for setting up custom NLP tasks that aims to greatly lower the technical knowledge and effort required for hosting and evaluating state-of-the-art NLP models, as well as for conducting model in the loop data collection with crowdworkers. Dynatask is integrated with Dynabench, a research platform for rethinking benchmarking in AI that facilitates human and model in the loop data collection and evaluation. To create a task, users only need to write a short task configuration file from which the relevant web interfaces and model hosting infrastructure are automatically generated. The system is available at https://dynabench.org/ and the full library can be found at https://github.com/facebookresearch/dynabench.

Mark Mazumder, Colby Banbury, Xiaozhe Yao et al.

Machine learning research has long focused on models rather than datasets, and prominent datasets are used for common ML tasks without regard to the breadth, difficulty, and faithfulness of the underlying problems. Neglecting the fundamental importance of data has given rise to inaccuracy, bias, and fragility in real-world applications, and research is hindered by saturation across existing dataset benchmarks. In response, we present DataPerf, a community-led benchmark suite for evaluating ML datasets and data-centric algorithms. We aim to foster innovation in data-centric AI through competition, comparability, and reproducibility. We enable the ML community to iterate on datasets, instead of just architectures, and we provide an open, online platform with multiple rounds of challenges to support this iterative development. The first iteration of DataPerf contains five benchmarks covering a wide spectrum of data-centric techniques, tasks, and modalities in vision, speech, acquisition, debugging, and diffusion prompting, and we support hosting new contributed benchmarks from the community. The benchmarks, online evaluation platform, and baseline implementations are open source, and the MLCommons Association will maintain DataPerf to ensure long-term benefits to academia and industry.

Tristan Thrush, Ryan Jiang, Max Bartolo et al.

We present a novel task and dataset for evaluating the ability of vision and language models to conduct visio-linguistic compositional reasoning, which we call Winoground. Given two images and two captions, the goal is to match them correctly - but crucially, both captions contain a completely identical set of words, only in a different order. The dataset was carefully hand-curated by expert annotators and is labeled with a rich set of fine-grained tags to assist in analyzing model performance. We probe a diverse range of state-of-the-art vision and language models and find that, surprisingly, none of them do much better than chance. Evidently, these models are not as skilled at visio-linguistic compositional reasoning as we might have hoped. We perform an extensive analysis to obtain insights into how future work might try to mitigate these models' shortcomings. We aim for Winoground to serve as a useful evaluation set for advancing the state of the art and driving further progress in the field. The dataset is available at https://huggingface.co/datasets/facebook/winoground.

Pranab Islam, Anand Kannappan, Douwe Kiela et al.

FinanceBench is a first-of-its-kind test suite for evaluating the performance of LLMs on open book financial question answering (QA). It comprises 10,231 questions about publicly traded companies, with corresponding answers and evidence strings. The questions in FinanceBench are ecologically valid and cover a diverse set of scenarios. They are intended to be clear-cut and straightforward to answer to serve as a minimum performance standard. We test 16 state of the art model configurations (including GPT-4-Turbo, Llama2 and Claude2, with vector stores and long context prompts) on a sample of 150 cases from FinanceBench, and manually review their answers (n=2,400). The cases are available open-source. We show that existing LLMs have clear limitations for financial QA. Notably, GPT-4-Turbo used with a retrieval system incorrectly answered or refused to answer 81% of questions. While augmentation techniques such as using longer context window to feed in relevant evidence improve performance, they are unrealistic for enterprise settings due to increased latency and cannot support larger financial documents. We find that all models examined exhibit weaknesses, such as hallucinations, that limit their suitability for use by enterprises.

Chris Chinenye Emezue, Sanchit Gandhi, Lewis Tunstall et al. · huggingface

The advancement of speech technologies has been remarkable, yet its integration with African languages remains limited due to the scarcity of African speech corpora. To address this issue, we present AfroDigits, a minimalist, community-driven dataset of spoken digits for African languages, currently covering 38 African languages. As a demonstration of the practical applications of AfroDigits, we conduct audio digit classification experiments on six African languages [Igbo (ibo), Yoruba (yor), Rundi (run), Oshiwambo (kua), Shona (sna), and Oromo (gax)] using the Wav2Vec2.0-Large and XLS-R models. Our experiments reveal a useful insight on the effect of mixing African speech corpora during finetuning. AfroDigits is the first published audio digit dataset for African languages and we believe it will, among other things, pave the way for Afro-centric speech applications such as the recognition of telephone numbers, and street numbers. We release the dataset and platform publicly at https://huggingface.co/datasets/chrisjay/crowd-speech-africa and https://huggingface.co/spaces/chrisjay/afro-speech respectively.

Rebecca Qian, Candace Ross, Jude Fernandes et al. · meta-ai

Unwanted and often harmful social biases are becoming ever more salient in NLP research, affecting both models and datasets. In this work, we ask whether training on demographically perturbed data leads to fairer language models. We collect a large dataset of human annotated text perturbations and train a neural perturbation model, which we show outperforms heuristic alternatives. We find that (i) language models (LMs) pre-trained on demographically perturbed corpora are typically more fair, and (ii) LMs finetuned on perturbed GLUE datasets exhibit less demographic bias on downstream tasks, and (iii) fairness improvements do not come at the expense of performance on downstream tasks. Lastly, we discuss outstanding questions about how best to evaluate the (un)fairness of large language models. We hope that this exploration of neural demographic perturbation will help drive more improvement towards fairer NLP.

