Adina Williams

Hugo Touvron, Louis Martin, Kevin Stone et al. · meta-ai

In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases. Our models outperform open-source chat models on most benchmarks we tested, and based on our human evaluations for helpfulness and safety, may be a suitable substitute for closed-source models. We provide a detailed description of our approach to fine-tuning and safety improvements of Llama 2-Chat in order to enable the community to build on our work and contribute to the responsible development of LLMs.

David Esiobu, Xiaoqing Tan, Saghar Hosseini et al. · meta-ai

As generative large language models (LLMs) grow more performant and prevalent, we must develop comprehensive enough tools to measure and improve their fairness. Different prompt-based datasets can be used to measure social bias across multiple text domains and demographic axes, meaning that testing LLMs on more datasets can potentially help us characterize their biases more fully, and better ensure equal and equitable treatment of marginalized demographic groups. In this work, our focus is two-fold: (1) Benchmarking: a comparison of 6 different prompt-based bias and toxicity metrics across 12 demographic axes and 5 families of generative LLMs. Out of those 6 metrics, AdvPromptSet and HolisticBiasR are novel datasets proposed in the paper. The comparison of those benchmarks gives us insights about the bias and toxicity of the compared models. Therefore, we explore the frequency of demographic terms in common LLM pre-training corpora and how this may relate to model biases. (2) Mitigation: we conduct a comprehensive study of how well 3 bias/toxicity mitigation techniques perform across our suite of measurements. ROBBIE aims to provide insights for practitioners while deploying a model, emphasizing the need to not only measure potential harms, but also understand how they arise by characterizing the data, mitigate harms once found, and balance any trade-offs. We open-source our analysis code in hopes of encouraging broader measurements of bias in future LLMs.

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.

Eric Michael Smith, Melissa Hall, Melanie Kambadur et al. · meta-ai

As language models grow in popularity, it becomes increasingly important to clearly measure all possible markers of demographic identity in order to avoid perpetuating existing societal harms. Many datasets for measuring bias currently exist, but they are restricted in their coverage of demographic axes and are commonly used with preset bias tests that presuppose which types of biases models can exhibit. In this work, we present a new, more inclusive bias measurement dataset, HolisticBias, which includes nearly 600 descriptor terms across 13 different demographic axes. HolisticBias was assembled in a participatory process including experts and community members with lived experience of these terms. These descriptors combine with a set of bias measurement templates to produce over 450,000 unique sentence prompts, which we use to explore, identify, and reduce novel forms of bias in several generative models. We demonstrate that HolisticBias is effective at measuring previously undetectable biases in token likelihoods from language models, as well as in an offensiveness classifier. We will invite additions and amendments to the dataset, which we hope will serve as a basis for more easy-to-use and standardized methods for evaluating bias in NLP models.

Benjamin Muller, Belen Alastruey, Prangthip Hansanti et al. · meta-ai, uw

Gender biases in language generation systems are challenging to mitigate. One possible source for these biases is gender representation disparities in the training and evaluation data. Despite recent progress in documenting this problem and many attempts at mitigating it, we still lack shared methodology and tooling to report gender representation in large datasets. Such quantitative reporting will enable further mitigation, e.g., via data augmentation. This paper describes the Gender-GAP Pipeline (for Gender-Aware Polyglot Pipeline), an automatic pipeline to characterize gender representation in large-scale datasets for 55 languages. The pipeline uses a multilingual lexicon of gendered person-nouns to quantify the gender representation in text. We showcase it to report gender representation in WMT training data and development data for the News task, confirming that current data is skewed towards masculine representation. Having unbalanced datasets may indirectly optimize our systems towards outperforming one gender over the others. We suggest introducing our gender quantification pipeline in current datasets and, ideally, modifying them toward a balanced representation.

Koustuv Sinha, Amirhossein Kazemnejad, Siva Reddy et al. · meta-ai, mila

Transformer language models encode the notion of word order using positional information. Most commonly, this positional information is represented by absolute position embeddings (APEs), that are learned from the pretraining data. However, in natural language, it is not absolute position that matters, but relative position, and the extent to which APEs can capture this type of information has not been investigated. In this work, we observe that models trained with APE over-rely on positional information to the point that they break-down when subjected to sentences with shifted position information. Specifically, when models are subjected to sentences starting from a non-zero position (excluding the effect of priming), they exhibit noticeably degraded performance on zero to full-shot tasks, across a range of model families and model sizes. Our findings raise questions about the efficacy of APEs to model the relativity of position information, and invite further introspection on the sentence and word order processing strategies employed by these models.

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.

Koustuv Sinha, Jon Gauthier, Aaron Mueller et al. · meta-ai, mit

Targeted syntactic evaluations of language models ask whether models show stable preferences for syntactically acceptable content over minimal-pair unacceptable inputs. Most targeted syntactic evaluation datasets ask models to make these judgements with just a single context-free sentence as input. This does not match language models' training regime, in which input sentences are always highly contextualized by the surrounding corpus. This mismatch raises an important question: how robust are models' syntactic judgements in different contexts? In this paper, we investigate the stability of language models' performance on targeted syntactic evaluations as we vary properties of the input context: the length of the context, the types of syntactic phenomena it contains, and whether or not there are violations of grammaticality. We find that model judgements are generally robust when placed in randomly sampled linguistic contexts. However, they are substantially unstable for contexts containing syntactic structures matching those in the critical test content. Among all tested models (GPT-2 and five variants of OPT), we significantly improve models' judgements by providing contexts with matching syntactic structures, and conversely significantly worsen them using unacceptable contexts with matching but violated syntactic structures. This effect is amplified by the length of the context, except for unrelated inputs. We show that these changes in model performance are not explainable by simple features matching the context and the test inputs, such as lexical overlap and dependency overlap. This sensitivity to highly specific syntactic features of the context can only be explained by the models' implicit in-context learning abilities.

