Dan Jurafsky

Peter Henderson, Mark S. Krass, Lucia Zheng et al. · stanford

One concern with the rise of large language models lies with their potential for significant harm, particularly from pretraining on biased, obscene, copyrighted, and private information. Emerging ethical approaches have attempted to filter pretraining material, but such approaches have been ad hoc and failed to take context into account. We offer an approach to filtering grounded in law, which has directly addressed the tradeoffs in filtering material. First, we gather and make available the Pile of Law, a 256GB (and growing) dataset of open-source English-language legal and administrative data, covering court opinions, contracts, administrative rules, and legislative records. Pretraining on the Pile of Law may help with legal tasks that have the promise to improve access to justice. Second, we distill the legal norms that governments have developed to constrain the inclusion of toxic or private content into actionable lessons for researchers and discuss how our dataset reflects these norms. Third, we show how the Pile of Law offers researchers the opportunity to learn such filtering rules directly from the data, providing an exciting new research direction in model-based processing.

Peter Henderson, Eric Mitchell, Christopher D. Manning et al. · stanford

A growing ecosystem of large, open-source foundation models has reduced the labeled data and technical expertise necessary to apply machine learning to many new problems. Yet foundation models pose a clear dual-use risk, indiscriminately reducing the costs of building both harmful and beneficial machine learning systems. Policy tools such as restricted model access and export controls are the primary methods currently used to mitigate such dual-use risks. In this work, we review potential safe-release strategies and argue that both policymakers and AI researchers would benefit from fundamentally new technologies enabling more precise control over the downstream usage of open-source foundation models. We propose one such approach: the task blocking paradigm, in which foundation models are trained with an additional mechanism to impede adaptation to harmful tasks without sacrificing performance on desirable tasks. We call the resulting models self-destructing models, inspired by mechanisms that prevent adversaries from using tools for harmful purposes. We present an algorithm for training self-destructing models leveraging techniques from meta-learning and adversarial learning, which we call meta-learned adversarial censoring (MLAC). In a small-scale experiment, we show MLAC can largely prevent a BERT-style model from being re-purposed to perform gender identification without harming the model's ability to perform profession classification.

Mirac Suzgun, Stuart M. Shieber, Dan Jurafsky · stanford

We introduce string2string, an open-source library that offers a comprehensive suite of efficient algorithms for a broad range of string-to-string problems. It includes traditional algorithmic solutions as well as recent advanced neural approaches to tackle various problems in string alignment, distance measurement, lexical and semantic search, and similarity analysis -- along with several helpful visualization tools and metrics to facilitate the interpretation and analysis of these methods. Notable algorithms featured in the library include the Smith-Waterman algorithm for pairwise local alignment, the Hirschberg algorithm for global alignment, the Wagner-Fisher algorithm for edit distance, BARTScore and BERTScore for similarity analysis, the Knuth-Morris-Pratt algorithm for lexical search, and Faiss for semantic search. Besides, it wraps existing efficient and widely-used implementations of certain frameworks and metrics, such as sacreBLEU and ROUGE, whenever it is appropriate and suitable. Overall, the library aims to provide extensive coverage and increased flexibility in comparison to existing libraries for strings. It can be used for many downstream applications, tasks, and problems in natural-language processing, bioinformatics, and computational social sciences. It is implemented in Python, easily installable via pip, and accessible through a simple API. Source code, documentation, and tutorials are all available on our GitHub page: https://github.com/stanfordnlp/string2string.

Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri et al. · stanford

Despite the success of large vision and language models (VLMs) in many downstream applications, it is unclear how well they encode compositional information. Here, we create the Attribution, Relation, and Order (ARO) benchmark to systematically evaluate the ability of VLMs to understand different types of relationships, attributes, and order. ARO consists of Visual Genome Attribution, to test the understanding of objects' properties; Visual Genome Relation, to test for relational understanding; and COCO & Flickr30k-Order, to test for order sensitivity. ARO is orders of magnitude larger than previous benchmarks of compositionality, with more than 50,000 test cases. We show where state-of-the-art VLMs have poor relational understanding, can blunder when linking objects to their attributes, and demonstrate a severe lack of order sensitivity. VLMs are predominantly trained and evaluated on large datasets with rich compositional structure in the images and captions. Yet, training on these datasets has not been enough to address the lack of compositional understanding, and evaluating on these datasets has failed to surface this deficiency. To understand why these limitations emerge and are not represented in the standard tests, we zoom into the evaluation and training procedures. We demonstrate that it is possible to perform well on retrieval over existing datasets without using the composition and order information. Given that contrastive pretraining optimizes for retrieval on datasets with similar shortcuts, we hypothesize that this can explain why the models do not need to learn to represent compositional information. This finding suggests a natural solution: composition-aware hard negative mining. We show that a simple-to-implement modification of contrastive learning significantly improves the performance on tasks requiring understanding of order and compositionality.

Federico Bianchi, Pratyusha Kalluri, Esin Durmus et al. · stanford

Machine learning models that convert user-written text descriptions into images are now widely available online and used by millions of users to generate millions of images a day. We investigate the potential for these models to amplify dangerous and complex stereotypes. We find a broad range of ordinary prompts produce stereotypes, including prompts simply mentioning traits, descriptors, occupations, or objects. For example, we find cases of prompting for basic traits or social roles resulting in images reinforcing whiteness as ideal, prompting for occupations resulting in amplification of racial and gender disparities, and prompting for objects resulting in reification of American norms. Stereotypes are present regardless of whether prompts explicitly mention identity and demographic language or avoid such language. Moreover, stereotypes persist despite mitigation strategies; neither user attempts to counter stereotypes by requesting images with specific counter-stereotypes nor institutional attempts to add system ``guardrails'' have prevented the perpetuation of stereotypes. Our analysis justifies concerns regarding the impacts of today's models, presenting striking exemplars, and connecting these findings with deep insights into harms drawn from social scientific and humanist disciplines. This work contributes to the effort to shed light on the uniquely complex biases in language-vision models and demonstrates the ways that the mass deployment of text-to-image generation models results in mass dissemination of stereotypes and resulting harms.

