Diyi Yang

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Matthew O. Jackson, Qiaozhu Me, Stephanie W. Wang et al.

We outline a foundation for a new field of ``AI Behavioral Science,'' covering three perspectives. First, as AI becomes ubiquitous and is increasingly proprietary and opaque, it becomes vital to develop techniques for assessing AI behavior. We outline how tools developed to assess people's behaviors by social scientists can be used to assess and infer AI's behaviors biases, tendencies, and heuristics. Second, we also discuss how AI can change the ways in which we learn about human behavior. Beyond its computational power, AI offers new techniques for simulating, inferring, and predicting human behaviors that we outline and discuss. Third, as humans and AI are interacting in increasingly complex and intertwined systems, we need to understand the implications for the resulting economic and political outcomes. We outline issues that are increasingly pressing concerning the future of human-AI interactions and potential changes and disruptions that can ensue.

Raj Sanjay Shah, Kunal Chawla, Dheeraj Eidnani et al. · gatech

Pre-trained language models have shown impressive performance on a variety of tasks and domains. Previous research on financial language models usually employs a generic training scheme to train standard model architectures, without completely leveraging the richness of the financial data. We propose a novel domain specific Financial LANGuage model (FLANG) which uses financial keywords and phrases for better masking, together with span boundary objective and in-filing objective. Additionally, the evaluation benchmarks in the field have been limited. To this end, we contribute the Financial Language Understanding Evaluation (FLUE), an open-source comprehensive suite of benchmarks for the financial domain. These include new benchmarks across 5 NLP tasks in financial domain as well as common benchmarks used in the previous research. Experiments on these benchmarks suggest that our model outperforms those in prior literature on a variety of NLP tasks. Our models, code and benchmark data are publicly available on Github and Huggingface.

Kaijie Zhu, Jiaao Chen, Jindong Wang et al. · gatech

Large language models (LLMs) have achieved remarkable performance in various evaluation benchmarks. However, concerns are raised about potential data contamination in their considerable volume of training corpus. Moreover, the static nature and fixed complexity of current benchmarks may inadequately gauge the advancing capabilities of LLMs. In this paper, we introduce DyVal, a general and flexible protocol for dynamic evaluation of LLMs. Based on our framework, we build graph-informed DyVal by leveraging the structural advantage of directed acyclic graphs to dynamically generate evaluation samples with controllable complexities. DyVal generates challenging evaluation sets on reasoning tasks including mathematics, logical reasoning, and algorithm problems. We evaluate various LLMs ranging from Flan-T5-large to GPT-3.5-Turbo and GPT-4. Experiments show that LLMs perform worse in DyVal-generated evaluation samples with different complexities, highlighting the significance of dynamic evaluation. We also analyze the failure cases and results of different prompting methods. Moreover, DyVal-generated samples are not only evaluation sets, but also helpful data for fine-tuning to improve the performance of LLMs on existing benchmarks. We hope that DyVal can shed light on future evaluation research of LLMs. Code is available at: https://github.com/microsoft/promptbench.

Clark Barrett, Brad Boyd, Elie Burzstein et al. · berkeley

Every major technical invention resurfaces the dual-use dilemma -- the new technology has the potential to be used for good as well as for harm. Generative AI (GenAI) techniques, such as large language models (LLMs) and diffusion models, have shown remarkable capabilities (e.g., in-context learning, code-completion, and text-to-image generation and editing). However, GenAI can be used just as well by attackers to generate new attacks and increase the velocity and efficacy of existing attacks. This paper reports the findings of a workshop held at Google (co-organized by Stanford University and the University of Wisconsin-Madison) on the dual-use dilemma posed by GenAI. This paper is not meant to be comprehensive, but is rather an attempt to synthesize some of the interesting findings from the workshop. We discuss short-term and long-term goals for the community on this topic. We hope this paper provides both a launching point for a discussion on this important topic as well as interesting problems that the research community can work to address.

Caleb Ziems, Jane A. Yu, Yi-Chia Wang et al.

Conversational agents have come increasingly closer to human competence in open-domain dialogue settings; however, such models can reflect insensitive, hurtful, or entirely incoherent viewpoints that erode a user's trust in the moral integrity of the system. Moral deviations are difficult to mitigate because moral judgments are not universal, and there may be multiple competing judgments that apply to a situation simultaneously. In this work, we introduce a new resource, not to authoritatively resolve moral ambiguities, but instead to facilitate systematic understanding of the intuitions, values and moral judgments reflected in the utterances of dialogue systems. The Moral Integrity Corpus, MIC, is such a resource, which captures the moral assumptions of 38k prompt-reply pairs, using 99k distinct Rules of Thumb (RoTs). Each RoT reflects a particular moral conviction that can explain why a chatbot's reply may appear acceptable or problematic. We further organize RoTs with a set of 9 moral and social attributes and benchmark performance for attribute classification. Most importantly, we show that current neural language models can automatically generate new RoTs that reasonably describe previously unseen interactions, but they still struggle with certain scenarios. Our findings suggest that MIC will be a useful resource for understanding and language models' implicit moral assumptions and flexibly benchmarking the integrity of conversational agents. To download the data, see https://github.com/GT-SALT/mic

Ying Xu, Dakuo Wang, Mo Yu et al. · cmu

Question answering (QA) is a fundamental means to facilitate assessment and training of narrative comprehension skills for both machines and young children, yet there is scarcity of high-quality QA datasets carefully designed to serve this purpose. In particular, existing datasets rarely distinguish fine-grained reading skills, such as the understanding of varying narrative elements. Drawing on the reading education research, we introduce FairytaleQA, a dataset focusing on narrative comprehension of kindergarten to eighth-grade students. Generated by educational experts based on an evidence-based theoretical framework, FairytaleQA consists of 10,580 explicit and implicit questions derived from 278 children-friendly stories, covering seven types of narrative elements or relations. Our dataset is valuable in two folds: First, we ran existing QA models on our dataset and confirmed that this annotation helps assess models' fine-grained learning skills. Second, the dataset supports question generation (QG) task in the education domain. Through benchmarking with QG models, we show that the QG model trained on FairytaleQA is capable of asking high-quality and more diverse questions.

