Large language models (LLMs) are helping millions of users write texts about diverse issues, and in doing so expose users to different ideas and perspectives. This creates concerns about issue bias, where an LLM tends to present just one perspective on a given issue, which in turn may influence how users think about this issue. So far, it has not been possible to measure which issue biases LLMs actually manifest in real user interactions, making it difficult to address the risks from biased LLMs. Therefore, we create IssueBench: a set of 2.49m realistic prompts for measuring issue bias in LLM writing assistance, which we construct based on 3.9k templates (e.g. "write a blog about") and 212 political issues (e.g. "AI regulation") from real user interactions. Using IssueBench, we show that issue biases are common and persistent in state-of-the-art LLMs. We also show that biases are remarkably similar across models, and that all models align more with US Democrat than Republican voter opinion on a subset of issues. IssueBench can easily be adapted to include other issues, templates, or tasks. By enabling robust and realistic measurement, we hope that IssueBench can bring a new quality of evidence to ongoing discussions about LLM biases and how to address them.

Large language models (LLMs) are trained and tested extensively on symbolic representations such as code and graphs, yet real-world user tasks are often specified in natural language. To what extent can LLMs generalize across these representations? Here, we approach this question by studying isomorphic tasks involving procedures represented in code, graphs, and natural language (e.g., scheduling steps in planning). We find that training LLMs with popular post-training methods on graphs or code data alone does not reliably generalize to corresponding natural language tasks, while training solely on natural language can lead to inefficient performance gains. To address this gap, we propose a two-stage data curriculum that first trains on symbolic, then natural language data. The curriculum substantially improves model performance across model families and tasks. Remarkably, a 1.5B Qwen model trained by our method can closely match zero-shot GPT-4o in naturalistic planning. Finally, our analysis suggests that successful cross-representation generalization can be interpreted as a form of generative analogy, which our curriculum effectively encourages.

Hundreds of millions of people rely on large language models (LLMs) for education, work, and even healthcare. Yet these models are known to reproduce and amplify social biases present in their training data. Moreover, text-based interfaces remain a barrier for many, for example, users with limited literacy, motor impairments, or mobile-only devices. Voice interaction promises to expand accessibility, but unlike text, speech carries identity cues that users cannot easily mask, raising concerns about whether accessibility gains may come at the cost of equitable treatment. Here we show that audio-enabled LLMs exhibit systematic gender discrimination, shifting responses toward gender-stereotyped adjectives and occupations solely on the basis of speaker voice, and amplifying bias beyond that observed in text-based interaction. Thus, voice interfaces do not merely extend text models to a new modality but introduce distinct bias mechanisms tied to paralinguistic cues. Complementary survey evidence (n=1,000) shows that infrequent chatbot users are most hesitant to undisclosed attribute inference and most likely to disengage when such practices are revealed. To demonstrate a potential mitigation strategy, we show that pitch manipulation can systematically regulate gender-discriminatory outputs. Overall, our findings reveal a critical tension in AI development: efforts to expand accessibility through voice interfaces simultaneously create new pathways for discrimination, demanding that fairness and accessibility be addressed in tandem.

Demographic cue-based evaluation is widely used to study how large language models (LLMs) adapt their responses to signaled demographic attributes within and across groups. This approach typically relies on a single cue (e.g., names) as a proxy for group membership, implicitly treating different cues as interchangeable operationalizations of a single underlying identity-conditioned behavior. We test this assumption in realistic advice-seeking interactions spanning 14.8 million prompts, focusing on race and gender in a U.S. context. We find that cues for the same group induce only partially overlapping changes in model responses, yielding inconsistent conclusions about personalization, while bias conclusions are unstable, with both magnitude and direction of group differences varying across cues. We further show that these inconsistencies reflect differences in cue-group association strength and linguistic features bundled within cues that shape model responses. Together, our findings suggest that demographic conditioning in LLMs is not a cue-invariant category-level parameter but depends fundamentally on how identity is cued, reflecting responses to linguistic signals rather than stable demographic categories. We therefore call for multi-cue, mechanism-aware evaluations as a foundation for robust and interpretable claims about demographic variation in LLM responses.

