Eunsol Choi

Anuj Diwan, Layne Berry, Eunsol Choi et al.

Recent visuolinguistic pre-trained models show promising progress on various end tasks such as image retrieval and video captioning. Yet, they fail miserably on the recently proposed Winoground dataset, which challenges models to match paired images and English captions, with items constructed to overlap lexically but differ in meaning (e.g., "there is a mug in some grass" vs. "there is some grass in a mug"). By annotating the dataset using new fine-grained tags, we show that solving the Winoground task requires not just compositional language understanding, but a host of other abilities like commonsense reasoning or locating small, out-of-focus objects in low-resolution images. In this paper, we identify the dataset's main challenges through a suite of experiments on related tasks (probing task, image retrieval task), data augmentation, and manual inspection of the dataset. Our analysis suggests that a main challenge in visuolinguistic models may lie in fusing visual and textual representations, rather than in compositional language understanding. We release our annotation and code at https://github.com/ajd12342/why-winoground-hard .

Zeyu Leo Liu, Shrey Pandit, Xi Ye et al.

Large language models (LLMs) are increasingly being used to synthesize and reason about source code. However, the static nature of these models' knowledge does not reflect the fact that libraries and API functions they invoke are continuously evolving, with functionality being added or changing. While numerous benchmarks evaluate how LLMs can generate code, no prior work has studied how an LLMs' knowledge about code API functions can be updated. To fill this gap, we present CodeUpdateArena, a benchmark for knowledge editing in the code domain. An instance in our benchmark consists of a synthetic API function update paired with a program synthesis example that uses the updated functionality; our goal is to update an LLM to be able to solve this program synthesis example without providing documentation of the update at inference time. Compared to knowledge editing for facts encoded in text, success here is more challenging: a code LLM must correctly reason about the semantics of the modified function rather than just reproduce its syntax. Our dataset is constructed by first prompting GPT-4 to generate atomic and executable function updates. Then, for each update, we generate program synthesis examples whose code solutions are prone to use the update. Our benchmark covers updates of various types to 54 functions from seven diverse Python packages, with a total of 670 program synthesis examples. Our experiments show that prepending documentation of the update to open-source code LLMs (i.e., DeepSeek, CodeLlama) does not allow them to incorporate changes for problem solving, and existing knowledge editing techniques also have substantial room for improvement. We hope our benchmark will inspire new methods for knowledge updating in code LLMs.

Xinyan Velocity Yu, Akari Asai, Trina Chatterjee et al. · uw

While the NLP community is generally aware of resource disparities among languages, we lack research that quantifies the extent and types of such disparity. Prior surveys estimating the availability of resources based on the number of datasets can be misleading as dataset quality varies: many datasets are automatically induced or translated from English data. To provide a more comprehensive picture of language resources, we examine the characteristics of 156 publicly available NLP datasets. We manually annotate how they are created, including input text and label sources and tools used to build them, and what they study, tasks they address and motivations for their creation. After quantifying the qualitative NLP resource gap across languages, we discuss how to improve data collection in low-resource languages. We survey language-proficient NLP researchers and crowd workers per language, finding that their estimated availability correlates with dataset availability. Through crowdsourcing experiments, we identify strategies for collecting high-quality multilingual data on the Mechanical Turk platform. We conclude by making macro and micro-level suggestions to the NLP community and individual researchers for future multilingual data development.

Shufan Wang, Fangyuan Xu, Laure Thompson et al. · princeton

Exemplification is a process by which writers explain or clarify a concept by providing an example. While common in all forms of writing, exemplification is particularly useful in the task of long-form question answering (LFQA), where a complicated answer can be made more understandable through simple examples. In this paper, we provide the first computational study of exemplification in QA, performing a fine-grained annotation of different types of examples (e.g., hypotheticals, anecdotes) in three corpora. We show that not only do state-of-the-art LFQA models struggle to generate relevant examples, but also that standard evaluation metrics such as ROUGE are insufficient to judge exemplification quality. We propose to treat exemplification as a \emph{retrieval} problem in which a partially-written answer is used to query a large set of human-written examples extracted from a corpus. Our approach allows a reliable ranking-type automatic metrics that correlates well with human evaluation. A human evaluation shows that our model's retrieved examples are more relevant than examples generated from a state-of-the-art LFQA model.