Caleb Ziems, Jiaao Chen, Camille Harris et al. · gatech

English Natural Language Understanding (NLU) systems have achieved great performances and even outperformed humans on benchmarks like GLUE and SuperGLUE. However, these benchmarks contain only textbook Standard American English (SAE). Other dialects have been largely overlooked in the NLP community. This leads to biased and inequitable NLU systems that serve only a sub-population of speakers. To understand disparities in current models and to facilitate more dialect-competent NLU systems, we introduce the VernAcular Language Understanding Evaluation (VALUE) benchmark, a challenging variant of GLUE that we created with a set of lexical and morphosyntactic transformation rules. In this initial release (V.1), we construct rules for 11 features of African American Vernacular English (AAVE), and we recruit fluent AAVE speakers to validate each feature transformation via linguistic acceptability judgments in a participatory design manner. Experiments show that these new dialectal features can lead to a drop in model performance. To run the transformation code and download both synthetic and gold-standard dialectal GLUE benchmarks, see https://github.com/SALT-NLP/value

Yijia Shao, Tianshi Li, Weiyan Shi et al. · harvard, stanford

As language models (LMs) are widely utilized in personalized communication scenarios (e.g., sending emails, writing social media posts) and endowed with a certain level of agency, ensuring they act in accordance with the contextual privacy norms becomes increasingly critical. However, quantifying the privacy norm awareness of LMs and the emerging privacy risk in LM-mediated communication is challenging due to (1) the contextual and long-tailed nature of privacy-sensitive cases, and (2) the lack of evaluation approaches that capture realistic application scenarios. To address these challenges, we propose PrivacyLens, a novel framework designed to extend privacy-sensitive seeds into expressive vignettes and further into agent trajectories, enabling multi-level evaluation of privacy leakage in LM agents' actions. We instantiate PrivacyLens with a collection of privacy norms grounded in privacy literature and crowdsourced seeds. Using this dataset, we reveal a discrepancy between LM performance in answering probing questions and their actual behavior when executing user instructions in an agent setup. State-of-the-art LMs, like GPT-4 and Llama-3-70B, leak sensitive information in 25.68% and 38.69% of cases, even when prompted with privacy-enhancing instructions. We also demonstrate the dynamic nature of PrivacyLens by extending each seed into multiple trajectories to red-team LM privacy leakage risk. Dataset and code are available at https://github.com/SALT-NLP/PrivacyLens.

Yanzhe Zhang, Xuezhi Wang, Diyi Yang · gatech

Continual learning is essential for real-world deployment when there is a need to quickly adapt the model to new tasks without forgetting knowledge of old tasks. Existing work on continual sequence generation either always reuses existing parameters to learn new tasks, which is vulnerable to catastrophic forgetting on dissimilar tasks, or blindly adds new parameters for every new task, which could prevent knowledge sharing between similar tasks. To get the best of both worlds, in this work, we propose continual sequence generation with adaptive compositional modules to adaptively add modules in transformer architectures and compose both old and new modules for new tasks. We also incorporate pseudo experience replay to facilitate knowledge transfer in those shared modules. Experiment results on various sequences of generation tasks show that our framework can adaptively add modules or reuse modules based on task similarity, outperforming state-of-the-art baselines in terms of both performance and parameter efficiency. We make our code public at https://github.com/GT-SALT/Adaptive-Compositional-Modules.

Albert Lu, Hongxin Zhang, Yanzhe Zhang et al. · gatech

The limits of open-ended generative models are unclear, yet increasingly important. What causes them to succeed and what causes them to fail? In this paper, we take a prompt-centric approach to analyzing and bounding the abilities of open-ended generative models. We present a generic methodology of analysis with two challenging prompt constraint types: structural and stylistic. These constraint types are categorized into a set of well-defined constraints that are analyzable by a single prompt. We then systematically create a diverse set of simple, natural, and useful prompts to robustly analyze each individual constraint. Using the GPT-3 text-davinci-002 model as a case study, we generate outputs from our collection of prompts and analyze the model's generative failures. We also show the generalizability of our proposed method on other large models like BLOOM and OPT. Our results and our in-context mitigation strategies reveal open challenges for future research. We have publicly released our code at https://github.com/SALT-NLP/Bound-Cap-LLM.

Hongxin Zhang, Yanzhe Zhang, Ruiyi Zhang et al. · gatech

Demonstration-based learning has shown great potential in stimulating pretrained language models' ability under limited data scenario. Simply augmenting the input with some demonstrations can significantly improve performance on few-shot NER. However, why such demonstrations are beneficial for the learning process remains unclear since there is no explicit alignment between the demonstrations and the predictions. In this paper, we design pathological demonstrations by gradually removing intuitively useful information from the standard ones to take a deep dive of the robustness of demonstration-based sequence labeling and show that (1) demonstrations composed of random tokens still make the model a better few-shot learner; (2) the length of random demonstrations and the relevance of random tokens are the main factors affecting the performance; (3) demonstrations increase the confidence of model predictions on captured superficial patterns. We have publicly released our code at https://github.com/SALT-NLP/RobustDemo.

Vishakh Padmakumar, Lujain Ibrahim, Zora Zhiruo Wang et al.

AI tools are increasingly integrated into real-world workflows. However, existing measures of reliance on these tools focus on AI output adoption or on self-reported indicators, rather than how task effort is distributed between users and tools. Here, we introduce offloading score, a measure of reliance that quantifies the fraction of cognitive effort offloaded to an AI tool. Offloading Score is simulation-based -- we construct a counterfactual workflow by estimating how the user would have completed the task without the tool, and then computing the fraction of steps saved by using the tool. We validate offloading score through intrinsic evaluations of metric validity, and a controlled user study ($n=40$) with developers performing programming tasks using AI tools. We vary time pressure to test whether reliance measures capture the known increase in reliance under time pressure. We show that offloading score detects significantly higher reliance in time-constrained settings ($+43\%$, $p=0.018$), while usage-based and self-reported baseline measures of reliance do not distinguish the conditions. We complement this with descriptive insights showing that higher reliance manifests as greater delegation of subtasks to the tool and more direct reuse of AI outputs. Finally, we demonstrate an approach of using offloading score in combination with target outcomes of a task (e.g., code understanding) to identify when reliance may be (in)appropriate. Our framework offers two contributions: an instrument users can apply to measure and reflect on their own reliance, and a quantitative signal that agent designers can utilize to mitigate overreliance.