Federico Bianchi, Mirac Suzgun, Giuseppe Attanasio et al. · stanford

Training large language models to follow instructions makes them perform better on a wide range of tasks and generally become more helpful. However, a perfectly helpful model will follow even the most malicious instructions and readily generate harmful content. In this paper, we raise concerns over the safety of models that only emphasize helpfulness, not harmlessness, in their instruction-tuning. We show that several popular instruction-tuned models are highly unsafe. Moreover, we show that adding just 3% safety examples (a few hundred demonstrations) when fine-tuning a model like LLaMA can substantially improve its safety. Our safety-tuning does not make models significantly less capable or helpful as measured by standard benchmarks. However, we do find exaggerated safety behaviours, where too much safety-tuning makes models refuse perfectly safe prompts if they superficially resemble unsafe ones. As a whole, our results illustrate trade-offs in training LLMs to be helpful and training them to be safe.

Peter Henderson, Xuechen Li, Dan Jurafsky et al. · stanford

Existing foundation models are trained on copyrighted material. Deploying these models can pose both legal and ethical risks when data creators fail to receive appropriate attribution or compensation. In the United States and several other countries, copyrighted content may be used to build foundation models without incurring liability due to the fair use doctrine. However, there is a caveat: If the model produces output that is similar to copyrighted data, particularly in scenarios that affect the market of that data, fair use may no longer apply to the output of the model. In this work, we emphasize that fair use is not guaranteed, and additional work may be necessary to keep model development and deployment squarely in the realm of fair use. First, we survey the potential risks of developing and deploying foundation models based on copyrighted content. We review relevant U.S. case law, drawing parallels to existing and potential applications for generating text, source code, and visual art. Experiments confirm that popular foundation models can generate content considerably similar to copyrighted material. Second, we discuss technical mitigations that can help foundation models stay in line with fair use. We argue that more research is needed to align mitigation strategies with the current state of the law. Lastly, we suggest that the law and technical mitigations should co-evolve. For example, coupled with other policy mechanisms, the law could more explicitly consider safe harbors when strong technical tools are used to mitigate infringement harms. This co-evolution may help strike a balance between intellectual property and innovation, which speaks to the original goal of fair use. But we emphasize that the strategies we describe here are not a panacea and more work is needed to develop policies that address the potential harms of foundation models.

Rishi Bommasani, Kathleen A. Creel, Ananya Kumar et al. · stanford

As the scope of machine learning broadens, we observe a recurring theme of algorithmic monoculture: the same systems, or systems that share components (e.g. training data), are deployed by multiple decision-makers. While sharing offers clear advantages (e.g. amortizing costs), does it bear risks? We introduce and formalize one such risk, outcome homogenization: the extent to which particular individuals or groups experience negative outcomes from all decision-makers. If the same individuals or groups exclusively experience undesirable outcomes, this may institutionalize systemic exclusion and reinscribe social hierarchy. To relate algorithmic monoculture and outcome homogenization, we propose the component-sharing hypothesis: if decision-makers share components like training data or specific models, then they will produce more homogeneous outcomes. We test this hypothesis on algorithmic fairness benchmarks, demonstrating that sharing training data reliably exacerbates homogenization, with individual-level effects generally exceeding group-level effects. Further, given the dominant paradigm in AI of foundation models, i.e. models that can be adapted for myriad downstream tasks, we test whether model sharing homogenizes outcomes across tasks. We observe mixed results: we find that for both vision and language settings, the specific methods for adapting a foundation model significantly influence the degree of outcome homogenization. We conclude with philosophical analyses of and societal challenges for outcome homogenization, with an eye towards implications for deployed machine learning systems.

Kaitlyn Zhou, Kawin Ethayarajh, Dallas Card et al. · stanford

Cosine similarity of contextual embeddings is used in many NLP tasks (e.g., QA, IR, MT) and metrics (e.g., BERTScore). Here, we uncover systematic ways in which word similarities estimated by cosine over BERT embeddings are understated and trace this effect to training data frequency. We find that relative to human judgements, cosine similarity underestimates the similarity of frequent words with other instances of the same word or other words across contexts, even after controlling for polysemy and other factors. We conjecture that this underestimation of similarity for high frequency words is due to differences in the representational geometry of high and low frequency words and provide a formal argument for the two-dimensional case.

Garrett Tanzer, Mirac Suzgun, Eline Visser et al. · stanford

Large language models (LLMs) can perform impressive feats with in-context learning or lightweight finetuning. It is natural to wonder how well these models adapt to genuinely new tasks, but how does one find tasks that are unseen in internet-scale training sets? We turn to a field that is explicitly motivated and bottlenecked by a scarcity of web data: low-resource languages. In this paper, we introduce MTOB (Machine Translation from One Book), a benchmark for learning to translate between English and Kalamang -- a language with less than 200 speakers and therefore virtually no presence on the web -- using several hundred pages of field linguistics reference materials. This task framing is novel in that it asks a model to learn a language from a single human-readable book of grammar explanations, rather than a large mined corpus of in-domain data, more akin to L2 learning than L1 acquisition. We demonstrate that baselines using current LLMs are promising but fall short of human performance, achieving 44.7 chrF on Kalamang to English translation and 45.8 chrF on English to Kalamang translation, compared to 51.6 and 57.0 chrF by a human who learned Kalamang from the same reference materials. We hope that MTOB will help measure LLM capabilities along a new dimension, and that the methods developed to solve it could help expand access to language technology for underserved communities by leveraging qualitatively different kinds of data than traditional machine translation.

Mirac Suzgun, Luke Melas-Kyriazi, Dan Jurafsky · stanford

We propose a method for arbitrary textual style transfer (TST)--the task of transforming a text into any given style--utilizing general-purpose pre-trained language models. Our method, Prompt-and-Rerank, is based on a mathematical formulation of the TST task, decomposing it into three constituent components: textual similarity, target style strength, and fluency. Specifically, our method first uses zero-shot or few-shot prompting to obtain a set of candidate generations in the target style, and then re-ranks these candidates according to a combination of the three components above. Empirically, our method enables small pre-trained language models to perform on par with state-of-the-art large-scale models while consuming two orders of magnitude less compute and memory. Finally, we conduct a systematic investigation of the effect of model size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on style transfer quality across seven diverse textual style transfer datasets.

Mirac Suzgun, Luke Melas-Kyriazi, Dan Jurafsky · stanford

In open-ended natural-language generation, existing text decoding methods typically struggle to produce text which is both diverse and high-quality. Greedy and beam search are known to suffer from text degeneration and linguistic diversity issues, while temperature, top-k, and nucleus sampling often yield diverse but low-quality outputs. In this work, we present crowd sampling, a family of decoding methods based on Bayesian risk minimization, to address this diversity-quality trade-off. Inspired by the principle of "the wisdom of the crowd," crowd sampling seeks to select a candidate from a pool of candidates that has the least expected risk (i.e., highest expected reward) under a generative model according to a given utility function. Crowd sampling can be seen as a generalization of numerous existing methods, including majority voting, and in practice, it can be used as a drop-in replacement for existing sampling methods. Extensive experiments show that crowd sampling delivers improvements of 3-7 ROUGE and BLEU points across a wide range of tasks, including summarization, data-to-text, translation, and textual style transfer, while achieving new state-of-the-art results on WebNLG and WMT'16.