Akari Asai, Shayne Longpre, Jungo Kasai et al. · uw

We present the results of the Workshop on Multilingual Information Access (MIA) 2022 Shared Task, evaluating cross-lingual open-retrieval question answering (QA) systems in 16 typologically diverse languages. In this task, we adapted two large-scale cross-lingual open-retrieval QA datasets in 14 typologically diverse languages, and newly annotated open-retrieval QA data in 2 underrepresented languages: Tagalog and Tamil. Four teams submitted their systems. The best system leveraging iteratively mined diverse negative examples and larger pretrained models achieves 32.2 F1, outperforming our baseline by 4.5 points. The second best system uses entity-aware contextualized representations for document retrieval, and achieves significant improvements in Tamil (20.8 F1), whereas most of the other systems yield nearly zero scores.

Fangyuan Xu, Weijia Shi, Eunsol Choi

Retrieving documents and prepending them in-context at inference time improves performance of language model (LMs) on a wide range of tasks. However, these documents, often spanning hundreds of words, make inference substantially more expensive. We propose compressing the retrieved documents into textual summaries prior to in-context integration. This not only reduces the computational costs but also relieves the burden of LMs to identify relevant information in long retrieved documents. We present two compressors -- an extractive compressor which selects useful sentences from retrieved documents and an abstractive compressor which generates summaries by synthesizing information from multiple documents. Both compressors are trained to improve LMs' performance on end tasks when the generated summaries are prepended to the LMs' input, while keeping the summary concise.If the retrieved documents are irrelevant to the input or offer no additional information to LM, our compressor can return an empty string, implementing selective augmentation.We evaluate our approach on language modeling task and open domain question answering task. We achieve a compression rate of as low as 6% with minimal loss in performance for both tasks, significantly outperforming the off-the-shelf summarization models. We show that our compressors trained for one LM can transfer to other LMs on the language modeling task and provide summaries largely faithful to the retrieved documents.

Jifan Chen, Aniruddh Sriram, Eunsol Choi et al.

Verifying complex political claims is a challenging task, especially when politicians use various tactics to subtly misrepresent the facts. Automatic fact-checking systems fall short here, and their predictions like "half-true" are not very useful in isolation, since we have no idea which parts of the claim are true and which are not. In this work, we focus on decomposing a complex claim into a comprehensive set of yes-no subquestions whose answers influence the veracity of the claim. We present ClaimDecomp, a dataset of decompositions for over 1000 claims. Given a claim and its verification paragraph written by fact-checkers, our trained annotators write subquestions covering both explicit propositions of the original claim and its implicit facets, such as asking about additional political context that changes our view of the claim's veracity. We study whether state-of-the-art models can generate such subquestions, showing that these models generate reasonable questions to ask, but predicting the comprehensive set of subquestions from the original claim without evidence remains challenging. We further show that these subquestions can help identify relevant evidence to fact-check the full claim and derive the veracity through their answers, suggesting that they can be useful pieces of a fact-checking pipeline.

Yasumasa Onoe, Michael J. Q. Zhang, Eunsol Choi et al.

Language models (LMs) are typically trained once on a large-scale corpus and used for years without being updated. However, in a dynamic world, new entities constantly arise. We propose a framework to analyze what LMs can infer about new entities that did not exist when the LMs were pretrained. We derive a dataset of entities indexed by their origination date and paired with their English Wikipedia articles, from which we can find sentences about each entity. We evaluate LMs' perplexity on masked spans within these sentences. We show that models more informed about the entities, such as those with access to a textual definition of them, achieve lower perplexity on this benchmark. Our experimental results demonstrate that making inferences about new entities remains difficult for LMs. Given its wide coverage on entity knowledge and temporal indexing, our dataset can be used to evaluate LMs and techniques designed to modify or extend their knowledge. Our automatic data collection pipeline can be easily used to continually update our benchmark.