Language model (LM) benchmarking faces several challenges: comprehensive evaluations are costly, benchmarks often fail to measure the intended capabilities, and evaluation quality can degrade due to labeling errors and benchmark saturation. Although various strategies have been proposed to mitigate these issues, they tend to address individual aspects in isolation, neglecting broader questions about overall evaluation quality. Here, we introduce Fluid Benchmarking, a new evaluation approach that advances LM benchmarking across multiple dimensions. Inspired by psychometrics, Fluid Benchmarking is based on the insight that the relative value of benchmark items depends on an LM's capability level, suggesting that evaluation should adapt to each LM. Methodologically, Fluid Benchmarking estimates an item response model based on existing LM evaluation results and uses the inferred quantities to select evaluation items dynamically, similar to computerized adaptive testing in education. In our experiments, we compare Fluid Benchmarking against the common practice of random item sampling as well as more sophisticated baselines, including alternative methods grounded in item response theory. We examine four dimensions – efficiency, validity, variance, and saturation – and find that Fluid Benchmarking achieves superior performance in all of them (e.g., higher validity and less variance on MMLU with fifty times fewer items). Our analysis shows that the two components of Fluid Benchmarking have distinct effects: item response theory, used to map performance into a latent ability space, increases validity, while dynamic item selection reduces variance. Overall, our results suggest that LM benchmarking can be substantially improved by moving beyond static evaluation.

Language is not monolithic. While benchmarks, including those designed for multiple languages, are often used as proxies to evaluate the performance of Large Language Models (LLMs), they tend to overlook the nuances of within-language variation and thus fail to model the experience of speakers of non-standard dialects. Focusing on African American Vernacular English (AAVE), we present the first study aimed at objectively assessing the fairness and robustness of LLMs in handling dialects across canonical reasoning tasks, including algorithm, math, logic, and integrated reasoning. We introduce ReDial (Reasoning with Dialect Queries), a benchmark containing 1.2K+ parallel query pairs in Standardized English and AAVE. We hire AAVE speakers, including experts with computer science backgrounds, to rewrite seven popular benchmarks, such as HumanEval and GSM8K. With ReDial, we evaluate widely used LLMs, including GPT, Claude, Llama, Mistral, and the Phi model families. Our findings reveal that almost all of these widely used models show significant brittleness and unfairness to queries in AAVE. Our work establishes a systematic and objective framework for analyzing LLM bias in dialectal queries. Moreover, it highlights how mainstream LLMs provide unfair service to dialect speakers in reasoning tasks, laying a critical foundation for future research.

Queer vernacular is rarely studied in NLP, despite advancements in resources and evaluation for other sociolects and informal language. Because of this, NLP systems often process queer language incorrectly, e.g., they misclassify it as hate speech or generate negative responses. To address this problem, we propose Slaying, the first real-world dataset of English queer slang. Slaying is community-validated, and includes over 500 queer slang terms that pertain to more than 20 queer subcommunities. We argue that queer language data resources have great potential in NLP – e.g., as components of large pretraining corpora and as the basis for benchmarks – and can improve queer users' experience of NLP systems. We leverage Slaying for two novel findings in support of this argument: (i) For a number of language models, we show that they are unbiased towards the queer community, but at the same time unable to process its language, i.e., absence of representation bias does not entail the absence of linguistic bias. (ii) Model performance on queer slang varies across queer subcommunities; it is generally worse for slang pertaining to African-American and Latine communities. These findings are relevant for both the queer NLP and the broader ML communities. Slaying is available to the public, and open to future revisions and extensions. Warning: This paper contains profane and potentially offensive language.

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

Dialects represent a significant component of human culture and are found across all regions of the world. In Germany, more than 40% of the population speaks a regional dialect (Adler and Hansen, 2022). However, despite cultural importance, individuals speaking dialects often face negative societal stereotypes. We examine whether such stereotypes are mirrored by large language models (LLMs). We draw on the sociolinguistic literature on dialect perception to analyze traits commonly associated with dialect speakers. Based on these traits, we assess the dialect naming bias and dialect usage bias expressed by LLMs in two tasks: association task and decision task. To assess a model’s dialect usage bias, we construct a novel evaluation corpus that pairs sentences from seven regional German dialects (e.g., Alemannic and Bavarian) with their standard German counterparts. We find that: (1) in the association task, all evaluated LLMs exhibit significant dialect naming and dialect usage bias against German dialect speakers, reflected in negative adjective associations; (2) all models reproduce these dialect naming and dialect usage biases in their decision making; and (3) contrary to prior work showing minimal bias with explicit demographic mentions, we find that explicitly labeling linguistic demographics—German dialect speakers—amplifies bias more than implicit cues like dialect usage.

We introduce Bolmo, the first family of competitive fully open byte-level language models (LMs) at the 1B and 7B parameter scales. In contrast to prior research on byte-level LMs, which focuses predominantly on training from scratch, we train Bolmo by byteifying existing subword-level LMs. Byteification enables overcoming the limitations of subword tokenization - such as insufficient character understanding and efficiency constraints due to the fixed subword vocabulary - while performing at the level of leading subword-level LMs. Bolmo is specifically designed for byteification: our architecture resolves a mismatch between the expressivity of prior byte-level architectures and subword-level LMs, which makes it possible to employ an effective exact distillation objective between Bolmo and the source subword model. This allows for converting a subword-level LM to a byte-level LM by investing less than 1% of a typical pretraining token budget. Bolmo substantially outperforms all prior byte-level LMs of comparable size, and outperforms the source subword-level LMs on character understanding and, in some cases, coding, while coming close to matching the original LMs' performance on other tasks. Furthermore, we show that Bolmo can achieve inference speeds competitive with subword-level LMs by training with higher token compression ratios, and can be cheaply and effectively post-trained by leveraging the existing ecosystem around the source subword-level LM. Our results finally make byte-level LMs a practical choice competitive with subword-level LMs across a wide set of use cases.