Alex Warstadt, Leshem Choshen, Aaron Mueller et al. · harvard, ibm-research

We present the call for papers for the BabyLM Challenge: Sample-efficient pretraining on a developmentally plausible corpus. This shared task is intended for participants with an interest in small scale language modeling, human language acquisition, low-resource NLP, and cognitive modeling. In partnership with CoNLL and CMCL, we provide a platform for approaches to pretraining with a limited-size corpus sourced from data inspired by the input to children. The task has three tracks, two of which restrict the training data to pre-released datasets of 10M and 100M words and are dedicated to explorations of approaches such as architectural variations, self-supervised objectives, or curriculum learning. The final track only restricts the amount of text used, allowing innovation in the choice of the data, its domain, and even its modality (i.e., data from sources other than text is welcome). We will release a shared evaluation pipeline which scores models on a variety of benchmarks and tasks, including targeted syntactic evaluations and natural language understanding.

Josef Valvoda, Naomi Saphra, Jonathan Rawski et al. · cambridge, cmu

Recombining known primitive concepts into larger novel combinations is a quintessentially human cognitive capability. Whether large neural models in NLP can acquire this ability while learning from data is an open question. In this paper, we investigate this problem from the perspective of formal languages. We use deterministic finite-state transducers to make an unbounded number of datasets with controllable properties governing compositionality. By randomly sampling over many transducers, we explore which of their properties contribute to learnability of a compositional relation by a neural network. We find that the models either learn the relations completely or not at all. The key is transition coverage, setting a soft learnability limit at 400 examples per transition.

Pranjal Aggarwal, Marjan Ghazvininejad, Seungone Kim et al. · meta-ai

The ability to precisely derive mathematical objects is a core requirement for downstream STEM applications, including mathematics, physics, and chemistry, where reasoning must culminate in formally structured expressions. Yet, current LM evaluations of mathematical and scientific reasoning rely heavily on simplified answer formats such as numerical values or multiple choice options due to the convenience of automated assessment. In this paper we provide three contributions for improving reasoning over mathematical objects: (i) we build and release training data and benchmarks for deriving mathematical objects, the Principia suite; (ii) we provide training recipes with strong LLM-judges and verifiers, where we show that on-policy judge training boosts performance; (iii) we show how on-policy training can also be used to scale test-time compute via aggregation. We find that strong LMs such as Qwen3-235B and o3 struggle on Principia, while our training recipes can bring significant improvements over different LLM backbones, while simultaneously improving results on existing numerical and MCQA tasks, demonstrating cross-format generalization of reasoning abilities.

Karolina Stańczak, Kevin Du, Adina Williams et al. · eth-zurich

How much meaning influences gender assignment across languages is an active area of research in linguistics and cognitive science. We can view current approaches as aiming to determine where gender assignment falls on a spectrum, from being fully arbitrarily determined to being largely semantically determined. For the latter case, there is a formulation of the neo-Whorfian hypothesis, which claims that even inanimate noun gender influences how people conceive of and talk about objects (using the choice of adjective used to modify inanimate nouns as a proxy for meaning). We offer a novel, causal graphical model that jointly represents the interactions between a noun's grammatical gender, its meaning, and adjective choice. In accordance with past results, we find a significant relationship between the gender of nouns and the adjectives that modify them. However, when we control for the meaning of the noun, the relationship between grammatical gender and adjective choice is near zero and insignificant.

Arjun Subramonian, Adina Williams, Maximilian Nickel et al. · meta-ai

The expressive power of graph neural networks is usually measured by comparing how many pairs of graphs or nodes an architecture can possibly distinguish as non-isomorphic to those distinguishable by the $k$-dimensional Weisfeiler-Leman ($k$-WL) test. In this paper, we uncover misalignments between graph machine learning practitioners' conceptualizations of expressive power and $k$-WL through a systematic analysis of the reliability and validity of $k$-WL. We conduct a survey ($n = 18$) of practitioners to surface their conceptualizations of expressive power and their assumptions about $k$-WL. In contrast to practitioners' beliefs, our analysis (which draws from graph theory and benchmark auditing) reveals that $k$-WL does not guarantee isometry, can be irrelevant to real-world graph tasks, and may not promote generalization or trustworthiness. We argue for extensional definitions and measurement of expressive power based on benchmarks. We further contribute guiding questions for constructing such benchmarks, which is critical for graph machine learning practitioners to develop and transparently communicate our understandings of expressive power.

Kaiser Sun, Adina Williams, Dieuwke Hupkes

NLP models have progressed drastically in recent years, according to numerous datasets proposed to evaluate performance. Questions remain, however, about how particular dataset design choices may impact the conclusions we draw about model capabilities. In this work, we investigate this question in the domain of compositional generalization. We examine the performance of six modeling approaches across 4 datasets, split according to 8 compositional splitting strategies, ranking models by 18 compositional generalization splits in total. Our results show that: i) the datasets, although all designed to evaluate compositional generalization, rank modeling approaches differently; ii) datasets generated by humans align better with each other than they with synthetic datasets, or than synthetic datasets among themselves; iii) generally, whether datasets are sampled from the same source is more predictive of the resulting model ranking than whether they maintain the same interpretation of compositionality; and iv) which lexical items are used in the data can strongly impact conclusions. Overall, our results demonstrate that much work remains to be done when it comes to assessing whether popular evaluation datasets measure what they intend to measure, and suggest that elucidating more rigorous standards for establishing the validity of evaluation sets could benefit the field.