Hung-Ting Chen, Michael J. Q. Zhang, Eunsol Choi

Question answering models can use rich knowledge sources -- up to one hundred retrieved passages and parametric knowledge in the large-scale language model (LM). Prior work assumes information in such knowledge sources is consistent with each other, paying little attention to how models blend information stored in their LM parameters with that from retrieved evidence documents. In this paper, we simulate knowledge conflicts (i.e., where parametric knowledge suggests one answer and different passages suggest different answers) and examine model behaviors. We find retrieval performance heavily impacts which sources models rely on, and current models mostly rely on non-parametric knowledge in their best-performing settings. We discover a troubling trend that contradictions among knowledge sources affect model confidence only marginally. To address this issue, we present a new calibration study, where models are discouraged from presenting any single answer when presented with multiple conflicting answer candidates in retrieved evidences.

Anuj Diwan, Eunsol Choi, David Harwath

We present the first unified study of the efficiency of self-attention-based Transformer variants spanning text, speech and vision. We identify input length thresholds (tipping points) at which efficient Transformer variants become more efficient than vanilla models, using a variety of efficiency metrics (latency, throughput, and memory). To conduct this analysis for speech, we introduce L-HuBERT, a novel local-attention variant of a self-supervised speech model. We observe that these thresholds are (a) much higher than typical dataset sequence lengths and (b) dependent on the metric and modality, showing that choosing the right model depends on modality, task type (long-form vs. typical context) and resource constraints (time vs. memory). By visualising the breakdown of the computational costs for transformer components, we also show that non-self-attention components exhibit significant computational costs. We release our profiling toolkit at https://github.com/ajd12342/profiling-transformers .

Shankar Padmanabhan, Yasumasa Onoe, Michael J. Q. Zhang et al.

Modern language models have the capacity to store and use immense amounts of knowledge about real-world entities, but it remains unclear how to update such knowledge stored in model parameters. While prior methods for updating knowledge in LMs successfully inject atomic facts, updated LMs fail to make inferences based on injected facts. In this work, we demonstrate that a context distillation-based approach can both impart knowledge about entities and propagate that knowledge to enable broader inferences. Our approach consists of two stages: transfer set generation and distillation on the transfer set. We first generate a transfer set by prompting a language model to generate continuations from the entity definition. Then, we update the model parameters so that the distribution of the LM (the student) matches the distribution of the LM conditioned on the definition (the teacher) on the transfer set. Our experiments demonstrate that this approach is more effective at propagating knowledge updates than fine-tuning and other gradient-based knowledge-editing methods. Moreover, it does not compromise performance in other contexts, even when injecting the definitions of up to 150 entities at once.

Anuj Diwan, Eunsol Choi, David Harwath

We introduce ParaSpeechCLAP, a dual-encoder contrastive model that maps speech and text style captions into a common embedding space, supporting a wide range of intrinsic (speaker-level) and situational (utterance-level) descriptors (such as pitch, texture and emotion) far beyond the narrow set handled by existing models. We train specialized ParaSpeechCLAP-Intrinsic and ParaSpeechCLAP-Situational models alongside a unified ParaSpeechCLAP-Combined model, finding that specialization yields stronger performance on individual style dimensions while the unified model excels on compositional evaluation. We further show that ParaSpeechCLAP-Intrinsic benefits from an additional classification loss and class-balanced training. We demonstrate our models' performance on style caption retrieval, speech attribute classification and as an inference-time reward model that improves style-prompted TTS without additional training. ParaSpeechCLAP outperforms baselines on most metrics across all three applications. Our models and code are released at https://github.com/ajd12342/paraspeechclap .