Chengwei Qin, Aston Zhang, Zhuosheng Zhang et al. · gatech

Spurred by advancements in scale, large language models (LLMs) have demonstrated the ability to perform a variety of natural language processing (NLP) tasks zero-shot -- i.e., without adaptation on downstream data. Recently, the debut of ChatGPT has drawn a great deal of attention from the natural language processing (NLP) community due to the fact that it can generate high-quality responses to human input and self-correct previous mistakes based on subsequent conversations. However, it is not yet known whether ChatGPT can serve as a generalist model that can perform many NLP tasks zero-shot. In this work, we empirically analyze the zero-shot learning ability of ChatGPT by evaluating it on 20 popular NLP datasets covering 7 representative task categories. With extensive empirical studies, we demonstrate both the effectiveness and limitations of the current version of ChatGPT. We find that ChatGPT performs well on many tasks favoring reasoning capabilities (e.g., arithmetic reasoning) while it still faces challenges when solving specific tasks such as sequence tagging. We additionally provide in-depth analysis through qualitative case studies.

Hui Chen, Wei Han, Diyi Yang et al.

This paper proposes a simple yet effective interpolation-based data augmentation approach termed DoubleMix, to improve the robustness of models in text classification. DoubleMix first leverages a couple of simple augmentation operations to generate several perturbed samples for each training data, and then uses the perturbed data and original data to carry out a two-step interpolation in the hidden space of neural models. Concretely, it first mixes up the perturbed data to a synthetic sample and then mixes up the original data and the synthetic perturbed data. DoubleMix enhances models' robustness by learning the "shifted" features in hidden space. On six text classification benchmark datasets, our approach outperforms several popular text augmentation methods including token-level, sentence-level, and hidden-level data augmentation techniques. Also, experiments in low-resource settings show our approach consistently improves models' performance when the training data is scarce. Extensive ablation studies and case studies confirm that each component of our approach contributes to the final performance and show that our approach exhibits superior performance on challenging counterexamples. Additionally, visual analysis shows that text features generated by our approach are highly interpretable. Our code for this paper can be found at https://github.com/declare-lab/DoubleMix.git.

James Seale Smith, Paola Cascante-Bonilla, Assaf Arbelle et al.

Recently, large-scale pre-trained Vision-and-Language (VL) foundation models have demonstrated remarkable capabilities in many zero-shot downstream tasks, achieving competitive results for recognizing objects defined by as little as short text prompts. However, it has also been shown that VL models are still brittle in Structured VL Concept (SVLC) reasoning, such as the ability to recognize object attributes, states, and inter-object relations. This leads to reasoning mistakes, which need to be corrected as they occur by teaching VL models the missing SVLC skills; often this must be done using private data where the issue was found, which naturally leads to a data-free continual (no task-id) VL learning setting. In this work, we introduce the first Continual Data-Free Structured VL Concepts Learning (ConStruct-VL) benchmark and show it is challenging for many existing data-free CL strategies. We, therefore, propose a data-free method comprised of a new approach of Adversarial Pseudo-Replay (APR) which generates adversarial reminders of past tasks from past task models. To use this method efficiently, we also propose a continual parameter-efficient Layered-LoRA (LaLo) neural architecture allowing no-memory-cost access to all past models at train time. We show this approach outperforms all data-free methods by as much as ~7% while even matching some levels of experience-replay (prohibitive for applications where data-privacy must be preserved). Our code is publicly available at https://github.com/jamessealesmith/ConStruct-VL

Aaron Reich, Jiaao Chen, Aastha Agrawal et al. · gatech

Named Entity Recognition (NER) systems often demonstrate great performance on in-distribution data, but perform poorly on examples drawn from a shifted distribution. One way to evaluate the generalization ability of NER models is to use adversarial examples, on which the specific variations associated with named entities are rarely considered. To this end, we propose leveraging expert-guided heuristics to change the entity tokens and their surrounding contexts thereby altering their entity types as adversarial attacks. Using expert-guided heuristics, we augmented the CoNLL 2003 test set and manually annotated it to construct a high-quality challenging set. We found that state-of-the-art NER systems trained on CoNLL 2003 training data drop performance dramatically on our challenging set. By training on adversarial augmented training examples and using mixup for regularization, we were able to significantly improve the performance on the challenging set as well as improve out-of-domain generalization which we evaluated by using OntoNotes data. We have publicly released our dataset and code at https://github.com/GT-SALT/Guided-Adversarial-Augmentation.

Mukund Rungta, Janvijay Singh, Saif M. Mohammad et al.

In a fair world, people have equitable opportunities to education, to conduct scientific research, to publish, and to get credit for their work, regardless of where they live. However, it is common knowledge among researchers that a vast number of papers accepted at top NLP venues come from a handful of western countries and (lately) China; whereas, very few papers from Africa and South America get published. Similar disparities are also believed to exist for paper citation counts. In the spirit of "what we do not measure, we cannot improve", this work asks a series of questions on the relationship between geographical location and publication success (acceptance in top NLP venues and citation impact). We first created a dataset of 70,000 papers from the ACL Anthology, extracted their meta-information, and generated their citation network. We then show that not only are there substantial geographical disparities in paper acceptance and citation but also that these disparities persist even when controlling for a number of variables such as venue of publication and sub-field of NLP. Further, despite some steps taken by the NLP community to improve geographical diversity, we show that the disparity in publication metrics across locations is still on an increasing trend since the early 2000s. We release our code and dataset here: https://github.com/iamjanvijay/acl-cite-net

Caleb Ziems, William Held, Omar Shaikh et al. · gatech

Large Language Models (LLMs) are capable of successfully performing many language processing tasks zero-shot (without training data). If zero-shot LLMs can also reliably classify and explain social phenomena like persuasiveness and political ideology, then LLMs could augment the Computational Social Science (CSS) pipeline in important ways. This work provides a road map for using LLMs as CSS tools. Towards this end, we contribute a set of prompting best practices and an extensive evaluation pipeline to measure the zero-shot performance of 13 language models on 25 representative English CSS benchmarks. On taxonomic labeling tasks (classification), LLMs fail to outperform the best fine-tuned models but still achieve fair levels of agreement with humans. On free-form coding tasks (generation), LLMs produce explanations that often exceed the quality of crowdworkers' gold references. We conclude that the performance of today's LLMs can augment the CSS research pipeline in two ways: (1) serving as zero-shot data annotators on human annotation teams, and (2) bootstrapping challenging creative generation tasks (e.g., explaining the underlying attributes of a text). In summary, LLMs are posed to meaningfully participate in social science analysis in partnership with humans.