Melissa Hall, Candace Ross, Adina Williams et al.

The unprecedented photorealistic results achieved by recent text-to-image generative systems and their increasing use as plug-and-play content creation solutions make it crucial to understand their potential biases. In this work, we introduce three indicators to evaluate the realism, diversity and prompt-generation consistency of text-to-image generative systems when prompted to generate objects from across the world. Our indicators complement qualitative analysis of the broader impact of such systems by enabling automatic and efficient benchmarking of geographic disparities, an important step towards building responsible visual content creation systems. We use our proposed indicators to analyze potential geographic biases in state-of-the-art visual content creation systems and find that: (1) models have less realism and diversity of generations when prompting for Africa and West Asia than Europe, (2) prompting with geographic information comes at a cost to prompt-consistency and diversity of generated images, and (3) models exhibit more region-level disparities for some objects than others. Perhaps most interestingly, our indicators suggest that progress in image generation quality has come at the cost of real-world geographic representation. Our comprehensive evaluation constitutes a crucial step towards ensuring a positive experience of visual content creation for everyone.

Angelika Romanou, Mark Ibrahim, Candace Ross et al.

Existing evaluation methods largely rely on clean, static benchmarks, which can overestimate true model performance by failing to capture the noise and variability inherent in real-world user inputs. This is especially true for language models, which can face human-generated text queries containing mistakes, typos, or alternative ways of phrasing the same question. In this work, we introduce a theoretical framework for quantifying model sensitivity to prompt variants, or brittleness, that can enable us to disentangle data-induced difficulty from prompt-related variability. Using this framework, we design a novel evaluation pipeline, Brittlebench, to holistically evaluate the sensitivity of frontier models. We apply semantics-preserving perturbations to a suite of popular benchmarks, and observe model performance to degrade as much as 12%. However, these perturbations do not affect all models equally: even a single perturbation alters the relative ranking of models in 63% of cases, impacting conclusions about comparative model performance. Decomposing the total variance of both state-of-the-art open-weight and commercial models, we find that semantics-preserving input perturbations can account for up to half of the performance variance for a given model. Brittlebench highlights the need for more robust evaluations and models, and allows us to systematically understand model brittleness.

Niklas Nolte, Ouail Kitouni, Adina Williams et al.

Despite their remarkable success in language modeling, transformers trained to predict the next token in a sequence struggle with long-term planning. This limitation is particularly evident in tasks requiring foresight to plan multiple steps ahead such as maze navigation. The standard next single token prediction objective, however, offers no explicit mechanism to predict multiple steps ahead - or revisit the path taken so far. Consequently, in this work we study whether explicitly predicting multiple steps ahead (and backwards) can improve transformers' maze navigation. We train parameter-matched transformers from scratch, under identical settings, to navigate mazes of varying types and sizes with standard next token prediction and MLM-U, an objective explicitly predicting multiple steps ahead and backwards. We find that MLM-U considerably improves transformers' ability to navigate mazes compared to standard next token prediction across maze types and complexities. We also find MLM-U training is 4x more sample efficient and converges 2x faster in terms of GPU training hours relative to next token training. Finally, for more complex mazes we find MLM-U benefits from scaling to larger transformers. Remarkably, we find transformers trained with MLM-U outperform larger transformers trained with next token prediction using additional supervision from A* search traces. We hope these findings underscore the promise of learning objectives to advance transformers' capacity for long-term planning. The code can be found at https://github.com/facebookresearch/maze_navigation_MLMU

Anaelia Ovalle, Candace Ross, Sebastian Ruder et al.

Large language models demonstrate strong reasoning capabilities through chain-of-thought prompting, but whether this reasoning quality transfers across languages remains underexplored. We introduce a human-validated framework to evaluate whether model-generated reasoning traces logically support their conclusions across languages. Analyzing 65k reasoning traces from GlobalMMLU questions across 6 languages and 6 frontier models, we uncover a critical blind spot: while models achieve high task accuracy, their reasoning can fail to support their conclusions. Reasoning traces in non-Latin scripts show at least twice as much misalignment between their reasoning and conclusions than those in Latin scripts. We develop an error taxonomy through human annotation to characterize these failures, finding they stem primarily from evidential errors (unsupported claims, ambiguous facts) followed by illogical reasoning steps. Our findings demonstrate that current multilingual evaluation practices provide an incomplete picture of model reasoning capabilities and highlight the need for reasoning-aware evaluation frameworks.

Florian Bordes, Candace Ross, Justine T Kao et al.

The rapid proliferation of benchmarks has created significant challenges in reproducibility, transparency, and informed decision-making. However, unlike datasets and models -- which benefit from structured documentation frameworks like Datasheets and Model Cards -- evaluation methodologies lack systematic documentation standards. We introduce Eval Factsheets, a structured, descriptive framework for documenting AI system evaluations through a comprehensive taxonomy and questionnaire-based approach. Our framework organizes evaluation characteristics across five fundamental dimensions: Context (Who made the evaluation and when?), Scope (What does it evaluate?), Structure (With what the evaluation is built?), Method (How does it work?) and Alignment (In what ways is it reliable/valid/robust?). We implement this taxonomy as a practical questionnaire spanning five sections with mandatory and recommended documentation elements. Through case studies on multiple benchmarks, we demonstrate that Eval Factsheets effectively captures diverse evaluation paradigms -- from traditional benchmarks to LLM-as-judge methodologies -- while maintaining consistency and comparability. We hope Eval Factsheets are incorporated into both existing and newly released evaluation frameworks and lead to more transparency and reproducibility.