Sterling Alic, Dorottya Demszky, Zid Mancenido et al. · stanford

Responsive teaching is a highly effective strategy that promotes student learning. In math classrooms, teachers might "funnel" students towards a normative answer or "focus" students to reflect on their own thinking, deepening their understanding of math concepts. When teachers focus, they treat students' contributions as resources for collective sensemaking, and thereby significantly improve students' achievement and confidence in mathematics. We propose the task of computationally detecting funneling and focusing questions in classroom discourse. We do so by creating and releasing an annotated dataset of 2,348 teacher utterances labeled for funneling and focusing questions, or neither. We introduce supervised and unsupervised approaches to differentiating these questions. Our best model, a supervised RoBERTa model fine-tuned on our dataset, has a strong linear correlation of .76 with human expert labels and with positive educational outcomes, including math instruction quality and student achievement, showing the model's potential for use in automated teacher feedback tools. Our unsupervised measures show significant but weaker correlations with human labels and outcomes, and they highlight interesting linguistic patterns of funneling and focusing questions. The high performance of the supervised measure indicates its promise for supporting teachers in their instruction.

Kawin Ethayarajh, Dan Jurafsky · stanford

Human ratings are the gold standard in NLG evaluation. The standard protocol is to collect ratings of generated text, average across annotators, and rank NLG systems by their average scores. However, little consideration has been given as to whether this approach faithfully captures human preferences. Analyzing this standard protocol through the lens of utility theory in economics, we identify the implicit assumptions it makes about annotators. These assumptions are often violated in practice, in which case annotator ratings cease to reflect their preferences. The most egregious violations come from using Likert scales, which provably reverse the direction of the true preference in certain cases. We suggest improvements to the standard protocol to make it more theoretically sound, but even in its improved form, it cannot be used to evaluate open-ended tasks like story generation. For the latter, we propose a new human evaluation protocol called $\textit{system-level probabilistic assessment}$ (SPA). When human evaluation of stories is done with SPA, we can recover the ordering of GPT-3 models by size, with statistically significant results. However, when human evaluation is done with the standard protocol, less than half of the expected preferences can be recovered (e.g., there is no significant difference between $\texttt{curie}$ and $\texttt{davinci}$, despite using a highly powered test).

Isabel Papadimitriou, Dan Jurafsky · stanford

Both humans and large language models are able to learn language without explicit structural supervision. What inductive biases make this learning possible? We address this fundamental cognitive question by leveraging transformer language models: we inject inductive bias into language models by pretraining on formally-structured data, and then evaluate the biased learners' ability to learn typologically-diverse natural languages. Our experimental setup creates a testbed for hypotheses about inductive bias in human language learning. We investigate the effect of injecting models with three types of inductive bias: 1) recursive, hierarchical processing, 2) crossing token-token relationships that can't be modeled by context-free grammars, and 3) a Zipfian power-law vocabulary distribution. We show that non-context-free relationships form the best inductive biases. Our study leverages the capabilities of transformer models to run controlled language learning experiments that are not possible to run on humans, and surfaces hypotheses about the structures that facilitate language learning in both humans and machines.

Connor Toups, Rishi Bommasani, Kathleen A. Creel et al. · stanford

Machine learning is traditionally studied at the model level: researchers measure and improve the accuracy, robustness, bias, efficiency, and other dimensions of specific models. In practice, the societal impact of machine learning is determined by the surrounding context of machine learning deployments. To capture this, we introduce ecosystem-level analysis: rather than analyzing a single model, we consider the collection of models that are deployed in a given context. For example, ecosystem-level analysis in hiring recognizes that a job candidate's outcomes are not only determined by a single hiring algorithm or firm but instead by the collective decisions of all the firms they applied to. Across three modalities (text, images, speech) and 11 datasets, we establish a clear trend: deployed machine learning is prone to systemic failure, meaning some users are exclusively misclassified by all models available. Even when individual models improve at the population level over time, we find these improvements rarely reduce the prevalence of systemic failure. Instead, the benefits of these improvements predominantly accrue to individuals who are already correctly classified by other models. In light of these trends, we consider medical imaging for dermatology where the costs of systemic failure are especially high. While traditional analyses reveal racial performance disparities for both models and humans, ecosystem-level analysis reveals new forms of racial disparity in model predictions that do not present in human predictions. These examples demonstrate ecosystem-level analysis has unique strengths for characterizing the societal impact of machine learning.

Tolúlopé Ògúnrèmí, Christopher D. Manning, Dan Jurafsky · stanford

While many speakers of low-resource languages regularly code-switch between their languages and other regional languages or English, datasets of codeswitched speech are too small to train bespoke acoustic models from scratch or do language model rescoring. Here we propose finetuning self-supervised speech representations such as wav2vec 2.0 XLSR to recognize code-switched data. We find that finetuning self-supervised multilingual representations and augmenting them with n-gram language models trained from transcripts reduces absolute word error rates by up to 20% compared to baselines of hybrid models trained from scratch on code-switched data. Our findings suggest that in circumstances with limited training data finetuning self-supervised representations is a better performing and viable solution.

Nay San, Martijn Bartelds, Blaine Billings et al. · stanford

Recent research using pre-trained transformer models suggests that just 10 minutes of transcribed speech may be enough to fine-tune such a model for automatic speech recognition (ASR) -- at least if we can also leverage vast amounts of text data (803 million tokens). But is that much text data necessary? We study the use of different amounts of text data, both for creating a lexicon that constrains ASR decoding to possible words (e.g. *dogz vs. dogs), and for training larger language models that bias the system toward probable word sequences (e.g. too dogs vs. two dogs). We perform experiments using 10 minutes of transcribed speech from English (for replicating prior work) and two additional pairs of languages differing in the availability of supplemental text data: Gronings and Frisian (~7.5M token corpora available), and Besemah and Nasal (only small lexica available). For all languages, we found that using only a lexicon did not appreciably improve ASR performance. For Gronings and Frisian, we found that lexica and language models derived from 'novel-length' 80k token subcorpora reduced the word error rate (WER) to 39% on average. Our findings suggest that where a text corpus in the upper tens of thousands of tokens or more is available, fine-tuning a transformer model with just tens of minutes of transcribed speech holds some promise towards obtaining human-correctable transcriptions near the 30% WER rule-of-thumb.