Christopher Mohri, Daniel Andor, Eunsol Choi et al.

We study the problem of classification with a reject option for a fixed predictor, applicable in natural language processing. We introduce a new problem formulation for this scenario, and an algorithm minimizing a new surrogate loss function. We provide a complete theoretical analysis of the surrogate loss function with a strong $H$-consistency guarantee. For evaluation, we choose the decontextualization task, and provide a manually-labelled dataset of $2\mathord,000$ examples. Our algorithm significantly outperforms the baselines considered, with a $\sim\!\!25\%$ improvement in coverage when halving the error rate, which is only $\sim\!\! 3 \%$ away from the theoretical limit.

Hung-Ting Chen, Fangyuan Xu, Shane Arora et al. · allen-ai

How retrieved documents are used in language models (LMs) for long-form generation task is understudied. We present two controlled studies on retrieval-augmented LM for long-form question answering (LFQA): one fixing the LM and varying evidence documents and the other fixing evidence documents and varying the LMs. We study various attributes of generated answers (e.g., fluency, length, variance), with an emphasis on the attribution of generated answers to in-context evidence documents. We collect a dataset (SALAD) containing human annotations of sentence-level answer attribution in LFQA and evaluate existing methods for automatically judging attribution. We find that while LMs can leverage relevant in-context documents, the generated answer is only partially attributable towards the documents, especially for LMs trained without retrieval augmentation. Together, our analysis reveals how retrieval augmentation impacts long knowledge-rich text generation and provide directions for future work.

Anirudh Srinivasan, Eunsol Choi

We study politeness phenomena in nine typologically diverse languages. Politeness is an important facet of communication and is sometimes argued to be cultural-specific, yet existing computational linguistic study is limited to English. We create TyDiP, a dataset containing three-way politeness annotations for 500 examples in each language, totaling 4.5K examples. We evaluate how well multilingual models can identify politeness levels -- they show a fairly robust zero-shot transfer ability, yet fall short of estimated human accuracy significantly. We further study mapping the English politeness strategy lexicon into nine languages via automatic translation and lexicon induction, analyzing whether each strategy's impact stays consistent across languages. Lastly, we empirically study the complicated relationship between formality and politeness through transfer experiments. We hope our dataset will support various research questions and applications, from evaluating multilingual models to constructing polite multilingual agents.

Fangyuan Xu, Junyi Jessy Li, Eunsol Choi

Long-form answers, consisting of multiple sentences, can provide nuanced and comprehensive answers to a broader set of questions. To better understand this complex and understudied task, we study the functional structure of long-form answers collected from three datasets, ELI5, WebGPT and Natural Questions. Our main goal is to understand how humans organize information to craft complex answers. We develop an ontology of six sentence-level functional roles for long-form answers, and annotate 3.9k sentences in 640 answer paragraphs. Different answer collection methods manifest in different discourse structures. We further analyze model-generated answers -- finding that annotators agree less with each other when annotating model-generated answers compared to annotating human-written answers. Our annotated data enables training a strong classifier that can be used for automatic analysis. We hope our work can inspire future research on discourse-level modeling and evaluation of long-form QA systems.

Ge Gao, Eunsol Choi, Yoav Artzi

We study learning from user feedback for extractive question answering by simulating feedback using supervised data. We cast the problem as contextual bandit learning, and analyze the characteristics of several learning scenarios with focus on reducing data annotation. We show that systems initially trained on a small number of examples can dramatically improve given feedback from users on model-predicted answers, and that one can use existing datasets to deploy systems in new domains without any annotation, but instead improving the system on-the-fly via user feedback.

Anuj Diwan, Ching-Feng Yeh, Wei-Ning Hsu et al.