Jingfeng Yang, Haoming Jiang, Qingyu Yin et al. · amazon-science, berkeley

Recent research showed promising results on combining pretrained language models (LMs) with canonical utterance for few-shot semantic parsing. The canonical utterance is often lengthy and complex due to the compositional structure of formal languages. Learning to generate such canonical utterance requires significant amount of data to reach high performance. Fine-tuning with only few-shot samples, the LMs can easily forget pretrained knowledge, overfit spurious biases, and suffer from compositionally out-of-distribution generalization errors. To tackle these issues, we propose a novel few-shot semantic parsing method -- SeqZero. SeqZero decomposes the problem into a sequence of sub-problems, which correspond to the sub-clauses of the formal language. Based on the decomposition, the LMs only need to generate short answers using prompts for predicting sub-clauses. Thus, SeqZero avoids generating a long canonical utterance at once. Moreover, SeqZero employs not only a few-shot model but also a zero-shot model to alleviate the overfitting. In particular, SeqZero brings out the merits from both models via ensemble equipped with our proposed constrained rescaling. SeqZero achieves SOTA performance of BART-based models on GeoQuery and EcommerceQuery, which are two few-shot datasets with compositional data split.

Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu et al. · gatech

Instruction tuning unlocks the superior capability of Large Language Models (LLM) to interact with humans. Furthermore, recent instruction-following datasets include images as visual inputs, collecting responses for image-based instructions. However, visual instruction-tuned models cannot comprehend textual details within images well. This work enhances the current visual instruction tuning pipeline with text-rich images (e.g., movie posters, book covers, etc.). Specifically, we first use publicly available OCR tools to collect results on 422K text-rich images from the LAION dataset. Moreover, we prompt text-only GPT-4 with recognized texts and image captions to generate 16K conversations, each containing question-answer pairs for text-rich images. By combining our collected data with previous multi-modal instruction-following data, our model, LLaVAR, substantially improves the LLaVA model's capability on text-based VQA datasets (up to 20% accuracy improvement) while achieving an accuracy of 91.42% on ScienceQA. The GPT-4-based instruction-following evaluation also demonstrates the improvement of our model on both natural images and text-rich images. Through qualitative analysis, LLaVAR shows promising interaction (e.g., reasoning, writing, and elaboration) skills with humans based on the latest real-world online content that combines text and images. We make our code/data/models publicly available at https://llavar.github.io/.

Omar Shaikh, Hongxin Zhang, William Held et al. · gatech

Generating a Chain of Thought (CoT) has been shown to consistently improve large language model (LLM) performance on a wide range of NLP tasks. However, prior work has mainly focused on logical reasoning tasks (e.g. arithmetic, commonsense QA); it remains unclear whether improvements hold for more diverse types of reasoning, especially in socially situated contexts. Concretely, we perform a controlled evaluation of zero-shot CoT across two socially sensitive domains: harmful questions and stereotype benchmarks. We find that zero-shot CoT reasoning in sensitive domains significantly increases a model's likelihood to produce harmful or undesirable output, with trends holding across different prompt formats and model variants. Furthermore, we show that harmful CoTs increase with model size, but decrease with improved instruction following. Our work suggests that zero-shot CoT should be used with caution on socially important tasks, especially when marginalized groups or sensitive topics are involved.

Zijun Liu, Yanzhe Zhang, Peng Li et al. · gatech, tsinghua

Recent studies show that collaborating multiple large language model (LLM) powered agents is a promising way for task solving. However, current approaches are constrained by using a fixed number of agents and static communication structures. In this work, we propose automatically selecting a team of agents from candidates to collaborate in a dynamic communication structure toward different tasks and domains. Specifically, we build a framework named Dynamic LLM-Powered Agent Network ($\textbf{DyLAN}$) for LLM-powered agent collaboration, operating a two-stage paradigm: (1) Team Optimization and (2) Task Solving. During the first stage, we utilize an $\textit{agent selection}$ algorithm, based on an unsupervised metric called $\textit{Agent Importance Score}$, enabling the selection of best agents according to their contributions in a preliminary trial, oriented to the given task. Then, in the second stage, the selected agents collaborate dynamically according to the query. Empirically, we demonstrate that DyLAN outperforms strong baselines in code generation, decision-making, general reasoning, and arithmetic reasoning tasks with moderate computational cost. On specific subjects in MMLU, selecting a team of agents in the team optimization stage improves accuracy by up to 25.0% in DyLAN.

Minzhi Li, Taiwei Shi, Caleb Ziems et al.

Annotated data plays a critical role in Natural Language Processing (NLP) in training models and evaluating their performance. Given recent developments in Large Language Models (LLMs), models such as ChatGPT demonstrate zero-shot capability on many text-annotation tasks, comparable with or even exceeding human annotators. Such LLMs can serve as alternatives for manual annotation, due to lower costs and higher scalability. However, limited work has leveraged LLMs as complementary annotators, nor explored how annotation work is best allocated among humans and LLMs to achieve both quality and cost objectives. We propose CoAnnotating, a novel paradigm for Human-LLM co-annotation of unstructured texts at scale. Under this framework, we utilize uncertainty to estimate LLMs' annotation capability. Our empirical study shows CoAnnotating to be an effective means to allocate work from results on different datasets, with up to 21% performance improvement over random baseline. For code implementation, see https://github.com/SALT-NLP/CoAnnotating.