William Berrios, Gautam Mittal, Tristan Thrush et al.

We propose LENS, a modular approach for tackling computer vision problems by leveraging the power of large language models (LLMs). Our system uses a language model to reason over outputs from a set of independent and highly descriptive vision modules that provide exhaustive information about an image. We evaluate the approach on pure computer vision settings such as zero- and few-shot object recognition, as well as on vision and language problems. LENS can be applied to any off-the-shelf LLM and we find that the LLMs with LENS perform highly competitively with much bigger and much more sophisticated systems, without any multimodal training whatsoever. We open-source our code at https://github.com/ContextualAI/lens and provide an interactive demo.

Selvan Sunitha Ravi, Bartosz Mielczarek, Anand Kannappan et al.

Retrieval Augmented Generation (RAG) techniques aim to mitigate hallucinations in Large Language Models (LLMs). However, LLMs can still produce information that is unsupported or contradictory to the retrieved contexts. We introduce LYNX, a SOTA hallucination detection LLM that is capable of advanced reasoning on challenging real-world hallucination scenarios. To evaluate LYNX, we present HaluBench, a comprehensive hallucination evaluation benchmark, consisting of 15k samples sourced from various real-world domains. Our experiment results show that LYNX outperforms GPT-4o, Claude-3-Sonnet, and closed and open-source LLM-as-a-judge models on HaluBench. We release LYNX, HaluBench and our evaluation code for public access.

Karel D'Oosterlinck, Winnie Xu, Chris Develder et al.

Large Language Models (LLMs) are often aligned using contrastive alignment objectives and preference pair datasets. The interaction between model, paired data, and objective makes alignment a complicated procedure, sometimes producing subpar results. We study this and find that (i) preference data gives a better learning signal when the underlying responses are contrastive, and (ii) alignment objectives lead to better performance when they specify more control over the model during training. Based on these insights, we introduce Contrastive Learning from AI Revisions (CLAIR), a data-creation method which leads to more contrastive preference pairs, and Anchored Preference Optimization (APO), a controllable and more stable alignment objective. We align Llama-3-8B-Instruct using various comparable datasets and alignment objectives and measure MixEval-Hard scores, which correlate highly with human judgments. The CLAIR preferences lead to the strongest performance out of all datasets, and APO consistently outperforms less controllable objectives. Our best model, trained on 32K CLAIR preferences with APO, improves Llama-3-8B-Instruct by 7.65%, closing the gap with GPT4-turbo by 45%. Our code is available at https://github.com/ContextualAI/CLAIR_and_APO.

Margaret Mitchell, Alexandra Sasha Luccioni, Nathan Lambert et al. · huggingface, salesforce

We identify the task of measuring data to quantitatively characterize the composition of machine learning data and datasets. Similar to an object's height, width, and volume, data measurements quantify different attributes of data along common dimensions that support comparison. Several lines of research have proposed what we refer to as measurements, with differing terminology; we bring some of this work together, particularly in fields of computer vision and language, and build from it to motivate measuring data as a critical component of responsible AI development. Measuring data aids in systematically building and analyzing machine learning (ML) data towards specific goals and gaining better control of what modern ML systems will learn. We conclude with a discussion of the many avenues of future work, the limitations of data measurements, and how to leverage these measurement approaches in research and practice.

Nick Ferguson, Josh Pennington, Narek Beghian et al.

Unstructured documents like PDFs contain valuable structured information, but downstream systems require this data in reliable, standardized formats. LLMs are increasingly deployed to automate this extraction, making accuracy and reliability paramount. However, progress is bottlenecked by two gaps. First, no end-to-end benchmark evaluates PDF-to-JSON extraction under enterprise-scale schema breadth. Second, no principled methodology captures the semantics of nested extraction, where fields demand different notions of correctness (exact match for identifiers, tolerance for quantities, semantic equivalence for names), arrays require alignment, and omission must be distinguished from hallucination. We address both gaps with ExtractBench, an open-source benchmark and evaluation framework for PDF-to-JSON structured extraction. The benchmark pairs 35 PDF documents with JSON Schemas and human-annotated gold labels across economically valuable domains, yielding 12,867 evaluatable fields spanning schema complexities from tens to hundreds of fields. The evaluation framework treats the schema as an executable specification: each field declares its scoring metric. Baseline evaluations reveal that frontier models (GPT-5/5.2, Gemini-3 Flash/Pro, Claude 4.5 Opus/Sonnet) remain unreliable on realistic schemas. Performance degrades sharply with schema breadth, culminating in 0% valid output on a 369-field financial reporting schema across all tested models. We release ExtractBench at https://github.com/ContextualAI/extract-bench.

Rajan Vivek, Kawin Ethayarajh, Diyi Yang et al.