Automatic speech recognition research focuses on training and evaluating on static datasets. Yet, as speech models are increasingly deployed on personal devices, such models encounter user-specific distributional shifts. To simulate this real-world scenario, we introduce LibriContinual, a continual learning benchmark for speaker-specific domain adaptation derived from LibriVox audiobooks, with data corresponding to 118 individual speakers and 6 train splits per speaker of different sizes. Additionally, current speech recognition models and continual learning algorithms are not optimized to be compute-efficient. We adapt a general-purpose training algorithm NetAug for ASR and create a novel Conformer variant called the DisConformer (Disentangled Conformer). This algorithm produces ASR models consisting of a frozen 'core' network for general-purpose use and several tunable 'augment' networks for speaker-specific tuning. Using such models, we propose a novel compute-efficient continual learning algorithm called DisentangledCL. Our experiments show that the DisConformer models significantly outperform baselines on general ASR i.e. LibriSpeech (15.58% rel. WER on test-other). On speaker-specific LibriContinual they significantly outperform trainable-parameter-matched baselines (by 20.65% rel. WER on test) and even match fully finetuned baselines in some settings.

Venelin Kovatchev, Trina Chatterjee, Venkata S Govindarajan et al.

Developing methods to adversarially challenge NLP systems is a promising avenue for improving both model performance and interpretability. Here, we describe the approach of the team "longhorns" on Task 1 of the The First Workshop on Dynamic Adversarial Data Collection (DADC), which asked teams to manually fool a model on an Extractive Question Answering task. Our team finished first, with a model error rate of 62%. We advocate for a systematic, linguistically informed approach to formulating adversarial questions, and we describe the results of our pilot experiments, as well as our official submission.

Nishant Balepur, Malachi Hamada, Varsha Kishore et al. · allen-ai

Deep Research (DR) tools (e.g. OpenAI DR) help researchers cope with ballooning publishing counts. Such tools can synthesize scientific papers to answer researchers' queries, but lack understanding of their users. We change that in MyScholarQA (MySQA), a personalized DR tool that: 1) infers a profile of a user's research interests; 2) proposes personalized actions for a user's input query; and 3) writes a multi-section report for the query that follows user-approved actions. We first test MySQA with NLP's standard protocol: we design a benchmark of synthetic users and LLM judges, where MySQA beats baselines in citation metrics and personalized action-following. However, we suspect this process does not cover all aspects of personalized DR users value, so we interview users in an online version of MySQA to unmask them. We reveal nine nuanced errors of personalized DR undetectable by our LLM judges, and we study qualitative feedback to form lessons for future DR design. In all, we argue for a pillar of personalization that easy-to-use LLM judges can lead NLP to overlook: real progress in personalization is only possible with real users.

Michael J. Q. Zhang, Eunsol Choi

Resolving ambiguities through interaction is a hallmark of natural language, and modeling this behavior is a core challenge in crafting AI assistants. In this work, we study such behavior in LMs by proposing a task-agnostic framework for resolving ambiguity by asking users clarifying questions. Our framework breaks down this objective into three subtasks: (1) determining when clarification is needed, (2) determining what clarifying question to ask, and (3) responding accurately with the new information gathered through clarification. We evaluate systems across three NLP applications: question answering, machine translation and natural language inference. For the first subtask, we present a novel uncertainty estimation approach, intent-sim, that determines the utility of querying for clarification by estimating the entropy over user intents. Our method consistently outperforms existing uncertainty estimation approaches at identifying predictions that will benefit from clarification. When only allowed to ask for clarification on 10% of examples, our system is able to double the performance gains over randomly selecting examples to clarify. Furthermore, we find that intent-sim is robust, demonstrating improvements across a wide range of NLP tasks and LMs. Together, our work lays foundation for studying clarifying interactions with LMs.

Mina Huh, Fangyuan Xu, Yi-Hao Peng et al.