Candace Ross, Florian Bordes, Adina Williams et al.

Multimodal language models possess a remarkable ability to handle an open-vocabulary's worth of objects. Yet the best models still suffer from hallucinations when reasoning about scenes in the real world, revealing a gap between their seemingly strong performance on existing perception benchmarks that are saturating and their reasoning in the real world. To address this gap, we build a novel benchmark of in-the-wild scenes that we call Common-O. With more than 10.5k examples using exclusively new images not found in web training data to avoid contamination, Common-O goes beyond just perception, inspired by cognitive tests for humans, to probe reasoning across scenes by asking "what's in common?". We evaluate leading multimodal language models, including models specifically trained to perform chain-of-thought reasoning. We find that perceiving objects in single images is tractable for most models, yet reasoning across scenes is very challenging even for the best models, including reasoning models. Despite saturating many leaderboards focusing on perception, the best performing model only achieves 35% on Common-O -- and on Common-O Complex, consisting of more complex scenes, the best model achieves only 1%. Curiously, we find models are more prone to hallucinate when similar objects are present in the scene, suggesting models may be relying on object co-occurrence seen during training. Among the models we evaluated, we found scale can provide modest improvements while models explicitly trained with multi-image inputs show bigger improvements, suggesting scaled multi-image training may offer promise. We make our benchmark publicly available to spur research into the challenge of hallucination when reasoning across scenes.

Da Ju, Hagen Blix, Adina Williams

Recent improvement in large language model performance have, in all likelihood, been accompanied by improvement in how well they can approximate the distribution of their training data. In this work, we explore the following question: which properties of text domains do LLMs faithfully approximate, and how well do they do so? Applying observational approaches familiar from corpus linguistics, we prompt a commonly used, opensource LLM to regenerate text from two domains of permissively licensed English text which are often contained in LLM training data -- Wikipedia and news text. This regeneration paradigm allows us to investigate whether LLMs can faithfully match the original human text domains in a fairly semantically-controlled setting. We investigate varying levels of syntactic abstraction, from more simple properties like sentence length, and article readability, to more complex and higher order properties such as dependency tag distribution, parse depth, and parse complexity. We find that the majority of the regenerated distributions show a shifted mean, a lower standard deviation, and a reduction of the long tail, as compared to the human originals.

Alex Warstadt, Aaron Mueller, Leshem Choshen et al. · ibm-research

Children can acquire language from less than 100 million words of input. Large language models are far less data-efficient: they typically require 3 or 4 orders of magnitude more data and still do not perform as well as humans on many evaluations. These intensive resource demands limit the ability of researchers to train new models and use existing models as developmentally plausible cognitive models. The BabyLM Challenge is a communal effort in which participants compete to optimize language model training on a fixed data budget. Submissions are compared on various evaluation tasks targeting grammatical ability, downstream task performance, and generalization. Participants can submit to up to three tracks with progressively looser data restrictions. From over 30 submissions, we extract concrete recommendations on how best to train data-efficient language models, and on where future efforts should (and perhaps should not) focus. The winning submissions using the LTG-BERT architecture (Samuel et al., 2023) outperformed models trained on trillions of words. Other submissions achieved strong results through training on shorter input sequences or training a student model on a pretrained teacher. Curriculum learning attempts, which accounted for a large number of submissions, were largely unsuccessful, though some showed modest improvements.

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.

Koustuv Sinha, Prasanna Parthasarathi, Joelle Pineau et al.

Recent investigations into the inner-workings of state-of-the-art large-scale pre-trained Transformer-based Natural Language Understanding (NLU) models indicate that they appear to know humanlike syntax, at least to some extent. We provide novel evidence that complicates this claim: we find that state-of-the-art Natural Language Inference (NLI) models assign the same labels to permuted examples as they do to the original, i.e. they are largely invariant to random word-order permutations. This behavior notably differs from that of humans; we struggle with ungrammatical sentences. To measure the severity of this issue, we propose a suite of metrics and investigate which properties of particular permutations lead models to be word-order invariant. In the MNLI dataset, for example, we find almost all (98.7%) examples contain at least one permutation which elicits the gold label. Models are sometimes even able to assign gold labels to permutations that they originally failed to predict correctly. We provide a comprehensive empirical evaluation of this phenomenon, and further show that this issue exists for both Transformers and pre-Transformer RNN / ConvNet based encoders, as well as across multiple languages (English and Mandarin Chinese). Our code and data are available at https://github.com/facebookresearch/unlu.

Hannah Rose Kirk, Alexander Whitefield, Paul Röttger et al. · oxford

Human feedback is central to the alignment of Large Language Models (LLMs). However, open questions remain about methods (how), domains (where), people (who) and objectives (to what end) of feedback processes. To navigate these questions, we introduce PRISM, a dataset that maps the sociodemographics and stated preferences of 1,500 diverse participants from 75 countries, to their contextual preferences and fine-grained feedback in 8,011 live conversations with 21 LLMs. With PRISM, we contribute (i) wider geographic and demographic participation in feedback; (ii) census-representative samples for two countries (UK, US); and (iii) individualised ratings that link to detailed participant profiles, permitting personalisation and attribution of sample artefacts. We target subjective and multicultural perspectives on value-laden and controversial issues, where we expect interpersonal and cross-cultural disagreement. We use PRISM in three case studies to demonstrate the need for careful consideration of which humans provide what alignment data.