Modern language models often exhibit powerful but brittle behavior, leading to the development of larger and more diverse benchmarks to reliably assess their behavior. Here, we suggest that model performance can be benchmarked and elucidated with much smaller evaluation sets. We first show that in six popular language classification benchmarks, model confidence in the correct class on many pairs of points is strongly correlated across models. We build upon this phenomenon to propose Anchor Point Selection, a technique to select small subsets of datasets that capture model behavior across the entire dataset. Anchor points reliably rank models: across 87 diverse language model-prompt pairs, evaluating models using 1-30 anchor points outperforms uniform sampling and other baselines at accurately ranking models. Moreover, just several anchor points can be used to estimate model per-class predictions on all other points in a dataset with low mean absolute error, sufficient for gauging where the model is likely to fail. Lastly, we present Anchor Point Maps for visualizing these insights and facilitating comparisons of the performance of different models on various regions within the dataset distribution.

Dat Tran, Douwe Kiela

Recent work reports strong performance from multi-agent LLM systems (MAS), but these gains are often confounded by increased test-time computation. When computation is normalized, single-agent systems (SAS) can match or outperform MAS, yet the theoretical basis and evaluation methodology behind this comparison remain unclear. We present an information-theoretic argument, grounded in the Data Processing Inequality, suggesting that under a fixed reasoning-token budget and with perfect context utilization, single-agent systems are more information-efficient. This perspective further predicts that multi-agent systems become competitive when a single agent's effective context utilization is degraded, or when more compute is expended. We test these predictions in a controlled empirical study across three model families (Qwen3, DeepSeek-R1-Distill-Llama, and Gemini 2.5), comparing SAS with multiple MAS architectures under matched budgets. We find that SAS consistently match or outperform MAS on multi-hop reasoning tasks when reasoning tokens are held constant. Beyond aggregate performance, we conduct a detailed diagnostic analysis of system behavior and evaluation methodology. We identify significant artifacts in API-based budget control (particularly in Gemini 2.5) and in standard benchmarks, both of which can inflate apparent gains from MAS. Overall, our results suggest that, for multi-hop reasoning tasks, many reported advantages of multi-agent systems are better explained by unaccounted computation and context effects rather than inherent architectural benefits, and highlight the importance of understanding and explicitly controlling the trade-offs between compute, context, and coordination in agentic systems.

Ard Snijders, Douwe Kiela, Katerina Margatina

In recent years, active learning has been successfully applied to an array of NLP tasks. However, prior work often assumes that training and test data are drawn from the same distribution. This is problematic, as in real-life settings data may stem from several sources of varying relevance and quality. We show that four popular active learning schemes fail to outperform random selection when applied to unlabelled pools comprised of multiple data sources on the task of natural language inference. We reveal that uncertainty-based strategies perform poorly due to the acquisition of collective outliers, i.e., hard-to-learn instances that hamper learning and generalization. When outliers are removed, strategies are found to recover and outperform random baselines. In further analysis, we find that collective outliers vary in form between sources, and show that hard-to-learn data is not always categorically harmful. Lastly, we leverage dataset cartography to introduce difficulty-stratified testing and find that different strategies are affected differently by example learnability and difficulty.

Niklas Muennighoff, Hongjin Su, Liang Wang et al. · microsoft-research

All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8x7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by > 60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm.

Nicholas Lui, Bryan Chia, William Berrios et al.

Computer vision models have been known to encode harmful biases, leading to the potentially unfair treatment of historically marginalized groups, such as people of color. However, there remains a lack of datasets balanced along demographic traits that can be used to evaluate the downstream fairness of these models. In this work, we demonstrate that diffusion models can be leveraged to create such a dataset. We first use a diffusion model to generate a large set of images depicting various occupations. Subsequently, each image is edited using inpainting to generate multiple variants, where each variant refers to a different perceived race. Using this dataset, we benchmark several vision-language models on a multi-class occupation classification task. We find that images generated with non-Caucasian labels have a significantly higher occupation misclassification rate than images generated with Caucasian labels, and that several misclassifications are suggestive of racial biases. We measure a model's downstream fairness by computing the standard deviation in the probability of predicting the true occupation label across the different perceived identity groups. Using this fairness metric, we find significant disparities between the evaluated vision-and-language models. We hope that our work demonstrates the potential value of diffusion methods for fairness evaluations.

Omri Uzan, Ron Polonsky, Douwe Kiela et al.

Document expansion is a classical technique for improving retrieval quality, and is attractive since it shifts computation offline, avoiding additional query-time processing. However, when applied to modern retrievers, it has been shown to degrade performance, often introducing noise that obfuscates the discriminative signal. We recast document expansion as a document optimization problem: a language model or a vision language model is fine-tuned to transform documents into representations that better align with the expected query distribution under a target retriever, using GRPO with the retriever's ranking improvements as rewards. This approach requires only black-box access to retrieval ranks, and is applicable across single-vector, multi-vector and lexical retrievers. We evaluate our approach on code retrieval and visual document retrieval (VDR) tasks. We find that learned document transformations yield retrieval gains and in many settings enable smaller, more efficient retrievers to outperform larger ones. For example, applying document optimization to OpenAI text-embedding-3-small model improves nDCG5 on code (58.7 to 66.8) and VDR (53.3 to 57.6), even slightly surpassing the 6.5X more expensive OpenAI text-embedding-3-large model (66.3 on code; 57.0 on VDR). When retriever weights are accessible, document optimization is often competitive with fine-tuning, and in most settings their combination performs best, improving Jina-ColBERT-V2 from 55.8 to 63.3 on VDR and from 48.6 to 61.8 on code retrieval.