Planning is a fundamental property of human intelligence. Reasoning about asynchronous plans is challenging since it requires sequential and parallel planning to optimize time costs. Can large language models (LLMs) succeed at this task? Here, we present the first large-scale study investigating this question. We find that a representative set of closed and open-source LLMs, including GPT-4 and LLaMA-2, behave poorly when not supplied with illustrations about the task-solving process in our benchmark AsyncHow. We propose a novel technique calledPlan Like a Graph(PLaG) that combines graphs with natural language prompts and achieves state-of-the-art results. We show that although PLaG can boost model performance, LLMs still suffer from drastic degradation when task complexity increases, highlighting the limits of utilizing LLMs for simulating digital devices. We see our study as an exciting step towards using LLMs as efficient autonomous agents. Our code and data are available at https://github.com/fangru-lin/graph-llm-asynchow-plan.

Information about pretraining corpora used to train the current best-performing language models is seldom discussed: commercial models rarely detail their data, and even open models are often released without accompanying training data or recipes to reproduce them. As a result, it is challenging to conduct and advance scientific research on language modeling, such as understanding how training data impacts model capabilities and limitations. To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials. We extensively document Dolma, including its design principles, details about its construction, and a summary of its contents. We present analyses and experimental results on intermediate states of Dolma to share what we have learned about important data curation practices. Finally, we open-source our data curation toolkit to enable reproduction of our work as well as support further research in large-scale data curation.

Much recent work seeks to evaluate values and opinions in large language models (LLMs) using multiple-choice surveys and questionnaires. Most of this work is motivated by concerns around real-world LLM applications. For example, politically-biased LLMs may subtly influence society when they are used by millions of people. Such real-world concerns, however, stand in stark contrast to the artificiality of current evaluations: real users do not typically ask LLMs survey questions. Motivated by this discrepancy, we challenge the prevailing constrained evaluation paradigm for values and opinions in LLMs and explore more realistic unconstrained evaluations. As a case study, we focus on the popular Political Compass Test (PCT). In a systematic review, we find that most prior work using the PCT forces models to comply with the PCT's multiple-choice format. We show that models give substantively different answers when not forced; that answers change depending on how models are forced; and that answers lack paraphrase robustness. Then, we demonstrate that models give different answers yet again in a more realistic open-ended answer setting. We distill these findings into recommendations and open challenges in evaluating values and opinions in LLMs.

While pretrained language models (PLMs) have been shown to possess a plethora of linguistic knowledge, the existing body of research has largely neglected extralinguistic knowledge, which is generally difficult to obtain by pretraining on text alone. Here, we contribute to closing this gap by examining geolinguistic knowledge, i.e., knowledge about geographic variation in language. We introduce geoadaptation, an intermediate training step that couples language modeling with geolocation prediction in a multi-task learning setup. We geoadapt four PLMs, covering language groups from three geographic areas, and evaluate them on five different tasks: fine-tuned (i.e., supervised) geolocation prediction, zero-shot (i.e., unsupervised) geolocation prediction, fine-tuned language identification, zero-shot language identification, and zero-shot prediction of dialect features. Geoadaptation is very successful at injecting geolinguistic knowledge into the PLMs: the geoadapted PLMs consistently outperform PLMs adapted using only language modeling (by especially wide margins on zero-shot prediction tasks), and we obtain new state-of-the-art results on two benchmarks for geolocation prediction and language identification. Furthermore, we show that the effectiveness of geoadaptation stems from its ability to geographically retrofit the representation space of the PLMs.

Hundreds of millions of people now interact with language models, with uses ranging from help with writing1,2 to informing hiring decisions3. However, these language models are known to perpetuate systematic racial prejudices, making their judgements biased in problematic ways about groups such as African Americans4-7. Although previous research has focused on overt racism in language models, social scientists have argued that racism with a more subtle character has developed over time, particularly in the United States after the civil rights movement8,9. It is unknown whether this covert racism manifests in language models. Here, we demonstrate that language models embody covert racism in the form of dialect prejudice, exhibiting raciolinguistic stereotypes about speakers of African American English (AAE) that are more negative than any human stereotypes about African Americans ever experimentally recorded. By contrast, the language models' overt stereotypes about African Americans are more positive. Dialect prejudice has the potential for harmful consequences: language models are more likely to suggest that speakers of AAE be assigned less-prestigious jobs, be convicted of crimes and be sentenced to death. Finally, we show that current practices of alleviating racial bias in language models, such as human preference alignment, exacerbate the discrepancy between covert and overt stereotypes, by superficially obscuring the racism that language models maintain on a deeper level. Our findings have far-reaching implications for the fair and safe use of language technology.