Vision language models can now generate long-form answers to questions about images - long-form visual question answers (LFVQA). We contribute VizWiz-LF, a dataset of long-form answers to visual questions posed by blind and low vision (BLV) users. VizWiz-LF contains 4.2k long-form answers to 600 visual questions, collected from human expert describers and six VQA models. We develop and annotate functional roles of sentences of LFVQA and demonstrate that long-form answers contain information beyond the question answer such as explanations and suggestions. We further conduct automatic and human evaluations with BLV and sighted people to evaluate long-form answers. BLV people perceive both human-written and generated long-form answers to be plausible, but generated answers often hallucinate incorrect visual details, especially for unanswerable visual questions (e.g., blurry or irrelevant images). To reduce hallucinations, we evaluate the ability of VQA models to abstain from answering unanswerable questions across multiple prompting strategies.

Xuewen Yao, Miriam Mikhelson, S. Craig Watkins et al.

In collaboration with Postpartum Support International (PSI), a non-profit organization dedicated to supporting caregivers with postpartum mood and anxiety disorders, we developed three chatbots to provide context-specific empathetic support to postpartum caregivers, leveraging both rule-based and generative models. We present and evaluate the performance of our chatbots using both machine-based metrics and human-based questionnaires. Overall, our rule-based model achieves the best performance, with outputs that are close to ground truth reference and contain the highest levels of empathy. Human users prefer the rule-based chatbot over the generative chatbot for its context-specific and human-like replies. Our generative chatbot also produced empathetic responses and was described by human users as engaging. However, limitations in the training dataset often result in confusing or nonsensical responses. We conclude by discussing practical benefits of rule-based vs. generative models for supporting individuals with mental health challenges. In light of the recent surge of ChatGPT and BARD, we also discuss the possibilities and pitfalls of large language models for digital mental healthcare.

Xuewen Yao, Miriam Mikhelson, Megan Micheletti et al.

Digital platforms, including online forums and helplines, have emerged as avenues of support for caregivers suffering from postpartum mental health distress. Understanding support seekers' experiences as shared on these platforms could provide crucial insight into caregivers' needs during this vulnerable time. In the current work, we provide a descriptive analysis of the concerns, psychological states, and motivations shared by healthy and distressed postpartum support seekers on two digital platforms, a one-on-one digital helpline and a publicly available online forum. Using a combination of human annotations, dictionary models and unsupervised techniques, we find stark differences between the experiences of distressed and healthy mothers. Distressed mothers described interpersonal problems and a lack of support, with 8.60% - 14.56% reporting severe symptoms including suicidal ideation. In contrast, the majority of healthy mothers described childcare issues, such as questions about breastfeeding or sleeping, and reported no severe mental health concerns. Across the two digital platforms, we found that distressed mothers shared similar content. However, the patterns of speech and affect shared by distressed mothers differed between the helpline vs. the online forum, suggesting the design of these platforms may shape meaningful measures of their support-seeking experiences. Our results provide new insight into the experiences of caregivers suffering from postpartum mental health distress. We conclude by discussing methodological considerations for understanding content shared by support seekers and design considerations for the next generation of support tools for postpartum parents.

Jeremy R. Cole, Palak Jain, Julian Martin Eisenschlos et al.

Identifying the difference between two versions of the same article is useful to update knowledge bases and to understand how articles evolve. Paired texts occur naturally in diverse situations: reporters write similar news stories and maintainers of authoritative websites must keep their information up to date. We propose representing factual changes between paired documents as question-answer pairs, where the answer to the same question differs between two versions. We find that question-answer pairs can flexibly and concisely capture the updated contents. Provided with paired documents, annotators identify questions that are answered by one passage but answered differently or cannot be answered by the other. We release DIFFQG which consists of 759 QA pairs and 1153 examples of paired passages with no factual change. These questions are intended to be both unambiguous and information-seeking and involve complex edits, pushing beyond the capabilities of current question generation and factual change detection systems. Our dataset summarizes the changes between two versions of the document as questions and answers, studying automatic update summarization in a novel way.

Nishant Balepur, Bhavya Rajasekaran, Jane Oh et al.