Sheshansh Agrawal, Thien Hang Nguyen, Douwe Kiela

Selecting the top $m$ from $n$ items via expensive $k$-wise comparisons is fundamental to settings ranging from LLM-based document reranking to crowdsourced evaluation and tournament design. Existing methods either rely on heuristics that fail to fully exploit the information each comparison reveals, or are inefficient when they do. We introduce a tournament graph framework that provides a principled foundation for $k$-wise ranking. Our key observation is that each $k$-item comparison reveals a complete tournament of $\binom{k}{2}$ pairwise preferences; aggregating these into a global preference graph and computing its transitive closure yields many additional orderings without further oracle calls. We formalize when an item's rank is certifiably determined and design a greedy query schedule that maximizes information gain towards identifying the top-$m$ items. The framework also gracefully handles non-transitive preferences (cycles induced by real-world oracles) by collapsing them into equivalence classes that yield principled tiered rankings. Applied to LLM reranking across 14 benchmarks and 5 models, our method achieves Pareto dominance over existing approaches: matching or exceeding accuracy while requiring 25-40% fewer tokens than comparable methods, and $7\times$ fewer than pairwise reranking at near-identical quality.

Tristan Thrush, Jared Moore, Miguel Monares et al.

Statements involving metalinguistic self-reference ("This paper has six sections.") are prevalent in many domains. Can current large language models (LLMs) handle such language? In this paper, we present "I am a Strange Dataset", a new dataset for addressing this question. There are two subtasks: generation and verification. In generation, models continue statements like "The penultimate word in this sentence is" (where a correct continuation is "is"). In verification, models judge the truth of statements like "The penultimate word in this sentence is sentence." (false). We also provide minimally different metalinguistic non-self-reference examples to complement the main dataset by probing for whether models can handle metalinguistic language at all. The dataset is hand-crafted by experts and validated by non-expert annotators. We test a variety of open-source LLMs (7B to 70B parameters) as well as closed-source LLMs through APIs. All models perform close to chance across both subtasks and even on the non-self-referential metalinguistic control data, though we find some steady improvement with model scale. GPT 4 is the only model to consistently do significantly better than chance, and it is still only in the 60% range, while our untrained human annotators score well in the 89-93% range. The dataset and evaluation toolkit are available at https://github.com/TristanThrush/i-am-a-strange-dataset.

Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff et al. · stanford

Kahneman & Tversky's $\textit{prospect theory}$ tells us that humans perceive random variables in a biased but well-defined manner (1992); for example, humans are famously loss-averse. We show that objectives for aligning LLMs with human feedback implicitly incorporate many of these biases -- the success of these objectives (e.g., DPO) over cross-entropy minimization can partly be ascribed to them belonging to a family of loss functions that we call $\textit{human-aware losses}$ (HALOs). However, the utility functions these methods attribute to humans still differ from those in the prospect theory literature. Using a Kahneman-Tversky model of human utility, we propose a HALO that directly maximizes the utility of generations instead of maximizing the log-likelihood of preferences, as current methods do. We call this approach KTO, and it matches or exceeds the performance of preference-based methods at scales from 1B to 30B, despite only learning from a binary signal of whether an output is desirable. More broadly, our work suggests that there is no one HALO that is universally superior; the best loss depends on the inductive biases most appropriate for a given setting, an oft-overlooked consideration.

Sheng Shen, Zhewei Yao, Douwe Kiela et al.

We demonstrate that, hidden within one-layer randomly weighted neural networks, there exist subnetworks that can achieve impressive performance, without ever modifying the weight initializations, on machine translation tasks. To find subnetworks for one-layer randomly weighted neural networks, we apply different binary masks to the same weight matrix to generate different layers. Hidden within a one-layer randomly weighted Transformer, we find that subnetworks that can achieve 29.45/17.29 BLEU on IWSLT14/WMT14. Using a fixed pre-trained embedding layer, the previously found subnetworks are smaller than, but can match 98%/92% (34.14/25.24 BLEU) of the performance of, a trained Transformer small/base on IWSLT14/WMT14. Furthermore, we demonstrate the effectiveness of larger and deeper transformers in this setting, as well as the impact of different initialization methods. We released the source code at https://github.com/sIncerass/one_layer_lottery_ticket.

Douwe Kiela, Max Bartolo, Yixin Nie et al.

We introduce Dynabench, an open-source platform for dynamic dataset creation and model benchmarking. Dynabench runs in a web browser and supports human-and-model-in-the-loop dataset creation: annotators seek to create examples that a target model will misclassify, but that another person will not. In this paper, we argue that Dynabench addresses a critical need in our community: contemporary models quickly achieve outstanding performance on benchmark tasks but nonetheless fail on simple challenge examples and falter in real-world scenarios. With Dynabench, dataset creation, model development, and model assessment can directly inform each other, leading to more robust and informative benchmarks. We report on four initial NLP tasks, illustrating these concepts and highlighting the promise of the platform, and address potential objections to dynamic benchmarking as a new standard for the field.