Da Ju, Karen Ulrich, Adina Williams

People tend to use language to mention surprising properties of events: for example, when a banana is blue, we are more likely to mention color than when it is yellow. This fact is taken to suggest that yellowness is somehow a typical feature of bananas, and blueness is exceptional. Similar to how a yellow color is typical of bananas, there may also be genders that are typical of occupations. In this work, we explore this question using information theoretic techniques coupled with corpus statistic analysis. In two distinct large corpora, we do not find strong evidence that occupations and gender display the same patterns of mentioning as do bananas and color. Instead, we find that gender mentioning is correlated with femaleness of occupation in particular, suggesting perhaps that woman-dominated occupations are seen as somehow ``more gendered'' than male-dominated ones, and thereby they encourage more gender mentioning overall.

Oscar Mañas, Pietro Astolfi, Melissa Hall et al.

Impressive advances in text-to-image (T2I) generative models have yielded a plethora of high performing models which are able to generate aesthetically appealing, photorealistic images. Despite the progress, these models still struggle to produce images that are consistent with the input prompt, oftentimes failing to capture object quantities, relations and attributes properly. Existing solutions to improve prompt-image consistency suffer from the following challenges: (1) they oftentimes require model fine-tuning, (2) they only focus on nearby prompt samples, and (3) they are affected by unfavorable trade-offs among image quality, representation diversity, and prompt-image consistency. In this paper, we address these challenges and introduce a T2I optimization-by-prompting framework, OPT2I, which leverages a large language model (LLM) to improve prompt-image consistency in T2I models. Our framework starts from a user prompt and iteratively generates revised prompts with the goal of maximizing a consistency score. Our extensive validation on two datasets, MSCOCO and PartiPrompts, shows that OPT2I can boost the initial consistency score by up to 24.9% in terms of DSG score while preserving the FID and increasing the recall between generated and real data. Our work paves the way toward building more reliable and robust T2I systems by harnessing the power of LLMs.

Bertie Vidgen, Adarsh Agrawal, Ahmed M. Ahmed et al. · deepmind, oxford

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

Leshem Choshen, Ryan Cotterell, Michael Y. Hu et al. · ibm-research

After last year's successful BabyLM Challenge, the competition will be hosted again in 2024/2025. The overarching goals of the challenge remain the same; however, some of the competition rules will be different. The big changes for this year's competition are as follows: First, we replace the loose track with a paper track, which allows (for example) non-model-based submissions, novel cognitively-inspired benchmarks, or analysis techniques. Second, we are relaxing the rules around pretraining data, and will now allow participants to construct their own datasets provided they stay within the 100M-word or 10M-word budget. Third, we introduce a multimodal vision-and-language track, and will release a corpus of 50% text-only and 50% image-text multimodal data as a starting point for LM model training. The purpose of this CfP is to provide rules for this year's challenge, explain these rule changes and their rationale in greater detail, give a timeline of this year's competition, and provide answers to frequently asked questions from last year's challenge.

Michael Y. Hu, Aaron Mueller, Candace Ross et al. · ibm-research

The BabyLM Challenge is a community effort to close the data-efficiency gap between human and computational language learners. Participants compete to optimize language model training on a fixed language data budget of 100 million words or less. This year, we released improved text corpora, as well as a vision-and-language corpus to facilitate research into cognitively plausible vision language models. Submissions were compared on evaluation tasks targeting grammatical ability, (visual) question answering, pragmatic abilities, and grounding, among other abilities. Participants could submit to a 10M-word text-only track, a 100M-word text-only track, and/or a 100M-word and image multimodal track. From 31 submissions employing diverse methods, a hybrid causal-masked language model architecture outperformed other approaches. No submissions outperformed the baselines in the multimodal track. In follow-up analyses, we found a strong relationship between training FLOPs and average performance across tasks, and that the best-performing submissions proposed changes to the training data, training objective, and model architecture. This year's BabyLM Challenge shows that there is still significant room for innovation in this setting, in particular for image-text modeling, but community-driven research can yield actionable insights about effective strategies for small-scale language modeling.

Lucas Charpentier, Leshem Choshen, Ryan Cotterell et al. · ibm-research

BabyLM aims to dissolve the boundaries between cognitive modeling and language modeling. We call for both workshop papers and for researchers to join the 3rd BabyLM competition. As in previous years, we call for participants in the data-efficient pretraining challenge in the general track. This year, we also offer a new track: INTERACTION. This new track encourages interactive behavior, learning from a teacher, and adapting the teaching material to the student. We also call for papers outside the competition in any relevant areas. These include training efficiency, cognitively plausible research, weak model evaluation, and more.

Maureen de Seyssel, Antony D'Avirro, Adina Williams et al.

We introduce EmphAssess, a prosodic benchmark designed to evaluate the capability of speech-to-speech models to encode and reproduce prosodic emphasis. We apply this to two tasks: speech resynthesis and speech-to-speech translation. In both cases, the benchmark evaluates the ability of the model to encode emphasis in the speech input and accurately reproduce it in the output, potentially across a change of speaker and language. As part of the evaluation pipeline, we introduce EmphaClass, a new model that classifies emphasis at the frame or word level.