Yanzhe Zhang, Lu Jiang, Greg Turk et al. · gatech

Text-to-image models, which can generate high-quality images based on textual input, have recently enabled various content-creation tools. Despite significantly affecting a wide range of downstream applications, the distributions of these generated images are still not fully understood, especially when it comes to the potential stereotypical attributes of different genders. In this work, we propose a paradigm (Gender Presentation Differences) that utilizes fine-grained self-presentation attributes to study how gender is presented differently in text-to-image models. By probing gender indicators in the input text (e.g., "a woman" or "a man"), we quantify the frequency differences of presentation-centric attributes (e.g., "a shirt" and "a dress") through human annotation and introduce a novel metric: GEP. Furthermore, we propose an automatic method to estimate such differences. The automatic GEP metric based on our approach yields a higher correlation with human annotations than that based on existing CLIP scores, consistently across three state-of-the-art text-to-image models. Finally, we demonstrate the generalization ability of our metrics in the context of gender stereotypes related to occupations.

Jiaao Chen, Aston Zhang, Xingjian Shi et al. · gatech

Parameter-efficient fine-tuning aims to achieve performance comparable to fine-tuning, using fewer trainable parameters. Several strategies (e.g., Adapters, prefix tuning, BitFit, and LoRA) have been proposed. However, their designs are hand-crafted separately, and it remains unclear whether certain design patterns exist for parameter-efficient fine-tuning. Thus, we present a parameter-efficient fine-tuning design paradigm and discover design patterns that are applicable to different experimental settings. Instead of focusing on designing another individual tuning strategy, we introduce parameter-efficient fine-tuning design spaces that parameterize tuning structures and tuning strategies. Specifically, any design space is characterized by four components: layer grouping, trainable parameter allocation, tunable groups, and strategy assignment. Starting from an initial design space, we progressively refine the space based on the model quality of each design choice and make greedy selection at each stage over these four components. We discover the following design patterns: (i) group layers in a spindle pattern; (ii) allocate the number of trainable parameters to layers uniformly; (iii) tune all the groups; (iv) assign proper tuning strategies to different groups. These design patterns result in new parameter-efficient fine-tuning methods. We show experimentally that these methods consistently and significantly outperform investigated parameter-efficient fine-tuning strategies across different backbone models and different tasks in natural language processing.

Jiaao Chen, Diyi Yang · gatech

Large language models (LLMs) have achieved significant progress from pre-training on and memorizing a wide range of textual data, however, this process might suffer from privacy issues and violations of data protection regulations. As a result, the ability to easily remove data related to individual users from such models while not deteriorating their predictive quality after the removal becomes increasingly important. To address these issues, in this work, we propose an efficient unlearning framework that could efficiently update LLMs without having to retrain the whole model after data removals, by introducing lightweight unlearning layers learned with a selective teacher-student objective into the transformers. In addition, we introduce a fusion mechanism to effectively combine different unlearning layers that learns to forget different sets of data to handle a sequence of forgetting operations. Experiments on classification and generation tasks demonstrate the effectiveness of our proposed methods compared to the state-of-the-art baselines.

Caleb Ziems, William Held, Jingfeng Yang et al. · amazon-science, gatech

Dialect differences caused by regional, social, and economic factors cause performance discrepancies for many groups of language technology users. Inclusive and equitable language technology must critically be dialect invariant, meaning that performance remains constant over dialectal shifts. Current systems often fall short of this ideal since they are designed and tested on a single dialect: Standard American English (SAE). We introduce a suite of resources for evaluating and achieving English dialect invariance. The resource is called Multi-VALUE, a controllable rule-based translation system spanning 50 English dialects and 189 unique linguistic features. Multi-VALUE maps SAE to synthetic forms of each dialect. First, we use this system to stress tests question answering, machine translation, and semantic parsing. Stress tests reveal significant performance disparities for leading models on non-standard dialects. Second, we use this system as a data augmentation technique to improve the dialect robustness of existing systems. Finally, we partner with native speakers of Chicano and Indian English to release new gold-standard variants of the popular CoQA task. To execute the transformation code, run model checkpoints, and download both synthetic and gold-standard dialectal benchmark datasets, see http://value-nlp.org.

Jiaao Chen, Aston Zhang, Mu Li et al. · gatech

Diffusion models that are based on iterative denoising have been recently proposed and leveraged in various generation tasks like image generation. Whereas, as a way inherently built for continuous data, existing diffusion models still have some limitations in modeling discrete data, e.g., languages. For example, the generally used Gaussian noise can not handle the discrete corruption well, and the objectives in continuous spaces fail to be stable for textual data in the diffusion process especially when the dimension is high. To alleviate these issues, we introduce a novel diffusion model for language modeling, Masked-Diffuse LM, with lower training cost and better performances, inspired by linguistic features in languages. Specifically, we design a linguistic-informed forward process which adds corruptions to the text through strategically soft-masking to better noise the textual data. Also, we directly predict the categorical distribution with cross-entropy loss function in every diffusion step to connect the continuous space and discrete space in a more efficient and straightforward way. Through experiments on 5 controlled generation tasks, we demonstrate that our Masked-Diffuse LM can achieve better generation quality than the state-of-the-art diffusion models with better efficiency.

Le Zhang, Zichao Yang, Diyi Yang · mila

Data augmentation is an effective approach to tackle over-fitting. Many previous works have proposed different data augmentations strategies for NLP, such as noise injection, word replacement, back-translation etc. Though effective, they missed one important characteristic of language--compositionality, meaning of a complex expression is built from its sub-parts. Motivated by this, we propose a compositional data augmentation approach for natural language understanding called TreeMix. Specifically, TreeMix leverages constituency parsing tree to decompose sentences into constituent sub-structures and the Mixup data augmentation technique to recombine them to generate new sentences. Compared with previous approaches, TreeMix introduces greater diversity to the samples generated and encourages models to learn compositionality of NLP data. Extensive experiments on text classification and SCAN demonstrate that TreeMix outperforms current state-of-the-art data augmentation methods.