Christopher Potts, Zhengxuan Wu, Atticus Geiger et al.

We introduce DynaSent ('Dynamic Sentiment'), a new English-language benchmark task for ternary (positive/negative/neutral) sentiment analysis. DynaSent combines naturally occurring sentences with sentences created using the open-source Dynabench Platform, which facilities human-and-model-in-the-loop dataset creation. DynaSent has a total of 121,634 sentences, each validated by five crowdworkers, and its development and test splits are designed to produce chance performance for even the best models we have been able to develop; when future models solve this task, we will use them to create DynaSent version 2, continuing the dynamic evolution of this benchmark. Here, we report on the dataset creation effort, focusing on the steps we took to increase quality and reduce artifacts. We also present evidence that DynaSent's Neutral category is more coherent than the comparable category in other benchmarks, and we motivate training models from scratch for each round over successive fine-tuning.

Shashwat Goel, Joschka Struber, Ilze Amanda Auzina et al.

As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as ''AI Oversight''. We study how model similarity affects both aspects of AI oversight by proposing Chance Adjusted Probabilistic Agreement (CAPA): a metric for LM similarity based on overlap in model mistakes. Using CAPA, we first show that LLM-as-a-judge scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from ''weak-to-strong generalization''. As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend -- model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.

Jon Saad-Falcon, Rajan Vivek, William Berrios et al.

As language models become integral to critical workflows, assessing their behavior remains a fundamental challenge -- human evaluation is costly and noisy, while automated metrics provide only coarse, difficult-to-interpret signals. We introduce natural language unit tests, a paradigm that decomposes response quality into explicit, testable criteria, along with a unified scoring model, LMUnit, which combines multi-objective training across preferences, direct ratings, and natural language rationales. Through controlled human studies, we show this paradigm significantly improves inter-annotator agreement and enables more effective LLM development workflows. LMUnit achieves state-of-the-art performance on evaluation benchmarks (FLASK, BigGenBench) and competitive results on RewardBench. These results validate both our proposed paradigm and scoring model, suggesting a promising path forward for language model evaluation and development.

Neil Chowdhury, Franklin Wang, Sumedh Shenoy et al.

Multimodal models leverage large-scale pre-training to achieve strong but still imperfect performance on tasks such as image captioning, visual question answering, and cross-modal retrieval. In this paper, we present a simple and efficient method for correcting errors in trained contrastive image-text retrieval models with no additional training, called Nearest Neighbor Normalization (NNN). We show an improvement on retrieval metrics in both text retrieval and image retrieval for all of the contrastive models that we tested (CLIP, BLIP, ALBEF, SigLIP, BEiT) and for both of the datasets that we used (MS-COCO and Flickr30k). NNN requires a reference database, but does not require any training on this database, and can even increase the retrieval accuracy of a model after finetuning.

Nikita Vassilyev, William Berrios, Ruowang Zhang et al.

Generally capable agents must learn from experience in ways that generalize across tasks and environments. The fundamental problems of learning, including credit assignment, overfitting, forgetting, local optima, and high-variance learning signals, persist whether the learned object lies in parameter space or context space. While these challenges are well understood in classical machine learning optimization, they remain underexplored in context space, leading current methods to be fragmented and ad hoc. We present Reflective Context Learning (RCL), a unified framework for agents that learn through repeated interaction, reflection on behavior and failure modes, and iterative updates to context. In RCL, reflection converts trajectories and current context into a directional update signal analogous to gradients, while mutation applies that signal to improve future behavior in context space. We recast recent context-optimization approaches as instances of this shared learning problem and systematically extend them with classical optimization primitives, including batching, improved credit-assignment signal, auxiliary losses, failure replay, and grouped rollouts for variance reduction. On AppWorld, BrowseComp+, and RewardBench2, these primitives improve over strong baselines, with their relative importance shifting across task regimes. We further analyze robustness to initialization, the effects of batch size, sampling and curriculum strategy, optimizer-state variants, and the impact of allocating stronger or weaker models to different optimization components. Our results suggest that learning through context updates should be treated not as a set of isolated algorithms, but as an optimization problem whose mechanisms can be studied systematically and improved through transferable principles.

Max Bartolo, Tristan Thrush, Sebastian Riedel et al.

In Dynamic Adversarial Data Collection (DADC), human annotators are tasked with finding examples that models struggle to predict correctly. Models trained on DADC-collected training data have been shown to be more robust in adversarial and out-of-domain settings, and are considerably harder for humans to fool. However, DADC is more time-consuming than traditional data collection and thus more costly per annotated example. In this work, we examine whether we can maintain the advantages of DADC, without incurring the additional cost. To that end, we introduce Generative Annotation Assistants (GAAs), generator-in-the-loop models that provide real-time suggestions that annotators can either approve, modify, or reject entirely. We collect training datasets in twenty experimental settings and perform a detailed analysis of this approach for the task of extractive question answering (QA) for both standard and adversarial data collection. We demonstrate that GAAs provide significant efficiency benefits with over a 30% annotation speed-up, while leading to over a 5x improvement in model fooling rates. In addition, we find that using GAA-assisted training data leads to higher downstream model performance on a variety of question answering tasks over adversarial data collection.