Kaitlyn Zhou, Kawin Ethayarajh, Dan Jurafsky · stanford

We examine whether some countries are more richly represented in embedding space than others. We find that countries whose names occur with low frequency in training corpora are more likely to be tokenized into subwords, are less semantically distinct in embedding space, and are less likely to be correctly predicted: e.g., Ghana (the correct answer and in-vocabulary) is not predicted for, "The country producing the most cocoa is [MASK].". Although these performance discrepancies and representational harms are due to frequency, we find that frequency is highly correlated with a country's GDP; thus perpetuating historic power and wealth inequalities. We analyze the effectiveness of mitigation strategies; recommend that researchers report training word frequencies; and recommend future work for the community to define and design representational guarantees.

Eva Portelance, Michael C. Frank, Dan Jurafsky · stanford

Interpreting a seemingly-simple function word like "or", "behind", or "more" can require logical, numerical, and relational reasoning. How are such words learned by children? Prior acquisition theories have often relied on positing a foundation of innate knowledge. Yet recent neural-network based visual question answering models apparently can learn to use function words as part of answering questions about complex visual scenes. In this paper, we study what these models learn about function words, in the hope of better understanding how the meanings of these words can be learnt by both models and children. We show that recurrent models trained on visually grounded language learn gradient semantics for function words requiring spatial and numerical reasoning. Furthermore, we find that these models can learn the meanings of logical connectives and and or without any prior knowledge of logical reasoning, as well as early evidence that they are sensitive to alternative expressions when interpreting language. Finally, we show that word learning difficulty is dependent on frequency in models' input. Our findings offer proof-of-concept evidence that it is possible to learn the nuanced interpretations of function words in visually grounded context by using non-symbolic general statistical learning algorithms, without any prior knowledge of linguistic meaning.

Nay San, Martijn Bartelds, Tolúlopé Ògúnrèmí et al. · stanford

Many archival recordings of speech from endangered languages remain unannotated and inaccessible to community members and language learning programs. One bottleneck is the time-intensive nature of annotation. An even narrower bottleneck occurs for recordings with access constraints, such as language that must be vetted or filtered by authorised community members before annotation can begin. We propose a privacy-preserving workflow to widen both bottlenecks for recordings where speech in the endangered language is intermixed with a more widely-used language such as English for meta-linguistic commentary and questions (e.g. What is the word for 'tree'?). We integrate voice activity detection (VAD), spoken language identification (SLI), and automatic speech recognition (ASR) to transcribe the metalinguistic content, which an authorised person can quickly scan to triage recordings that can be annotated by people with lower levels of access. We report work-in-progress processing 136 hours archival audio containing a mix of English and Muruwari. Our collaborative work with the Muruwari custodian of the archival materials show that this workflow reduces metalanguage transcription time by 20% even given only minimal amounts of annotated training data: 10 utterances per language for SLI and for ASR at most 39 minutes, and possibly as little as 39 seconds.

Yiwei Luo, Kristina Gligorić, Dan Jurafsky · stanford

Identifying implicit attitudes toward food can mitigate social prejudice due to food's salience as a marker of ethnic identity. Stereotypes about food are representational harms that may contribute to racialized discourse and negatively impact economic outcomes for restaurants. Understanding the presence of representational harms in online corpora in particular is important, given the increasing use of large language models (LLMs) for text generation and their tendency to reproduce attitudes in their training data. Through careful linguistic analyses, we evaluate social theories about attitudes toward immigrant cuisine in a large-scale study of framing differences in 2.1M English language Yelp reviews. Controlling for factors such as restaurant price and neighborhood racial diversity, we find that immigrant cuisines are more likely to be othered using socially constructed frames of authenticity (e.g., "authentic," "traditional"), and that non-European cuisines (e.g., Indian, Mexican) in particular are described as more exotic compared to European ones (e.g., French). We also find that non-European cuisines are more likely to be described as cheap and dirty, even after controlling for price, and even among the most expensive restaurants. Finally, we show that reviews generated by LLMs reproduce similar framing tendencies, pointing to the downstream retention of these representational harms. Our results corroborate social theories of gastronomic stereotyping, revealing racialized evaluative processes and linguistic strategies through which they manifest.

Anjalie Field, Prateek Verma, Nay San et al. · stanford

Police body-worn cameras have the potential to improve accountability and transparency in policing. Yet in practice, they result in millions of hours of footage that is never reviewed. We investigate the potential of large pre-trained speech models for facilitating reviews, focusing on ASR and officer speech detection in footage from traffic stops. Our proposed pipeline includes training data alignment and filtering, fine-tuning with resource constraints, and combining officer speech detection with ASR for a fully automated approach. We find that (1) fine-tuning strongly improves ASR performance on officer speech (WER=12-13%), (2) ASR on officer speech is much more accurate than on community member speech (WER=43.55-49.07%), (3) domain-specific tasks like officer speech detection and diarization remain challenging. Our work offers practical applications for reviewing body camera footage and general guidance for adapting pre-trained speech models to noisy multi-speaker domains.

Kaitlyn Zhou, Jena D. Hwang, Xiang Ren et al. · allen-ai, cmu

The ability to communicate uncertainty, risk, and limitation is crucial for the safety of large language models. However, current evaluations of these abilities rely on simple calibration, asking whether the language generated by the model matches appropriate probabilities. Instead, evaluation of this aspect of LLM communication should focus on the behaviors of their human interlocutors: how much do they rely on what the LLM says? Here we introduce an interaction-centered evaluation framework called Rel-A.I. (pronounced "rely"}) that measures whether humans rely on LLM generations. We use this framework to study how reliance is affected by contextual features of the interaction (e.g, the knowledge domain that is being discussed), or the use of greetings communicating warmth or competence (e.g., "I'm happy to help!"). We find that contextual characteristics significantly affect human reliance behavior. For example, people rely 10% more on LMs when responding to questions involving calculations and rely 30% more on LMs that are perceived as more competent. Our results show that calibration and language quality alone are insufficient in evaluating the risks of human-LM interactions, and illustrate the need to consider features of the interactional context.