Shaona Ghosh, Heather Frase, Adina Williams et al. · deepmind, stanford

The rapid advancement and deployment of AI systems have created an urgent need for standard safety-evaluation frameworks. This paper introduces AILuminate v1.0, the first comprehensive industry-standard benchmark for assessing AI-product risk and reliability. Its development employed an open process that included participants from multiple fields. The benchmark evaluates an AI system's resistance to prompts designed to elicit dangerous, illegal, or undesirable behavior in 12 hazard categories, including violent crimes, nonviolent crimes, sex-related crimes, child sexual exploitation, indiscriminate weapons, suicide and self-harm, intellectual property, privacy, defamation, hate, sexual content, and specialized advice (election, financial, health, legal). Our method incorporates a complete assessment standard, extensive prompt datasets, a novel evaluation framework, a grading and reporting system, and the technical as well as organizational infrastructure for long-term support and evolution. In particular, the benchmark employs an understandable five-tier grading scale (Poor to Excellent) and incorporates an innovative entropy-based system-response evaluation. In addition to unveiling the benchmark, this report also identifies limitations of our method and of building safety benchmarks generally, including evaluator uncertainty and the constraints of single-turn interactions. This work represents a crucial step toward establishing global standards for AI risk and reliability evaluation while acknowledging the need for continued development in areas such as multiturn interactions, multimodal understanding, coverage of additional languages, and emerging hazard categories. Our findings provide valuable insights for model developers, system integrators, and policymakers working to promote safer AI deployment.

Melissa Hall, Samuel J. Bell, Candace Ross et al. · meta-ai

Rapid progress in text-to-image generative models coupled with their deployment for visual content creation has magnified the importance of thoroughly evaluating their performance and identifying potential biases. In pursuit of models that generate images that are realistic, diverse, visually appealing, and consistent with the given prompt, researchers and practitioners often turn to automated metrics to facilitate scalable and cost-effective performance profiling. However, commonly-used metrics often fail to account for the full diversity of human preference; often even in-depth human evaluations face challenges with subjectivity, especially as interpretations of evaluation criteria vary across regions and cultures. In this work, we conduct a large, cross-cultural study to study how much annotators in Africa, Europe, and Southeast Asia vary in their perception of geographic representation, visual appeal, and consistency in real and generated images from state-of-the art public APIs. We collect over 65,000 image annotations and 20 survey responses. We contrast human annotations with common automated metrics, finding that human preferences vary notably across geographic location and that current metrics do not fully account for this diversity. For example, annotators in different locations often disagree on whether exaggerated, stereotypical depictions of a region are considered geographically representative. In addition, the utility of automatic evaluations is dependent on assumptions about their set-up, such as the alignment of feature extractors with human perception of object similarity or the definition of "appeal" captured in reference datasets used to ground evaluations. We recommend steps for improved automatic and human evaluations.

Vipul Gupta, Candace Ross, David Pantoja et al.

One of the most challenging problems facing NLP today is evaluation. Some of the most pressing issues pertain to benchmark saturation, data contamination, and diversity in the quality of test examples. To address these concerns, we propose Selection Methodology for Accurate, Reduced, and Targeted (SMART) filtering, a novel approach to select a high-quality subset of examples from existing benchmark datasets by systematically removing less informative and less challenging examples. Our approach applies three filtering criteria, removing (i) easy examples, (ii) data-contaminated examples, and (iii) examples that are similar to each other based on distance in an embedding space. We demonstrate the effectiveness of SMART on three multiple choice QA datasets, where our methodology increases efficiency by reducing dataset size by 48\% on average, while increasing Pearson correlation with rankings from ChatBot Arena, a more open-ended human evaluation setting. Our method enables us to be more efficient, whether using SMART to make new benchmarks more challenging or to revitalize older datasets, while still preserving the relative model rankings.

Florian Bordes, Quentin Garrido, Justine T Kao et al.

We present IntPhys 2, a video benchmark designed to evaluate the intuitive physics understanding of deep learning models. Building on the original IntPhys benchmark, IntPhys 2 focuses on four core principles related to macroscopic objects: Permanence, Immutability, Spatio-Temporal Continuity, and Solidity. These conditions are inspired by research into intuitive physical understanding emerging during early childhood. IntPhys 2 offers a comprehensive suite of tests, based on the violation of expectation framework, that challenge models to differentiate between possible and impossible events within controlled and diverse virtual environments. Alongside the benchmark, we provide performance evaluations of several state-of-the-art models. Our findings indicate that while these models demonstrate basic visual understanding, they face significant challenges in grasping intuitive physics across the four principles in complex scenes, with most models performing at chance levels (50%), in stark contrast to human performance, which achieves near-perfect accuracy. This underscores the gap between current models and human-like intuitive physics understanding, highlighting the need for advancements in model architectures and training methodologies.

Candace Ross, Melissa Hall, Adriana Romero Soriano et al.

Language models are increasingly being incorporated as components in larger AI systems for various purposes, from prompt optimization to automatic evaluation. In this work, we analyze the construct validity of four recent, commonly used methods for measuring text-to-image consistency - CLIPScore, TIFA, VPEval, and DSG - which rely on language models and/or VQA models as components. We define construct validity for text-image consistency metrics as a set of desiderata that text-image consistency metrics should have, and find that no tested metric satisfies all of them. We find that metrics lack sufficient sensitivity to language and visual properties. Next, we find that TIFA, VPEval and DSG contribute novel information above and beyond CLIPScore, but also that they correlate highly with each other. We also ablate different aspects of the text-image consistency metrics and find that not all model components are strictly necessary, also a symptom of insufficient sensitivity to visual information. Finally, we show that all three VQA-based metrics likely rely on familiar text shortcuts (such as yes-bias in QA) that call their aptitude as quantitative evaluations of model performance into question.