Amanpreet Singh, Ronghang Hu, Vedanuj Goswami et al.

State-of-the-art vision and vision-and-language models rely on large-scale visio-linguistic pretraining for obtaining good performance on a variety of downstream tasks. Generally, such models are often either cross-modal (contrastive) or multi-modal (with earlier fusion) but not both; and they often only target specific modalities or tasks. A promising direction would be to use a single holistic universal model, as a "foundation", that targets all modalities at once -- a true vision and language foundation model should be good at vision tasks, language tasks, and cross- and multi-modal vision and language tasks. We introduce FLAVA as such a model and demonstrate impressive performance on a wide range of 35 tasks spanning these target modalities.

Eric Wallace, Adina Williams, Robin Jia et al.

To create models that are robust across a wide range of test inputs, training datasets should include diverse examples that span numerous phenomena. Dynamic adversarial data collection (DADC), where annotators craft examples that challenge continually improving models, holds promise as an approach for generating such diverse training sets. Prior work has shown that running DADC over 1-3 rounds can help models fix some error types, but it does not necessarily lead to better generalization beyond adversarial test data. We argue that running DADC over many rounds maximizes its training-time benefits, as the different rounds can together cover many of the task-relevant phenomena. We present the first study of longer-term DADC, where we collect 20 rounds of NLI examples for a small set of premise paragraphs, with both adversarial and non-adversarial approaches. Models trained on DADC examples make 26% fewer errors on our expert-curated test set compared to models trained on non-adversarial data. Our analysis shows that DADC yields examples that are more difficult, more lexically and syntactically diverse, and contain fewer annotation artifacts compared to non-adversarial examples.

Sasha Sheng, Amanpreet Singh, Vedanuj Goswami et al.

Performance on the most commonly used Visual Question Answering dataset (VQA v2) is starting to approach human accuracy. However, in interacting with state-of-the-art VQA models, it is clear that the problem is far from being solved. In order to stress test VQA models, we benchmark them against human-adversarial examples. Human subjects interact with a state-of-the-art VQA model, and for each image in the dataset, attempt to find a question where the model's predicted answer is incorrect. We find that a wide range of state-of-the-art models perform poorly when evaluated on these examples. We conduct an extensive analysis of the collected adversarial examples and provide guidance on future research directions. We hope that this Adversarial VQA (AdVQA) benchmark can help drive progress in the field and advance the state of the art.

Divyansh Kaushik, Douwe Kiela, Zachary C. Lipton et al.

In adversarial data collection (ADC), a human workforce interacts with a model in real time, attempting to produce examples that elicit incorrect predictions. Researchers hope that models trained on these more challenging datasets will rely less on superficial patterns, and thus be less brittle. However, despite ADC's intuitive appeal, it remains unclear when training on adversarial datasets produces more robust models. In this paper, we conduct a large-scale controlled study focused on question answering, assigning workers at random to compose questions either (i) adversarially (with a model in the loop); or (ii) in the standard fashion (without a model). Across a variety of models and datasets, we find that models trained on adversarial data usually perform better on other adversarial datasets but worse on a diverse collection of out-of-domain evaluation sets. Finally, we provide a qualitative analysis of adversarial (vs standard) data, identifying key differences and offering guidance for future research.

Ethan Perez, Douwe Kiela, Kyunghyun Cho

Pretrained language models (LMs) perform well on many tasks even when learning from a few examples, but prior work uses many held-out examples to tune various aspects of learning, such as hyperparameters, training objectives, and natural language templates ("prompts"). Here, we evaluate the few-shot ability of LMs when such held-out examples are unavailable, a setting we call true few-shot learning. We test two model selection criteria, cross-validation and minimum description length, for choosing LM prompts and hyperparameters in the true few-shot setting. On average, both marginally outperform random selection and greatly underperform selection based on held-out examples. Moreover, selection criteria often prefer models that perform significantly worse than randomly-selected ones. We find similar results even when taking into account our uncertainty in a model's true performance during selection, as well as when varying the amount of computation and number of examples used for selection. Overall, our findings suggest that prior work significantly overestimated the true few-shot ability of LMs given the difficulty of few-shot model selection.

Zhiyi Ma, Kawin Ethayarajh, Tristan Thrush et al.

We introduce Dynaboard, an evaluation-as-a-service framework for hosting benchmarks and conducting holistic model comparison, integrated with the Dynabench platform. Our platform evaluates NLP models directly instead of relying on self-reported metrics or predictions on a single dataset. Under this paradigm, models are submitted to be evaluated in the cloud, circumventing the issues of reproducibility, accessibility, and backwards compatibility that often hinder benchmarking in NLP. This allows users to interact with uploaded models in real time to assess their quality, and permits the collection of additional metrics such as memory use, throughput, and robustness, which -- despite their importance to practitioners -- have traditionally been absent from leaderboards. On each task, models are ranked according to the Dynascore, a novel utility-based aggregation of these statistics, which users can customize to better reflect their preferences, placing more/less weight on a particular axis of evaluation or dataset. As state-of-the-art NLP models push the limits of traditional benchmarks, Dynaboard offers a standardized solution for a more diverse and comprehensive evaluation of model quality.