Junshen K. Chen, Dallas Card, Dan Jurafsky · stanford

Off-the-shelf models are widely used by computational social science researchers to measure properties of text, such as sentiment. However, without access to source data it is difficult to account for domain shift, which represents a threat to validity. Here, we treat domain adaptation as a modular process that involves separate model producers and model consumers, and show how they can independently cooperate to facilitate more accurate measurements of text. We introduce two lightweight techniques for this scenario, and demonstrate that they reliably increase out-of-domain accuracy on four multi-domain text classification datasets when used with linear and contextual embedding models. We conclude with recommendations for model producers and consumers, and release models and replication code to accompany this paper.

Cole Simmons, Richard Diehl Martinez, Dan Jurafsky

Sumerian transliteration is a conventional system for representing a scholar's interpretation of a tablet in the Latin script. Thanks to visionary digital Assyriology projects such as ETCSL, CDLI, and Oracc, a large number of Sumerian transliterations have been published online, and these data are well-structured for a variety of search and analysis tasks. However, the absence of a comprehensive, accessible dataset pairing transliterations with a digital representation of the tablet's cuneiform glyphs has prevented the application of modern Natural Language Processing (NLP) methods to the task of Sumerian transliteration. To address this gap, we present SumTablets, a dataset pairing Unicode representations of 91,606 Sumerian cuneiform tablets (totaling 6,970,407 glyphs) with the associated transliterations published by Oracc. We construct SumTablets by first preprocessing and standardizing the Oracc transliterations before mapping each reading back to the Unicode representation of the source glyph. Further, we retain parallel structural information (e.g., surfaces, newlines, broken segments) through the use of special tokens. We release SumTablets as a Hugging Face Dataset (CC BY 4.0) and open source data preparation code via GitHub. Additionally, we leverage SumTablets to implement and evaluate two transliteration baselines: (1) weighted sampling from a glyph's possible readings, and (2) fine-tuning an autoregressive language model. Our fine-tuned language model achieves an average transliteration character-level F-score (chrF) of 97.55, demonstrating the immediate potential of transformer-based transliteration models in allowing experts to rapidly verify generated transliterations rather than manually transliterating tablets one-by-one.

Moussa Koulako Bala Doumbouya, Ananjan Nandi, Gabriel Poesia et al.

Despite their demonstrated valuable capabilities, state-of-the-art (SOTA) widely deployed large language models (LLMs) still have the potential to cause harm to society due to the ineffectiveness of their safety filters, which can be bypassed by prompt transformations called jailbreak attacks. Current approaches to LLM safety assessment, which employ datasets of templated prompts and benchmarking pipelines, fail to cover sufficiently large and diverse sets of jailbreak attacks, leading to the widespread deployment of unsafe LLMs. Recent research showed that novel jailbreak attacks could be derived by composition; however, a formal composable representation for jailbreak attacks, which, among other benefits, could enable the exploration of a large compositional space of jailbreak attacks through program synthesis methods, has not been previously proposed. We introduce h4rm3l, a novel approach that addresses this gap with a human-readable domain-specific language (DSL). Our framework comprises: (1) The h4rm3l DSL, which formally expresses jailbreak attacks as compositions of parameterized string transformation primitives. (2) A synthesizer with bandit algorithms that efficiently generates jailbreak attacks optimized for a target black box LLM. (3) The h4rm3l red-teaming software toolkit that employs the previous two components and an automated harmful LLM behavior classifier that is strongly aligned with human judgment. We demonstrate h4rm3l's efficacy by synthesizing a dataset of 2656 successful novel jailbreak attacks targeting 6 SOTA open-source and proprietary LLMs, and by benchmarking those models against a subset of these synthesized attacks. Our results show that h4rm3l's synthesized attacks are diverse and more successful than existing jailbreak attacks in literature, with success rates exceeding 90% on SOTA LLMs.

Myra Cheng, Isabel Sieh, Humishka Zope et al. · stanford

LLMs can be socially sycophantic, affirming users when they ask questions like "am I in the wrong?" rather than providing genuine assessment. We hypothesize that this behavior arises from incorrect assumptions about the user, like underestimating how often users are seeking information over reassurance. We present Verbalized Assumptions, a framework for eliciting these assumptions from LLMs. Verbalized Assumptions provide insight into LLM sycophancy, delusion, and other safety issues, e.g., the top bigram in LLMs' assumptions on social sycophancy datasets is ``seeking validation.'' We provide evidence for a causal link between Verbalized Assumptions and sycophantic model behavior: our assumption probes (linear probes trained on internal representations of these assumptions) enable interpretable fine-grained steering of social sycophancy. We explore why LLMs default to sycophantic assumptions: on identical queries, people expect more objective and informative responses from AI than from other humans, but LLMs trained on human-human conversation do not account for this difference in expectations. Our work contributes a new understanding of assumptions as a mechanism for sycophancy.

Kaitlyn Zhou, Dan Jurafsky, Tatsunori Hashimoto

The increased deployment of LMs for real-world tasks involving knowledge and facts makes it important to understand model epistemology: what LMs think they know, and how their attitudes toward that knowledge are affected by language use in their inputs. Here, we study an aspect of model epistemology: how epistemic markers of certainty, uncertainty, or evidentiality like "I'm sure it's", "I think it's", or "Wikipedia says it's" affect models, and whether they contribute to model failures. We develop a typology of epistemic markers and inject 50 markers into prompts for question answering. We find that LMs are highly sensitive to epistemic markers in prompts, with accuracies varying more than 80%. Surprisingly, we find that expressions of high certainty result in a 7% decrease in accuracy as compared to low certainty expressions; similarly, factive verbs hurt performance, while evidentials benefit performance. Our analysis of a popular pretraining dataset shows that these markers of uncertainty are associated with answers on question-answering websites, while markers of certainty are associated with questions. These associations may suggest that the behavior of LMs is based on mimicking observed language use, rather than truly reflecting epistemic uncertainty.

Kristina Gligorić, Tijana Zrnic, Cinoo Lee et al.