Multiple-choice question answering (MCQA) is standard in NLP, but benchmarks lack rigorous quality control. We present BenchMarker, an education-inspired toolkit using LLM judges to flag three common MCQ flaws: 1) contamination - items appearing exactly online; 2) shortcuts - cues in the choices that enable guessing; and 3) writing errors - structural/grammatical issues based on a 19-rule education rubric. We validate BenchMarker with human annotations, then run the tool to audit 12 benchmarks, revealing: 2) contaminated MCQs tend to inflate accuracy, while writing errors tend to lower it and change rankings beyond random; and 3) prior benchmark repairs address their targeted issues (i.e., lowering accuracy with LLM-written distractors), but inadvertently add new flaws (i.e. implausible distractors, many correct answers). Overall, flaws in MCQs degrade NLP evaluation, but education research offers a path forward. We release BenchMarker to bridge the fields and improve MCQA benchmark design.

Ayush Rajesh Jhaveri, Anthony GX-Chen, Ilia Sucholutsky et al.

Confirmation bias, the tendency to seek evidence that supports rather than challenges one's belief, hinders one's reasoning ability. We examine whether large language models (LLMs) exhibit confirmation bias by adapting the rule-discovery study from human psychology: given a sequence of three numbers (a "triple"), an agent engages in an interactive feedback loop where it (1) proposes a new triple, (2) receives feedback on whether it satisfies the hidden rule, and (3) guesses the rule. Across eleven LLMs of multiple families and scales, we find that LLMs exhibit confirmation bias, often proposing triples to confirm their hypothesis rather than trying to falsify it. This leads to slower and less frequent discovery of the hidden rule. We further explore intervention strategies (e.g., encouraging the agent to consider counter examples) developed for humans. We find prompting LLMs with such instruction consistently decreases confirmation bias in LLMs, improving rule discovery rates from 42% to 56% on average. Lastly, we mitigate confirmation bias by distilling intervention-induced behavior into LLMs, showing promising generalization to a new task, the Blicket test. Our work shows that confirmation bias is a limitation of LLMs in hypothesis exploration, and that it can be mitigated via injecting interventions designed for humans.

Zhenheng Tang, Xiang Liu, Qian Wang et al.

As Large Language Models (LLMs) become more powerful and autonomous, they increasingly face conflicts and dilemmas in many scenarios. We first summarize and taxonomize these diverse conflicts. Then, we model the LLM's preferences to make different choices as a priority graph, where instructions and values are nodes, and the edges represent context-specific priorities determined by the model's output distribution. This graph reveals that a unified stable LLM alignment is very challenging, because the graph is neither static nor necessarily consistent in different contexts. Besides, it also reveals a potential vulnerability: priority hacking, where adversaries can craft deceptive contexts to manipulate the graph and bypass safety alignments. To counter this, we propose a runtime verification mechanism, enabling LLMs to query external sources to ground their context and resist manipulation. While this approach enhances robustness, we also acknowledge that many ethical and value dilemmas are philosophically irreducible, posing a long-term, open challenge for the future of AI alignment.

Anuj Diwan, Zhisheng Zheng, David Harwath et al.

We introduce Paralinguistic Speech Captions (ParaSpeechCaps), a large-scale dataset that annotates speech utterances with rich style captions. While rich abstract tags (e.g. guttural, nasal, pained) have been explored in small-scale human-annotated datasets, existing large-scale datasets only cover basic tags (e.g. low-pitched, slow, loud). We combine off-the-shelf text and speech embedders, classifiers and an audio language model to automatically scale rich tag annotations for the first time. ParaSpeechCaps covers a total of 59 style tags, including both speaker-level intrinsic tags and utterance-level situational tags. It consists of 342 hours of human-labelled data (PSC-Base) and 2427 hours of automatically annotated data (PSC-Scaled). We finetune Parler-TTS, an open-source style-prompted TTS model, on ParaSpeechCaps, and achieve improved style consistency (+7.9% Consistency MOS) and speech quality (+15.5% Naturalness MOS) over the best performing baseline that combines existing rich style tag datasets. We ablate several of our dataset design choices to lay the foundation for future work in this space. Our dataset, models and code are released at https://github.com/ajd12342/paraspeechcaps .