Omar Shaikh, Caleb Ziems, William Held et al. · gatech

Pragmatic reference enables efficient interpersonal communication. Prior work uses simple reference games to test models of pragmatic reasoning, often with unidentified speakers and listeners. In practice, however, speakers' sociocultural background shapes their pragmatic assumptions. For example, readers of this paper assume NLP refers to "Natural Language Processing," and not "Neuro-linguistic Programming." This work introduces the Cultural Codes dataset, which operationalizes sociocultural pragmatic inference in a simple word reference game. Cultural Codes is based on the multi-turn collaborative two-player game, Codenames Duet. Our dataset consists of 794 games with 7,703 turns, distributed across 153 unique players. Alongside gameplay, we collect information about players' personalities, values, and demographics. Utilizing theories of communication and pragmatics, we predict each player's actions via joint modeling of their sociocultural priors and the game context. Our experiments show that accounting for background characteristics significantly improves model performance for tasks related to both clue giving and guessing, indicating that sociocultural priors play a vital role in gameplay decisions.

Bolin Lai, Hongxin Zhang, Miao Liu et al. · gatech

Persuasion modeling is a key building block for conversational agents. Existing works in this direction are limited to analyzing textual dialogue corpus. We argue that visual signals also play an important role in understanding human persuasive behaviors. In this paper, we introduce the first multimodal dataset for modeling persuasion behaviors. Our dataset includes 199 dialogue transcriptions and videos captured in a multi-player social deduction game setting, 26,647 utterance level annotations of persuasion strategy, and game level annotations of deduction game outcomes. We provide extensive experiments to show how dialogue context and visual signals benefit persuasion strategy prediction. We also explore the generalization ability of language models for persuasion modeling and the role of persuasion strategies in predicting social deduction game outcomes. Our dataset, code, and models can be found at https://persuasion-deductiongame.socialai-data.org.

Jingfeng Yang, Le Zhang, Diyi Yang · amazon-science, mila

Although sequence-to-sequence models often achieve good performance in semantic parsing for i.i.d. data, their performance is still inferior in compositional generalization. Several data augmentation methods have been proposed to alleviate this problem. However, prior work only leveraged superficial grammar or rules for data augmentation, which resulted in limited improvement. We propose to use subtree substitution for compositional data augmentation, where we consider subtrees with similar semantic functions as exchangeable. Our experiments showed that such augmented data led to significantly better performance on SCAN and GeoQuery, and reached new SOTA on compositional split of GeoQuery.

Jing Huang, Diyi Yang, Christopher Potts · stanford

Large Language Models (LLMs) frequently memorize long sequences verbatim, often with serious legal and privacy implications. Much prior work has studied such verbatim memorization using observational data. To complement such work, we develop a framework to study verbatim memorization in a controlled setting by continuing pre-training from Pythia checkpoints with injected sequences. We find that (1) non-trivial amounts of repetition are necessary for verbatim memorization to happen; (2) later (and presumably better) checkpoints are more likely to verbatim memorize sequences, even for out-of-distribution sequences; (3) the generation of memorized sequences is triggered by distributed model states that encode high-level features and makes important use of general language modeling capabilities. Guided by these insights, we develop stress tests to evaluate unlearning methods and find they often fail to remove the verbatim memorized information, while also degrading the LM. Overall, these findings challenge the hypothesis that verbatim memorization stems from specific model weights or mechanisms. Rather, verbatim memorization is intertwined with the LM's general capabilities and thus will be very difficult to isolate and suppress without degrading model quality.

Mingyu Derek Ma, Alexander K. Taylor, Nuan Wen et al. · cmu, harvard

We present MIDDAG, an intuitive, interactive system that visualizes the information propagation paths on social media triggered by COVID-19-related news articles accompanied by comprehensive insights, including user/community susceptibility level, as well as events and popular opinions raised by the crowd while propagating the information. Besides discovering information flow patterns among users, we construct communities among users and develop the propagation forecasting capability, enabling tracing and understanding of how information is disseminated at a higher level.

Ryan Louie, Ananjan Nandi, William Fang et al.

Recent works leverage LLMs to roleplay realistic social scenarios, aiding novices in practicing their social skills. However, simulating sensitive interactions, such as in mental health, is challenging. Privacy concerns restrict data access, and collecting expert feedback, although vital, is laborious. To address this, we develop Roleplay-doh, a novel human-LLM collaboration pipeline that elicits qualitative feedback from a domain-expert, which is transformed into a set of principles, or natural language rules, that govern an LLM-prompted roleplay. We apply this pipeline to enable senior mental health supporters to create customized AI patients for simulated practice partners for novice counselors. After uncovering issues in GPT-4 simulations not adhering to expert-defined principles, we also introduce a novel principle-adherence prompting pipeline which shows 30% improvements in response quality and principle following for the downstream task. Via a user study with 25 counseling experts, we demonstrate that the pipeline makes it easy and effective to create AI patients that more faithfully resemble real patients, as judged by creators and third-party counselors. See our project website at https://roleplay-doh.github.io/ for code and data.

Jiaao Chen, Mohan Dodda, Diyi Yang · gatech

Abstractive dialogue summarization has received increasing attention recently. Despite the fact that most of the current dialogue summarization systems are trained to maximize the likelihood of human-written summaries and have achieved significant results, there is still a huge gap in generating high-quality summaries as determined by humans, such as coherence and faithfulness, partly due to the misalignment in maximizing a single human-written summary. To this end, we propose to incorporate different levels of human feedback into the training process. This will enable us to guide the models to capture the behaviors humans care about for summaries. Specifically, we ask humans to highlight the salient information to be included in summaries to provide the local feedback , and to make overall comparisons among summaries in terms of coherence, accuracy, coverage, concise and overall quality, as the global feedback. We then combine both local and global feedback to fine-tune the dialog summarization policy with Reinforcement Learning. Experiments conducted on multiple datasets demonstrate the effectiveness and generalization of our methods over the state-of-the-art supervised baselines, especially in terms of human judgments.

Yuting Ning, Zhehao Zhang, Yash Kumar Lal et al.