Guy Davidson, Todd M. Gureckis, Brenden M. Lake et al.

Demonstrations and instructions are two primary approaches for prompting language models to perform in-context learning (ICL) tasks. Do identical tasks elicited in different ways result in similar representations of the task? An improved understanding of task representation mechanisms would offer interpretability insights and may aid in steering models. We study this through \textit{function vectors} (FVs), recently proposed as a mechanism to extract few-shot ICL task representations. We generalize FVs to alternative task presentations, focusing on short textual instruction prompts, and successfully extract instruction function vectors that promote zero-shot task accuracy. We find evidence that demonstration- and instruction-based function vectors leverage different model components, and offer several controls to dissociate their contributions to task performance. Our results suggest that different task promptings forms do not induce a common task representation through FVs but elicit different, partly overlapping mechanisms. Our findings offer principled support to the practice of combining instructions and task demonstrations, imply challenges in universally monitoring task inference across presentation forms, and encourage further examinations of LLM task inference mechanisms.

Anaelia Ovalle, Krunoslav Lehman Pavasovic, Louis Martin et al.

Natural-language assistants are designed to provide users with helpful responses while avoiding harmful outputs, largely achieved through alignment to human preferences. Yet there is limited understanding of whether alignment techniques may inadvertently perpetuate or even amplify harmful biases inherited from their pre-aligned base models. This issue is compounded by the choice of bias evaluation benchmarks in popular preference-finetuned models, which predominantly focus on dominant social categories, such as binary gender, thereby limiting insights into biases affecting underrepresented groups. Towards addressing this gap, we center transgender, nonbinary, and other gender-diverse identities to investigate how alignment procedures interact with pre-existing gender-diverse bias in LLMs. Our key contributions include: 1) a comprehensive survey of bias evaluation modalities across leading preference-finetuned LLMs, highlighting critical gaps in gender-diverse representation, 2) systematic evaluation of gender-diverse biases across 16 models spanning Direct Preference Optimization (DPO) stages, uncovering harms popular bias benchmarks fail to detect, and 3) a flexible framework for measuring harmful biases in implicit reward signals applicable to other social contexts. Our findings reveal that DPO-aligned models are particularly sensitive to supervised finetuning (SFT), and can amplify two forms of real-world gender-diverse harms from their base models: stigmatization and gender non-affirmative language. We conclude with recommendations tailored to DPO and broader alignment practices, advocating for the adoption of community-informed bias evaluation frameworks to more effectively identify and address underrepresented harms in LLMs.

Mingda Chen, Yang Li, Xilun Chen et al.

Long-form factuality evaluation assesses the ability of models to generate accurate, comprehensive responses to short prompts. Existing benchmarks often lack human verification, leading to potential quality issues. To address this limitation, we introduce FACTORY, a large-scale, human-verified prompt set. Developed using a model-in-the-loop approach and refined by humans, FACTORY includes challenging prompts that are fact-seeking, answerable, and unambiguous. We conduct human evaluations on 6 state-of-the-art language models using FACTORY and existing datasets. Our results show that FACTORY is a challenging benchmark: approximately 40% of the claims made in the responses of SOTA models are not factual, compared to only 10% for other datasets. Our analysis identifies the strengths of FACTORY over prior benchmarks, emphasizing its reliability and the necessity for models to reason across long-tailed facts.

Bhaktipriya Radharapu, Manon Revel, Megan Ung et al.

The increasing use of LLMs as substitutes for humans in ``aligning'' LLMs has raised questions about their ability to replicate human judgments and preferences, especially in ambivalent scenarios where humans disagree. This study examines the biases and limitations of LLMs in three roles: answer generator, judge, and debater. These roles loosely correspond to previously described alignment frameworks: preference alignment (judge) and scalable oversight (debater), with the answer generator reflecting the typical setting with user interactions. We develop a ``no-consensus'' benchmark by curating examples that encompass a variety of a priori ambivalent scenarios, each presenting two possible stances. Our results show that while LLMs can provide nuanced assessments when generating open-ended answers, they tend to take a stance on no-consensus topics when employed as judges or debaters. These findings underscore the necessity for more sophisticated methods for aligning LLMs without human oversight, highlighting that LLMs cannot fully capture human disagreement even on topics where humans themselves are divided.

Yanli Zhou, Brenden M. Lake, Adina Williams

The ability to learn and compose functions is foundational to efficient learning and reasoning in humans, enabling flexible generalizations such as creating new dishes from known cooking processes. Beyond sequential chaining of functions, existing linguistics literature indicates that humans can grasp more complex compositions with interacting functions, where output production depends on context changes induced by different function orderings. Extending the investigation into the visual domain, we developed a function learning paradigm to explore the capacity of humans and neural network models in learning and reasoning with compositional functions under varied interaction conditions. Following brief training on individual functions, human participants were assessed on composing two learned functions, in ways covering four main interaction types, including instances in which the application of the first function creates or removes the context for applying the second function. Our findings indicate that humans can make zero-shot generalizations on novel visual function compositions across interaction conditions, demonstrating sensitivity to contextual changes. A comparison with a neural network model on the same task reveals that, through the meta-learning for compositionality (MLC) approach, a standard sequence-to-sequence Transformer can mimic human generalization patterns in composing functions.