Large language models (LLMs) have shown high agreement with human raters across a variety of tasks, demonstrating potential to ease the challenges of human data collection. In computational social science (CSS), researchers are increasingly leveraging LLM annotations to complement slow and expensive human annotations. Still, guidelines for collecting and using LLM annotations, without compromising the validity of downstream conclusions, remain limited. We introduce Confidence-Driven Inference: a method that combines LLM annotations and LLM confidence indicators to strategically select which human annotations should be collected, with the goal of producing accurate statistical estimates and provably valid confidence intervals while reducing the number of human annotations needed. Our approach comes with safeguards against LLM annotations of poor quality, guaranteeing that the conclusions will be both valid and no less accurate than if we only relied on human annotations. We demonstrate the effectiveness of Confidence-Driven Inference over baselines in statistical estimation tasks across three CSS settings--text politeness, stance, and bias--reducing the needed number of human annotations by over 25% in each. Although we use CSS settings for demonstration, Confidence-Driven Inference can be used to estimate most standard quantities across a broad range of NLP problems.

Omar Shaikh, Kristina Gligorić, Ashna Khetan et al.

Effective conversation requires common ground: a shared understanding between the participants. Common ground, however, does not emerge spontaneously in conversation. Speakers and listeners work together to both identify and construct a shared basis while avoiding misunderstanding. To accomplish grounding, humans rely on a range of dialogue acts, like clarification (What do you mean?) and acknowledgment (I understand.). However, it is unclear whether large language models (LLMs) generate text that reflects human grounding. To this end, we curate a set of grounding acts and propose corresponding metrics that quantify attempted grounding. We study whether LLM generations contain grounding acts, simulating turn-taking from several dialogue datasets and comparing results to humans. We find that -- compared to humans -- LLMs generate language with less conversational grounding, instead generating text that appears to simply presume common ground. To understand the roots of the identified grounding gap, we examine the role of instruction tuning and preference optimization, finding that training on contemporary preference data leads to a reduction in generated grounding acts. Altogether, we highlight the need for more research investigating conversational grounding in human-AI interaction.

Rishi Bommasani, Sarah H. Bana, Kathleen A. Creel et al.

Many employers screen job applicants with algorithms built by the same few algorithm vendors. We hypothesize that algorithmic monoculture leads to the same individuals and members of the same racial groups facing rejection. We acquire and analyze a novel dataset of 3 million applicants submitting 4 million applications where all the applications are screened by algorithms built by the same vendor. We find clear racial disparities in applicant outcomes. Of all applications submitted by Asian and Black applicants, 14.74% and 25.87% are submitted to positions that adversely impact Asian and Black applicants, respectively, according to U.S. employment discrimination standards. Individuals also receive homogeneous outcomes: 4% of all applicants who apply to 10 positions are recommended for rejection from all positions, a rate higher than expected by chance. To better understand this homogeneity, we leverage the deterministic replicability of hiring algorithms to generate the outcomes applicants would have received if they applied to all positions. We show that applicants would need to apply widely in order to ensure their applications are considered by a human

Antón de la Fuente, Dan Jurafsky

This study asks how self-supervised speech models represent suprasegmental categories like Mandarin lexical tone, English lexical stress, and English phrasal accents. Through a series of probing tasks, we make layer-wise comparisons of English and Mandarin 12 layer monolingual models. Our findings suggest that 1) English and Mandarin wav2vec 2.0 models learn contextual representations of abstract suprasegmental categories which are strongest in the middle third of the network. 2) Models are better at representing features that exist in the language of their training data, and this difference is driven by enriched context in transformer blocks, not local acoustic representation. 3) Fine-tuned wav2vec 2.0 improves performance in later layers compared to pre-trained models mainly for lexically contrastive features like tone and stress, 4) HuBERT and WavLM learn similar representations to wav2vec 2.0, differing mainly in later layer performance. Our results extend previous understanding of how models represent suprasegmentals and offer new insights into the language-specificity and contextual nature of these representations.

Zhengxuan Wu, Aryaman Arora, Zheng Wang et al. · stanford

Parameter-efficient finetuning (PEFT) methods seek to adapt large neural models via updates to a small number of weights. However, much prior interpretability work has shown that representations encode rich semantic information, suggesting that editing representations might be a more powerful alternative. We pursue this hypothesis by developing a family of Representation Finetuning (ReFT) methods. ReFT methods operate on a frozen base model and learn task-specific interventions on hidden representations. We define a strong instance of the ReFT family, Low-rank Linear Subspace ReFT (LoReFT), and we identify an ablation of this method that trades some performance for increased efficiency. Both are drop-in replacements for existing PEFTs and learn interventions that are 15x--65x more parameter-efficient than LoRA. We showcase LoReFT on eight commonsense reasoning tasks, four arithmetic reasoning tasks, instruction-tuning, and GLUE. In all these evaluations, our ReFTs deliver the best balance of efficiency and performance, and almost always outperform state-of-the-art PEFTs. We release a generic ReFT training library publicly at https://github.com/stanfordnlp/pyreft.

Bianca Datta, Markus J. Buehler, Yvonne Chow et al.

Global food systems must deliver nutritious, sustainable foods while sharply reducing environmental impact. Yet, food innovation remains slow, empirical, and fragmented. Artificial intelligence (AI) offers a transformative path to link molecular composition to functional performance, connect chemical structure to sensory outcomes, and accelerate cross-disciplinary innovation across the production pipeline. While broadly applicable to food systems, we focus on sustainable proteins--plant-based, fermentation-derived, and cultivated--as a high-impact testbed for AI-driven closed-loop design. We review the applications, opportunities, and challenges of AI for Food as an emerging discipline that integrates ingredient design, formulation development, fermentation and production, texture analysis, sensory science, manufacturing, and recipe generation. We identify four priorities: advancing scientific machine learning with embedded domain priors, treating food as a programmable biomaterial, building self-driving laboratories for automated discovery, and developing deep reasoning models that integrate nutrition and sustainability. Integrating AI responsibly into the food innovation cycle can accelerate the transition to sustainable food systems and establish a predictive, design-driven science of food for human and planetary health.

Sunny Yu, Myra Cheng, Ahmad Jabbar et al.

Large language models (LLMs) have the potential to boost human productivity by speeding up task completion -- provided users know when to offload cognitive work to them. But we do not know if users are well-calibrated in estimating these potential time savings. We conducted a preregistered large-scale behavioral study (N = 1237) to characterize mismatches between expectations and reality, with a focus on simple cognitive tasks. While actual completion times between independent completion and AI-assisted completion did not differ, participants predicted AI to be significantly faster. The same bias was not observed when imagining help from another human participant. We identify a speedup illusion where people have accurate forecasts of independent completion times but significantly underestimate AI-assisted times. Additionally, time and effort dissociate: participants reported lower subjective effort with AI despite equivalent completion times. This suggests that completion time itself is not sufficient to characterize efficiency gains.

Mirac Suzgun, Emily Shen, Federico Bianchi et al.