Max Bartolo, Tristan Thrush, Robin Jia et al.

Despite recent progress, state-of-the-art question answering models remain vulnerable to a variety of adversarial attacks. While dynamic adversarial data collection, in which a human annotator tries to write examples that fool a model-in-the-loop, can improve model robustness, this process is expensive which limits the scale of the collected data. In this work, we are the first to use synthetic adversarial data generation to make question answering models more robust to human adversaries. We develop a data generation pipeline that selects source passages, identifies candidate answers, generates questions, then finally filters or re-labels them to improve quality. Using this approach, we amplify a smaller human-written adversarial dataset to a much larger set of synthetic question-answer pairs. By incorporating our synthetic data, we improve the state-of-the-art on the AdversarialQA dataset by 3.7F1 and improve model generalisation on nine of the twelve MRQA datasets. We further conduct a novel human-in-the-loop evaluation to show that our models are considerably more robust to new human-written adversarial examples: crowdworkers can fool our model only 8.8% of the time on average, compared to 17.6% for a model trained without synthetic data.

Shir Gur, Natalia Neverova, Chris Stauffer et al.

Recent advances in using retrieval components over external knowledge sources have shown impressive results for a variety of downstream tasks in natural language processing. Here, we explore the use of unstructured external knowledge sources of images and their corresponding captions for improving visual question answering (VQA). First, we train a novel alignment model for embedding images and captions in the same space, which achieves substantial improvement in performance on image-caption retrieval w.r.t. similar methods. Second, we show that retrieval-augmented multi-modal transformers using the trained alignment model improve results on VQA over strong baselines. We further conduct extensive experiments to establish the promise of this approach, and examine novel applications for inference time such as hot-swapping indices.

Chuan Guo, Alexandre Sablayrolles, Hervé Jégou et al.

We propose the first general-purpose gradient-based attack against transformer models. Instead of searching for a single adversarial example, we search for a distribution of adversarial examples parameterized by a continuous-valued matrix, hence enabling gradient-based optimization. We empirically demonstrate that our white-box attack attains state-of-the-art attack performance on a variety of natural language tasks. Furthermore, we show that a powerful black-box transfer attack, enabled by sampling from the adversarial distribution, matches or exceeds existing methods, while only requiring hard-label outputs.

Kurt Shuster, Spencer Poff, Moya Chen et al.

Despite showing increasingly human-like conversational abilities, state-of-the-art dialogue models often suffer from factual incorrectness and hallucination of knowledge (Roller et al., 2020). In this work we explore the use of neural-retrieval-in-the-loop architectures - recently shown to be effective in open-domain QA (Lewis et al., 2020b; Izacard and Grave, 2020) - for knowledge-grounded dialogue, a task that is arguably more challenging as it requires querying based on complex multi-turn dialogue context and generating conversationally coherent responses. We study various types of architectures with multiple components - retrievers, rankers, and encoder-decoders - with the goal of maximizing knowledgeability while retaining conversational ability. We demonstrate that our best models obtain state-of-the-art performance on two knowledge-grounded conversational tasks. The models exhibit open-domain conversational capabilities, generalize effectively to scenarios not within the training data, and, as verified by human evaluations, substantially reduce the well-known problem of knowledge hallucination in state-of-the-art chatbots.

Koustuv Sinha, Robin Jia, Dieuwke Hupkes et al.

A possible explanation for the impressive performance of masked language model (MLM) pre-training is that such models have learned to represent the syntactic structures prevalent in classical NLP pipelines. In this paper, we propose a different explanation: MLMs succeed on downstream tasks almost entirely due to their ability to model higher-order word co-occurrence statistics. To demonstrate this, we pre-train MLMs on sentences with randomly shuffled word order, and show that these models still achieve high accuracy after fine-tuning on many downstream tasks -- including on tasks specifically designed to be challenging for models that ignore word order. Our models perform surprisingly well according to some parametric syntactic probes, indicating possible deficiencies in how we test representations for syntactic information. Overall, our results show that purely distributional information largely explains the success of pre-training, and underscore the importance of curating challenging evaluation datasets that require deeper linguistic knowledge.

Kalesha Bullard, Douwe Kiela, Franziska Meier et al.

Effective communication is an important skill for enabling information exchange in multi-agent settings and emergent communication is now a vibrant field of research, with common settings involving discrete cheap-talk channels. Since, by definition, these settings involve arbitrary encoding of information, typically they do not allow for the learned protocols to generalize beyond training partners. In contrast, in this work, we present a novel problem setting and the Quasi-Equivalence Discovery (QED) algorithm that allows for zero-shot coordination (ZSC), i.e., discovering protocols that can generalize to independently trained agents. Real world problem settings often contain costly communication channels, e.g., robots have to physically move their limbs, and a non-uniform distribution over intents. We show that these two factors lead to unique optimal ZSC policies in referential games, where agents use the energy cost of the messages to communicate intent. Other-Play was recently introduced for learning optimal ZSC policies, but requires prior access to the symmetries of the problem. Instead, QED can iteratively discovers the symmetries in this setting and converges to the optimal ZSC policy.