Hundreds of millions of people now interact with language models, with uses ranging from serving as a writing aid to informing hiring decisions. Yet these language models are known to perpetuate systematic racial prejudices, making their judgments biased in problematic ways about groups like African Americans. While prior research has focused on overt racism in language models, social scientists have argued that racism with a more subtle character has developed over time. It is unknown whether this covert racism manifests in language models. Here, we demonstrate that language models embody covert racism in the form of dialect prejudice: we extend research showing that Americans hold raciolinguistic stereotypes about speakers of African American English and find that language models have the same prejudice, exhibiting covert stereotypes that are more negative than any human stereotypes about African Americans ever experimentally recorded, although closest to the ones from before the civil rights movement. By contrast, the language models' overt stereotypes about African Americans are much more positive. We demonstrate that dialect prejudice has the potential for harmful consequences by asking language models to make hypothetical decisions about people, based only on how they speak. Language models are more likely to suggest that speakers of African American English be assigned less prestigious jobs, be convicted of crimes, and be sentenced to death. Finally, we show that existing methods for alleviating racial bias in language models such as human feedback training do not mitigate the dialect prejudice, but can exacerbate the discrepancy between covert and overt stereotypes, by teaching language models to superficially conceal the racism that they maintain on a deeper level. Our findings have far-reaching implications for the fair and safe employment of language technology.

Construction Grammar (CxG) is a paradigm from cognitive linguistics emphasizing the connection between syntax and semantics. Rather than rules that operate on lexical items, it posits constructions as the central building blocks of language, i.e., linguistic units of different granularity that combine syntax and semantics. As a first step toward assessing the compatibility of CxG with the syntactic and semantic knowledge demonstrated by state-of-the-art pretrained language models (PLMs), we present an investigation of their capability to classify and understand one of the most commonly studied constructions, the English comparative correlative (CC). We conduct experiments examining the classification accuracy of a syntactic probe on the one hand and the models' behavior in a semantic application task on the other, with BERT, RoBERTa, and DeBERTa as the example PLMs. Our results show that all three investigated PLMs, as well as OPT, are able to recognize the structure of the CC but fail to use its meaning. While human-like performance of PLMs on many NLP tasks has been alleged, this indicates that PLMs still suffer from substantial shortcomings in central domains of linguistic knowledge.

Large language models (LLMs) have recently reached an impressive level of linguistic capability, prompting comparisons with human language skills. However, there have been relatively few systematic inquiries into the linguistic capabilities of the latest generation of LLMs, and those studies that do exist (i) ignore the remarkable ability of humans to generalize, (ii) focus only on English, and (iii) investigate syntax or semantics and overlook other capabilities that lie at the heart of human language, like morphology. Here, we close these gaps by conducting the first rigorous analysis of the morphological capabilities of ChatGPT in four typologically varied languages (specifically, English, German, Tamil, and Turkish). We apply a version of Berko's (1958) wug test to ChatGPT, using novel, uncontaminated datasets for the four examined languages. We find that ChatGPT massively underperforms purpose-built systems, particularly in English. Overall, our results -- through the lens of morphology -- cast a new light on the linguistic capabilities of ChatGPT, suggesting that claims of human-like language skills are premature and misleading.

We propose a fully unsupervised method to detect bias in contextualized embeddings. The method leverages the assortative information latently encoded by social networks and combines orthogonality regularization, structured sparsity learning, and graph neural networks to find the embedding subspace capturing this information. As a concrete example, we focus on the phenomenon of ideological bias: we introduce the concept of an ideological subspace, show how it can be found by applying our method to online discussion forums, and present techniques to probe it. Our experiments suggest that the ideological subspace encodes abstract evaluative semantics and reflects changes in the political left-right spectrum during the presidency of Donald Trump.

We introduce CaMEL (Case Marker Extraction without Labels), a novel and challenging task in computational morphology that is especially relevant for low-resource languages. We propose a first model for CaMEL that uses a massively multilingual corpus to extract case markers in 83 languages based only on a noun phrase chunker and an alignment system. To evaluate CaMEL, we automatically construct a silver standard from UniMorph. The case markers extracted by our model can be used to detect and visualise similarities and differences between the case systems of different languages as well as to annotate fine-grained deep cases in languages in which they are not overtly marked.