AI chatbots are rapidly shaping how people encounter the news, yet no prior study has systematically measured how accurately these systems, with their proprietary search integrations and retrieval-synthesis pipelines, handle emerging facts across languages and regions. We present a 14-day (February 9-22, 2026) evaluation of six AI chatbots (Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini) on 2,100 factual questions derived from same-day BBC News reporting across six regional services (US & Canada, Arabic, Afrique, Hindi, Russian, Turkish). The best systems achieve over 90% multiple-choice accuracy on questions about events reported hours earlier. The same systems, however, lose 11-13% under free-response evaluation, and 16-17% across the cohort. We further characterize three failure patterns. First, every model achieves its lowest accuracy on Hindi (79% vs. 89-91% elsewhere) and citations indicate an Anglophone retrieval bias (e.g., models answering Hindi queries cite English Wikipedia more than any Hindi outlet). Second, retrieval, not reasoning, failures drive over 70% of all errors. When models retrieve a correct source, they often extract the correct answer; the problem is to land on the right source in the first place. Third, models achieving 88-96% accuracy on well-formed questions drop to 19-70% when questions contain subtle false premises, with the most vulnerable model accepting fabricated facts 64% of the time. We also identify a detection-accuracy paradox: the best false-premise detector ranks second in adversarial accuracy (abstention rate), while a weaker detector ranks first, showing that premise detection and answer recovery are partially independent capabilities. Overall, these suggest that high accuracy can mask systematic regional inequity, near-total dependence on retrieval infrastructure, and vulnerability to imperfect queries real users pose.

Sunny Yu, Myra Cheng, Ahmad Jabbar et al.

People are increasingly turning to AI assistance for simple tasks, e.g., arithmetic, spell-check, and answering simple questions. But does AI assistance actually save users time and effort? We investigate people's propensity to use AI for cognitively simple tasks and assess whether their reliance is well-calibrated. Across three pre-registered user studies (N = 2691), we find that people frequently choose to use AI even when doing so is inefficient (i.e. provides no meaningful time or effort savings). We identify systematic miscalibration at two levels: (1) a self-estimate miscalibration where people on average believe that they are using AI less than they actually are, and (2) efficiency-gain illusions where people overestimate how much time and effort savings AI use affords. We also identify a session-level carryover effect where a participant's prior AI use leads to further AI adoption and entrenches their miscalibration about time savings. Our results shed light on the mechanisms and biases underlying people's choice of whether to use AI as well as the risk of an overreliance feedback loop.

Laya Iyer, Pranav Somani, Alice Guo et al.

The next-token prediction (NTP) objective has been foundational in the development of modern large language models (LLMs), driving advances in fluency and generalization. However, NTP operates at the \textit{token} level, treating deviations from a single reference continuation as errors even when alternative continuations are equally plausible or semantically equivalent (e.g., ``mom'' vs. ``mother''). As a result, token-level loss can penalize valid abstractions, paraphrases, or conceptually correct reasoning paths, biasing models toward surface form rather than underlying meaning. This mismatch between the training signal and semantic correctness motivates learning objectives that operate over higher-level representations. We propose a shift from token-level to concept-level prediction, where concepts group multiple surface forms of the same idea (e.g., ``mom,'' ``mommy,'' ``mother'' $\rightarrow$ \textit{MOTHER}). We introduce various methods for integrating conceptual supervision into LLM training and show that concept-aware models achieve lower perplexity, improved robustness under domain shift, and stronger performance than NTP-based models on diverse NLP benchmarks. This suggests \textit{concept-level supervision} as an improved training signal that better aligns LLMs with human semantic abstractions.

Heidi C. Zhang, Shabnam Behzad, Kawin Ethayarajh et al.

Model checklists (Ribeiro et al., 2020) have emerged as a useful tool for understanding the behavior of LLMs, analogous to unit-testing in software engineering. However, despite datasets being a key determinant of model behavior, evaluating datasets, e.g., for the existence of annotation artifacts, is largely done ad hoc, once a problem in model behavior has already been found downstream. In this work, we take a more principled approach to unit-testing datasets by proposing a taxonomy based on the V-information literature. We call a collection of such unit tests a data checklist. Using a checklist, not only are we able to recover known artifacts in well-known datasets such as SNLI, but we also discover previously unknown artifacts in preference datasets for LLM alignment. Data checklists further enable a new kind of data filtering, which we use to improve the efficacy and data efficiency of preference alignment.

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.

Chen Shani, Yuval Reif, Nathan Roll et al.

Multilingual language models (LMs) promise broader NLP access, yet current systems deliver uneven performance across the world's languages. This survey examines why these gaps persist and whether they reflect intrinsic linguistic difficulty or modeling artifacts. We organize the literature around two questions: do linguistic disparities arise from representation and allocation choices (e.g., tokenization, encoding, data exposure, parameter sharing) rather than inherent complexity; and which design choices mitigate inequities across typologically diverse languages. We review linguistic features, such as orthography, morphology, lexical diversity, syntax, information density, and typological distance, linking each to concrete modeling mechanisms. Gaps often shrink when segmentation, encoding, and data exposure are normalized, suggesting much apparent difficulty stems from current modeling choices. We synthesize these insights into design recommendations for tokenization, sampling, architectures, and evaluation to support more balanced multilingual LMs.

Andrew Lanpouthakoun, Aryaman Arora, Zhengxuan Wu et al.

Large language models can now be personalised efficiently at scale using parameter efficient finetuning methods (PEFTs), but serving user-specific PEFTs harms throughput, even with specialised kernels and memory management techniques. This is because, theoretically and empirically, a mismatch exists between prefill (processing a large number of tokens at once) and decode (generating a single token autoregressively): the latter has far lower throughput when serving multiple adapters. Rather than optimising performance relative to parameter count, for efficient multi-adapter serving, we instead ought to optimise performance relative to serving throughput. We therefore propose PreFT (Prefill-only Finetuning), wherein we only apply the adapter to prefill tokens and discard it afterwards. PreFT significantly increases throughput with minimal effect on performance. We develop and release an efficient implementation of two prefill-only PEFTs, LoRA and ReFT, on the vLLM inference engine. We first show that serving multi-user PreFTs is more efficient than traditional PEFTs ($1.9\times$ the throughput when serving $512$ adapters on Llama 3.1 70B). Then, we compare the performance of prefill-only vs. all-token adapters on a variety of supervised finetuning and reinforcement learning tasks with LMs at varying scales. On SFT, we observe that the evaluation loss of PreFTs is higher than PEFTs, but can be compensated by increasing rank with nearly no reduction in throughput. On RL, we consistently find that PreFTs approach parity with standard PEFTs. Together, this work validates prefill-only adaptation of LLMs as a more favourable accuracy-throughput tradeoff than existing PEFTs for personalised serving.