Ethan Perez, Douwe Kiela, Kyunghyun Cho

We introduce a method to determine if a certain capability helps to achieve an accurate model of given data. We view labels as being generated from the inputs by a program composed of subroutines with different capabilities, and we posit that a subroutine is useful if and only if the minimal program that invokes it is shorter than the one that does not. Since minimum program length is uncomputable, we instead estimate the labels' minimum description length (MDL) as a proxy, giving us a theoretically-grounded method for analyzing dataset characteristics. We call the method Rissanen Data Analysis (RDA) after the father of MDL, and we showcase its applicability on a wide variety of settings in NLP, ranging from evaluating the utility of generating subquestions before answering a question, to analyzing the value of rationales and explanations, to investigating the importance of different parts of speech, and uncovering dataset gender bias.

Bertie Vidgen, Tristan Thrush, Zeerak Waseem et al.

We present a human-and-model-in-the-loop process for dynamically generating datasets and training better performing and more robust hate detection models. We provide a new dataset of ~40,000 entries, generated and labelled by trained annotators over four rounds of dynamic data creation. It includes ~15,000 challenging perturbations and each hateful entry has fine-grained labels for the type and target of hate. Hateful entries make up 54% of the dataset, which is substantially higher than comparable datasets. We show that model performance is substantially improved using this approach. Models trained on later rounds of data collection perform better on test sets and are harder for annotators to trick. They also perform better on HateCheck, a suite of functional tests for online hate detection. We provide the code, dataset and annotation guidelines for other researchers to use. Accepted at ACL 2021.

Sheng Shen, Alexei Baevski, Ari S. Morcos et al.

We demonstrate that transformers obtain impressive performance even when some of the layers are randomly initialized and never updated. Inspired by old and well-established ideas in machine learning, we explore a variety of non-linear "reservoir" layers interspersed with regular transformer layers, and show improvements in wall-clock compute time until convergence, as well as overall performance, on various machine translation and (masked) language modelling tasks.

Yixin Nie, Mary Williamson, Mohit Bansal et al.

To quantify how well natural language understanding models can capture consistency in a general conversation, we introduce the DialoguE COntradiction DEtection task (DECODE) and a new conversational dataset containing both human-human and human-bot contradictory dialogues. We then compare a structured utterance-based approach of using pre-trained Transformer models for contradiction detection with the typical unstructured approach. Results reveal that: (i) our newly collected dataset is notably more effective at providing supervision for the dialogue contradiction detection task than existing NLI data including those aimed to cover the dialogue domain; (ii) the structured utterance-based approach is more robust and transferable on both analysis and out-of-distribution dialogues than its unstructured counterpart. We also show that our best contradiction detection model correlates well with human judgments and further provide evidence for its usage in both automatically evaluating and improving the consistency of state-of-the-art generative chatbots.

Grusha Prasad, Yixin Nie, Mohit Bansal et al.

Given the increasingly prominent role NLP models (will) play in our lives, it is important for human expectations of model behavior to align with actual model behavior. Using Natural Language Inference (NLI) as a case study, we investigate the extent to which human-generated explanations of models' inference decisions align with how models actually make these decisions. More specifically, we define three alignment metrics that quantify how well natural language explanations align with model sensitivity to input words, as measured by integrated gradients. Then, we evaluate eight different models (the base and large versions of BERT, RoBERTa and ELECTRA, as well as anRNN and bag-of-words model), and find that the BERT-base model has the highest alignment with human-generated explanations, for all alignment metrics. Focusing in on transformers, we find that the base versions tend to have higher alignment with human-generated explanations than their larger counterparts, suggesting that increasing the number of model parameters leads, in some cases, to worse alignment with human explanations. Finally, we find that a model's alignment with human explanations is not predicted by the model's accuracy, suggesting that accuracy and alignment are complementary ways to evaluate models.

Kalesha Bullard, Franziska Meier, Douwe Kiela et al.

Effective communication is an important skill for enabling information exchange and cooperation in multi-agent settings. Indeed, emergent communication is now a vibrant field of research, with common settings involving discrete cheap-talk channels. One limitation of this setting is that it does not allow for the emergent protocols to generalize beyond the training partners. Furthermore, so far emergent communication has primarily focused on the use of symbolic channels. In this work, we extend this line of work to a new modality, by studying agents that learn to communicate via actuating their joints in a 3D environment. We show that under realistic assumptions, a non-uniform distribution of intents and a common-knowledge energy cost, these agents can find protocols that generalize to novel partners. We also explore and analyze specific difficulties associated with finding these solutions in practice. Finally, we propose and evaluate initial training improvements to address these challenges, involving both specific training curricula and providing the latent feature that can be coordinated on during training.