Agent skills occupy a privileged position in the agent workflow, as agents are expected to implicitly follow and execute them, rendering third-party skills a vulnerable attack surface. Existing studies have revealed unsafe agent behaviors induced by skill-based attacks, but they primarily evaluate poisoned skills within a single task execution and enumerate harms through ad-hoc risk lists. To bridge these gaps, we introduce SkillHarm, a benchmark of skill-based attacks across the skill-use lifecycle, paired with a systematic taxonomy of skill-relevant risks. SkillHarm evaluates two attack scenarios: Fixed-Payload Poisoning (FPP), where a fixed poisoned skill package directly compromises any task session that invokes it, and Self-Mutating Poisoning (SMP), where an initially benign execution silently mutates persistent skill content, deferring harm until a subsequent reuse. It further defines 12 risk types based on the agent workflow component targeted by the harm: data pipelines, system environments, and agent autonomy. To instantiate these attacks at scale, we build AutoSkillHarm, an automated construction pipeline with coding agents driven by natural-language harnesses. The resulting benchmark contains 879 attack samples across 71 skills. Experiments show that current agents remain vulnerable with attack success rates up to 86.3% in FPP and 69.3% in SMP. Our analysis further reveals a latent risk: many apparent attack failures stem from the agent failing to engage with the poisoned file rather than genuine resistance, and current defenses still fail to reliably mitigate the threat.

Omar Shaikh, Valentino Chai, Michele J. Gelfand et al.

Interpersonal conflict is an uncomfortable but unavoidable fact of life. Navigating conflict successfully is a skill -- one that can be learned through deliberate practice -- but few have access to effective training or feedback. To expand this access, we introduce Rehearsal, a system that allows users to rehearse conflicts with a believable simulated interlocutor, explore counterfactual "what if?" scenarios to identify alternative conversational paths, and learn through feedback on how and when to apply specific conflict strategies. Users can utilize Rehearsal to practice handling a variety of predefined conflict scenarios, from office disputes to relationship issues, or they can choose to create their own setting. To enable Rehearsal, we develop IRP prompting, a method of conditioning output of a large language model on the influential Interest-Rights-Power (IRP) theory from conflict resolution. Rehearsal uses IRP to generate utterances grounded in conflict resolution theory, guiding users towards counterfactual conflict resolution strategies that help de-escalate difficult conversations. In a between-subjects evaluation, 40 participants engaged in an actual conflict with a confederate after training. Compared to a control group with lecture material covering the same IRP theory, participants with simulated training from Rehearsal significantly improved their performance in the unaided conflict: they reduced their use of escalating competitive strategies by an average of 67%, while doubling their use of cooperative strategies. Overall, Rehearsal highlights the potential effectiveness of language models as tools for learning and practicing interpersonal skills.

William Held, Diyi Yang · gatech

Multilingual transformer-based models demonstrate remarkable zero and few-shot transfer across languages by learning and reusing language-agnostic features. However, as a fixed-size model acquires more languages, its performance across all languages degrades, a phenomenon termed interference. Often attributed to limited model capacity, interference is commonly addressed by adding additional parameters despite evidence that transformer-based models are overparameterized. In this work, we show that it is possible to reduce interference by instead identifying and pruning language-specific parameters. First, we use Shapley Values, a credit allocation metric from coalitional game theory, to identify attention heads that introduce interference. Then, we show that removing identified attention heads from a fixed model improves performance for a target language on both sentence classification and structural prediction, seeing gains as large as 24.7\%. Finally, we provide insights on language-agnostic and language-specific attention heads using attention visualization.

Zora Zhiruo Wang, Sanidhya Vijayvargiya, Aspen Chen et al. · cmu

AI agents are increasingly developed and evaluated on benchmarks relevant to human work, yet it remains unclear how representative these benchmarking efforts are of the labor market as a whole. In this work, we systematically study the relationship between agent development efforts and the distribution of real-world human work by mapping benchmark instances to work domains and skills. We first analyze 43 benchmarks and 72,342 tasks, measuring their alignment with human employment and capital allocation across all 1,016 real-world occupations in the U.S. labor market. We reveal substantial mismatches between agent development that tends to be programming-centric, and the categories in which human labor and economic value are concentrated. Within work areas that agents currently target, we further characterize current agent utility by measuring their autonomy levels, providing practical guidance for agent interaction strategies across work scenarios. Building on these findings, we propose three measurable principles for designing benchmarks that better capture socially important and technically challenging forms of work: coverage, realism, and granular evaluation.

Raj Sanjay Shah, Faye Holt, Shirley Anugrah Hayati et al.

Millions of people participate in online peer-to-peer support sessions, yet there has been little prior research on systematic psychology-based evaluations of fine-grained peer-counselor behavior in relation to client satisfaction. This paper seeks to bridge this gap by mapping peer-counselor chat-messages to motivational interviewing (MI) techniques. We annotate 14,797 utterances from 734 chat conversations using 17 MI techniques and introduce four new interviewing codes such as chit-chat and inappropriate to account for the unique conversational patterns observed on online platforms. We automate the process of labeling peer-counselor responses to MI techniques by fine-tuning large domain-specific language models and then use these automated measures to investigate the behavior of the peer counselors via correlational studies. Specifically, we study the impact of MI techniques on the conversation ratings to investigate the techniques that predict clients' satisfaction with their counseling sessions. When counselors use techniques such as reflection and affirmation, clients are more satisfied. Examining volunteer counselors' change in usage of techniques suggest that counselors learn to use more introduction and open questions as they gain experience. This work provides a deeper understanding of the use of motivational interviewing techniques on peer-to-peer counselor platforms and sheds light on how to build better training programs for volunteer counselors on online platforms.

William Held, Christopher Hidey, Fei Liu et al. · gatech

Modern virtual assistants use internal semantic parsing engines to convert user utterances to actionable commands. However, prior work has demonstrated that semantic parsing is a difficult multilingual transfer task with low transfer efficiency compared to other tasks. In global markets such as India and Latin America, this is a critical issue as switching between languages is prevalent for bilingual users. In this work we dramatically improve the zero-shot performance of a multilingual and codeswitched semantic parsing system using two stages of multilingual alignment. First, we show that constrastive alignment pretraining improves both English performance and transfer efficiency. We then introduce a constrained optimization approach for hyperparameter-free adversarial alignment during finetuning. Our Doubly Aligned Multilingual Parser (DAMP) improves mBERT transfer performance by 3x, 6x, and 81x on the Spanglish, Hinglish and Multilingual Task Oriented Parsing benchmarks respectively and outperforms XLM-R and mT5-Large using 3.2x fewer parameters.

Caleb Ziems, Minzhi Li, Anthony Zhang et al.