Christine Zhang, Dan Jurafsky, Chen Shani

The next-token prediction (NTP) objective trains language models to predict a single continuation token at each step. In natural language, however, a prefix can be continued in many valid ways, and even similar meanings may differ in surface form. For example, the sentence ``this website is safe to \underline{browse}'' could plausibly continue with words such as browse, search, visit, surf, or navigate. While standard NTP training treats these alternatives as mutually exclusive targets, we explore a framework that instead predicts concepts, approximated as sets of semantically related tokens. We show that models trained with concept supervision exhibit stronger alignment with human semantic similarity judgments on multiple lexical benchmarks. These gains are accompanied by lower perplexity on semantically meaningful words (definition in Section 3.1), and a modest increase in global token-level perplexity, reflecting a tradeoff between standard NTP optimization and concept-level supervision. Our results suggest that concept-level objectives can improve semantic alignment while maintaining competitive language modeling performance.

Zhengxuan Wu, Aryaman Arora, Atticus Geiger et al. · stanford

Fine-grained steering of language model outputs is essential for safety and reliability. Prompting and finetuning are widely used to achieve these goals, but interpretability researchers have proposed a variety of representation-based techniques as well, including sparse autoencoders (SAEs), linear artificial tomography, supervised steering vectors, linear probes, and representation finetuning. At present, there is no benchmark for making direct comparisons between these proposals. Therefore, we introduce AxBench, a large-scale benchmark for steering and concept detection, and report experiments on Gemma-2-2B and 9B. For steering, we find that prompting outperforms all existing methods, followed by finetuning. For concept detection, representation-based methods such as difference-in-means, perform the best. On both evaluations, SAEs are not competitive. We introduce a novel weakly-supervised representational method (Rank-1 Representation Finetuning; ReFT-r1), which is competitive on both tasks while providing the interpretability advantages that prompting lacks. Along with AxBench, we train and publicly release SAE-scale feature dictionaries for ReFT-r1 and DiffMean.

Isabel Papadimitriou, Kezia Lopez, Dan Jurafsky

While multilingual language models can improve NLP performance on low-resource languages by leveraging higher-resource languages, they also reduce average performance on all languages (the 'curse of multilinguality'). Here we show another problem with multilingual models: grammatical structures in higher-resource languages bleed into lower-resource languages, a phenomenon we call grammatical structure bias. We show this bias via a novel method for comparing the fluency of multilingual models to the fluency of monolingual Spanish and Greek models: testing their preference for two carefully-chosen variable grammatical structures (optional pronoun-drop in Spanish and optional Subject-Verb ordering in Greek). We find that multilingual BERT is biased toward the English-like setting (explicit pronouns and Subject-Verb-Object ordering) as compared to our monolingual control language model. With our case studies, we hope to bring to light the fine-grained ways in which multilingual models can be biased,and encourage more linguistically-aware fluency evaluation.

Federico Bianchi, Patrick John Chia, Mert Yuksekgonul et al.

Negotiation is the basis of social interactions; humans negotiate everything from the price of cars to how to share common resources. With rapidly growing interest in using large language models (LLMs) to act as agents on behalf of human users, such LLM agents would also need to be able to negotiate. In this paper, we study how well LLMs can negotiate with each other. We develop NegotiationArena: a flexible framework for evaluating and probing the negotiation abilities of LLM agents. We implemented three types of scenarios in NegotiationArena to assess LLM's behaviors in allocating shared resources (ultimatum games), aggregate resources (trading games) and buy/sell goods (price negotiations). Each scenario allows for multiple turns of flexible dialogues between LLM agents to allow for more complex negotiations. Interestingly, LLM agents can significantly boost their negotiation outcomes by employing certain behavioral tactics. For example, by pretending to be desolate and desperate, LLMs can improve their payoffs by 20\% when negotiating against the standard GPT-4. We also quantify irrational negotiation behaviors exhibited by the LLM agents, many of which also appear in humans. Together, \NegotiationArena offers a new environment to investigate LLM interactions, enabling new insights into LLM's theory of mind, irrationality, and reasoning abilities.

Valentin Hofmann, Pratyusha Ria Kalluri, Dan Jurafsky et al.

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.

Aryaman Arora, Dan Jurafsky, Christopher Potts · stanford

Language models (LMs) have proven to be powerful tools for psycholinguistic research, but most prior work has focused on purely behavioural measures (e.g., surprisal comparisons). At the same time, research in model interpretability has begun to illuminate the abstract causal mechanisms shaping LM behavior. To help bring these strands of research closer together, we introduce CausalGym. We adapt and expand the SyntaxGym suite of tasks to benchmark the ability of interpretability methods to causally affect model behaviour. To illustrate how CausalGym can be used, we study the pythia models (14M--6.9B) and assess the causal efficacy of a wide range of interpretability methods, including linear probing and distributed alignment search (DAS). We find that DAS outperforms the other methods, and so we use it to study the learning trajectory of two difficult linguistic phenomena in pythia-1b: negative polarity item licensing and filler--gap dependencies. Our analysis shows that the mechanism implementing both of these tasks is learned in discrete stages, not gradually.

Myra Cheng, Kristina Gligoric, Tiziano Piccardi et al.

Anthropomorphism, or the attribution of human-like characteristics to non-human entities, has shaped conversations about the impacts and possibilities of technology. We present AnthroScore, an automatic metric of implicit anthropomorphism in language. We use a masked language model to quantify how non-human entities are implicitly framed as human by the surrounding context. We show that AnthroScore corresponds with human judgments of anthropomorphism and dimensions of anthropomorphism described in social science literature. Motivated by concerns of misleading anthropomorphism in computer science discourse, we use AnthroScore to analyze 15 years of research papers and downstream news articles. In research papers, we find that anthropomorphism has steadily increased over time, and that papers related to language models have the most anthropomorphism. Within ACL papers, temporal increases in anthropomorphism are correlated with key neural advancements. Building upon concerns of scientific misinformation in mass media, we identify higher levels of anthropomorphism in news headlines compared to the research papers they cite. Since AnthroScore is lexicon-free, it can be directly applied to a wide range of text sources.