Megan Ung, Alicia Sun, Samuel J. Bell et al.

Large language models (LLMs) are often fine-tuned for use on downstream tasks, though this can degrade capabilities learned during previous training. This phenomenon, often referred to as catastrophic forgetting, has important potential implications for the safety of deployed models. In this work, we first show that models trained on downstream tasks forget their safety tuning to a greater extent than models trained in the opposite order. Second, we show that forgetting disproportionately impacts safety information about certain groups. To quantify this phenomenon, we define a new metric we term biased forgetting. We conduct a systematic evaluation of the effects of task ordering on forgetting and apply mitigations that can help the model recover from the forgetting observed. We hope our findings can better inform methods for chaining the finetuning of LLMs in continual learning settings to enable training of safer and less toxic models.

Samuel J. Bell, Mariano Coria Meglioli, Megan Richards et al.

Text toxicity detection systems exhibit significant biases, producing disproportionate rates of false positives on samples mentioning demographic groups. But what about toxicity detection in speech? To investigate the extent to which text-based biases are mitigated by speech-based systems, we produce a set of high-quality group annotations for the multilingual MuTox dataset, and then leverage these annotations to systematically compare speech- and text-based toxicity classifiers. Our findings indicate that access to speech data during inference supports reduced bias against group mentions, particularly for ambiguous and disagreement-inducing samples. Our results also suggest that improving classifiers, rather than transcription pipelines, is more helpful for reducing group bias. We publicly release our annotations and provide recommendations for future toxicity dataset construction.

Vipul Gupta, David Pantoja, Candace Ross et al.

As large language models (LLMs) have grown in prevalence, particular benchmarks have become essential for the evaluation of these models and for understanding model capabilities. Most commonly, we use test accuracy averaged across multiple subtasks in order to rank models on leaderboards, to determine which model is best for our purposes. In this paper, we investigate the robustness of the accuracy measurement on a widely used multiple choice question answering dataset, MMLU. When shuffling the answer label contents, we find that all explored models decrease in accuracy on MMLU, but not every model is equally sensitive. These findings suggest a possible adjustment to the standard practice of leaderboard testing, where we additionally consider the percentage of examples each model answers correctly by random chance.

Abhishek Sureddy, Dishant Padalia, Nandhinee Periyakaruppa et al.

Recent work has identified substantial disparities in generated images of different geographic regions, including stereotypical depictions of everyday objects like houses and cars. However, existing measures for these disparities have been limited to either human evaluations, which are time-consuming and costly, or automatic metrics evaluating full images, which are unable to attribute these disparities to specific parts of the generated images. In this work, we introduce a new set of metrics, Decomposed Indicators of Disparities in Image Generation (Decomposed-DIG), that allows us to separately measure geographic disparities in the depiction of objects and backgrounds in generated images. Using Decomposed-DIG, we audit a widely used latent diffusion model and find that generated images depict objects with better realism than backgrounds and that backgrounds in generated images tend to contain larger regional disparities than objects. We use Decomposed-DIG to pinpoint specific examples of disparities, such as stereotypical background generation in Africa, struggling to generate modern vehicles in Africa, and unrealistically placing some objects in outdoor settings. Informed by our metric, we use a new prompting structure that enables a 52% worst-region improvement and a 20% average improvement in generated background diversity.

Ouail Kitouni, Niklas Nolte, Diane Bouchacourt et al.

Today's best language models still struggle with hallucinations: factually incorrect generations, which impede their ability to reliably retrieve information seen during training. The reversal curse, where models cannot recall information when probed in a different order than was encountered during training, exemplifies this in information retrieval. We reframe the reversal curse as a factorization curse - a failure of models to learn the same joint distribution under different factorizations. Through a series of controlled experiments with increasing levels of realism including WikiReversal, a setting we introduce to closely simulate a knowledge intensive finetuning task, we find that the factorization curse is an inherent failure of the next-token prediction objective used in popular large language models. Moreover, we demonstrate reliable information retrieval cannot be solved with scale, reversed tokens, or even naive bidirectional-attention training. Consequently, various approaches to finetuning on specialized data would necessarily provide mixed results on downstream tasks, unless the model has already seen the right sequence of tokens. Across five tasks of varying levels of complexity, our results uncover a promising path forward: factorization-agnostic objectives can significantly mitigate the reversal curse and hint at improved knowledge storage and planning capabilities.

Karolina Stańczak, Lucas Torroba Hennigen, Adina Williams et al.

The success of pre-trained contextualized representations has prompted researchers to analyze them for the presence of linguistic information. Indeed, it is natural to assume that these pre-trained representations do encode some level of linguistic knowledge as they have brought about large empirical improvements on a wide variety of NLP tasks, which suggests they are learning true linguistic generalization. In this work, we focus on intrinsic probing, an analysis technique where the goal is not only to identify whether a representation encodes a linguistic attribute but also to pinpoint where this attribute is encoded. We propose a novel latent-variable formulation for constructing intrinsic probes and derive a tractable variational approximation to the log-likelihood. Our results show that our model is versatile and yields tighter mutual information estimates than two intrinsic probes previously proposed in the literature. Finally, we find empirical evidence that pre-trained representations develop a cross-lingually entangled notion of morphosyntax.