Sentiment transfer is one popular example of a text style transfer task, where the goal is to reverse the sentiment polarity of a text. With a sentiment reversal comes also a reversal in meaning. We introduce a different but related task called positive reframing in which we neutralize a negative point of view and generate a more positive perspective for the author without contradicting the original meaning. Our insistence on meaning preservation makes positive reframing a challenging and semantically rich task. To facilitate rapid progress, we introduce a large-scale benchmark, Positive Psychology Frames, with 8,349 sentence pairs and 12,755 structured annotations to explain positive reframing in terms of six theoretically-motivated reframing strategies. Then we evaluate a set of state-of-the-art text style transfer models, and conclude by discussing key challenges and directions for future work.

Patsorn Sangkloy, Wittawat Jitkrittum, Diyi Yang et al.

We address the problem of retrieving images with both a sketch and a text query. We present TASK-former (Text And SKetch transformer), an end-to-end trainable model for image retrieval using a text description and a sketch as input. We argue that both input modalities complement each other in a manner that cannot be achieved easily by either one alone. TASK-former follows the late-fusion dual-encoder approach, similar to CLIP, which allows efficient and scalable retrieval since the retrieval set can be indexed independently of the queries. We empirically demonstrate that using an input sketch (even a poorly drawn one) in addition to text considerably increases retrieval recall compared to traditional text-based image retrieval. To evaluate our approach, we collect 5,000 hand-drawn sketches for images in the test set of the COCO dataset. The collected sketches are available a https://janesjanes.github.io/tsbir/.

William Held, Camille Harris, Michael Best et al. · gatech

Coloniality, the continuation of colonial harms beyond "official" colonization, has pervasive effects across society and scientific fields. Natural Language Processing (NLP) is no exception to this broad phenomenon. In this work, we argue that coloniality is implicitly embedded in and amplified by NLP data, algorithms, and software. We formalize this analysis using Actor-Network Theory (ANT): an approach to understanding social phenomena through the network of relationships between human stakeholders and technology. We use our Actor-Network to guide a quantitative survey of the geography of different phases of NLP research, providing evidence that inequality along colonial boundaries increases as NLP builds on itself. Based on this, we argue that combating coloniality in NLP requires not only changing current values but also active work to remove the accumulation of colonial ideals in our foundational data and algorithms.

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.

John Yang, Kilian Lieret, Joyce Yang et al.

Current benchmarks for coding evaluate language models (LMs) on concrete, well-specified tasks such as fixing specific bugs or writing targeted tests. However, human programmers do not spend all day incessantly addressing isolated tasks. Instead, real-world software development is grounded in the pursuit of high-level goals, like improving user retention or reducing costs. Evaluating whether LMs can also iteratively develop code to better accomplish open-ended objectives without any explicit guidance remains an open challenge. To address this, we introduce CodeClash, a benchmark where LMs compete in multi-round tournaments to build the best codebase for achieving a competitive objective. Each round proceeds in two phases: agents edit their code, then their codebases compete head-to-head in a code arena that determines winners based on objectives like score maximization, resource acquisition, or survival. Whether it's writing notes, scrutinizing documentation, analyzing competition logs, or creating test suites, models must decide for themselves how to improve their codebases both absolutely and against their opponents. We run 1680 tournaments (25,200 rounds total) to evaluate 8 LMs across 6 arenas. Our results reveal that while models exhibit diverse development styles, they share fundamental limitations in strategic reasoning. Models also struggle with long-term codebase maintenance, as repositories become progressively messy and redundant. These limitations are stark: top models lose every round against expert human programmers. We open-source CodeClash to advance the study of autonomous, goal-oriented code development.

Eric Zelikman, Wanjing Anya Ma, Jasmine E. Tran et al.

Developing an educational test can be expensive and time-consuming, as each item must be written by experts and then evaluated by collecting hundreds of student responses. Moreover, many tests require multiple distinct sets of questions administered throughout the school year to closely monitor students' progress, known as parallel tests. In this study, we focus on tests of silent sentence reading efficiency, used to assess students' reading ability over time. To generate high-quality parallel tests, we propose to fine-tune large language models (LLMs) to simulate how previous students would have responded to unseen items. With these simulated responses, we can estimate each item's difficulty and ambiguity. We first use GPT-4 to generate new test items following a list of expert-developed rules and then apply a fine-tuned LLM to filter the items based on criteria from psychological measurements. We also propose an optimal-transport-inspired technique for generating parallel tests and show the generated tests closely correspond to the original test's difficulty and reliability based on crowdworker responses. Our evaluation of a generated test with 234 students from grades 2 to 8 produces test scores highly correlated (r=0.93) to those of a standard test form written by human experts and evaluated across thousands of K-12 students.

Omar Shaikh, Valentin Teutschbein, Kanishk Gandhi et al.

Truly proactive AI systems must anticipate what we will do next. This foresight demands far richer information than the sparse signals we type into our prompts -- it demands reasoning over the entire context of what we see and do. We formalize this as next action prediction (NAP): given a sequence of a user's multimodal interactions with a computer (screenshots, clicks, sensor data), predict that user's next action. Progress on this task requires both new data and modeling approaches. To scale data, we annotate longitudinal, naturalistic computer use with vision-language models. We release an open-source pipeline for performing this labeling on private infrastructure, and label over 360K actions across one month of continuous phone usage from 20 users, amounting to 1,800 hours of screen time. We then introduce LongNAP, a user model that combines parametric and in-context learning to reason over long interaction histories. LongNAP is trained via policy gradient methods to generate user-specific reasoning traces given some context; retrieve relevant traces from a library of past traces; and then apply retrieved traces in-context to predict future actions. Using an LLM-as-judge evaluation metric (0-1 similarity to ground truth), LongNAP significantly outperforms supervised finetuning and prompted baselines on held-out data (by 79% and 39% respectively). Additionally, LongNAP generalizes to held out users when trained across individuals. The space of next actions a user might take at any moment is unbounded, spanning thousands of possible outcomes. Despite this, 17.1% of LongNAP's predicted trajectories are well-aligned with what a user does next (LLM-judge score $\geq$ 0.5). This rises to 26% when we filter to highly confident predictions. In sum, we argue that learning from the full context of user behavior to anticipate user needs is now a viable task with substantial opportunity.