Nishant Balepur, Malachi Hamada, Varsha Kishore et al.

Scientific Deep Research (DR) agents answer user queries by synthesizing research papers into multi-section reports. User feedback can improve their utility, but existing protocols only score the final report, making it hard to study and learn which intermediate actions DR agents should take to improve reports. We collect DRACULA, the first dataset with user feedback on intermediate actions for DR. Over five weeks, nineteen expert CS researchers ask queries to a DR system that proposes actions (e.g., "Add a section on datasets"). Our users select actions they prefer, then judge whether an output report applied their selections successfully, yielding 8,103 action preferences and 5,230 execution judgments. After confirming a DR agent can execute DRACULA's actions, we study the predictability of user-preferred actions via simulation-how well LLMs predict the actions users select-a step toward learning to generate useful actions. We discover: (1) LLM judges initially struggle to predict action selections, but improve most when using a user's full selection history, rather than self-reported or extrapolated user context signals; (2) Users' selections for the same query differ based on unstated goals, bottlenecking simulation and motivating affordances that let users steer reports; and (3) Our simulation results inform an online intervention that generates new actions based on the user's past interactions, which users pick most often in follow-up studies. Overall, while work extensively studies execution, DRACULA reveals a key challenge is deciding which actions to execute in the first place. We open-source DRACULA's study design, user feedback, and simulation tasks to spur future work on action feedback for long-horizon agents.

Fangyuan Xu, Rujun Han, Yanfei Chen et al.

Deep search agents, which aim to answer complex questions requiring reasoning across multiple documents, can significantly speed up the information-seeking process. Collecting human annotations for this application is prohibitively expensive due to long and complex exploration trajectories. We propose an agentic pipeline that automatically generates high quality, difficulty-controlled deep search question-answer pairs for a given corpus and a target difficulty level. Our pipeline, SAGE, consists of a data generator which proposes QA pairs and a search agent which attempts to solve the generated question and provide execution feedback for the data generator. The two components interact over multiple rounds to iteratively refine the question-answer pairs until they satisfy the target difficulty level. Our intrinsic evaluation shows SAGE generates questions that require diverse reasoning strategies, while significantly increases the correctness and difficulty of the generated data. Our extrinsic evaluation demonstrates up to 23% relative performance gain on popular deep search benchmarks by training deep search agents with our synthetic data. Additional experiments show that agents trained on our data can adapt from fixed-corpus retrieval to Google Search at inference time, without further training.

Hung-Ting Chen, Eunsol Choi

We study retrieving a set of documents that covers various perspectives on a complex and contentious question (e.g., will ChatGPT do more harm than good?). We curate a Benchmark for Retrieval Diversity for Subjective questions (BERDS), where each example consists of a question and diverse perspectives associated with the question, sourced from survey questions and debate websites. On this data, retrievers paired with a corpus are evaluated to surface a document set that contains diverse perspectives. Our framing diverges from most retrieval tasks in that document relevancy cannot be decided by simple string matches to references. Instead, we build a language model-based automatic evaluator that decides whether each retrieved document contains a perspective. This allows us to evaluate the performance of three different types of corpus (Wikipedia, web snapshot, and corpus constructed on the fly with retrieved pages from the search engine) paired with retrievers. Retrieving diverse documents remains challenging, with the outputs from existing retrievers covering all perspectives on only 40% of the examples. We further study the effectiveness of query expansion and diversity-focused reranking approaches and analyze retriever sycophancy.

Hung-Ting Chen, Xiang Liu, Shauli Ravfogel et al.

Most text retrievers generate \emph{one} query vector to retrieve relevant documents. Yet, the conditional distribution of relevant documents for the query may be multimodal, e.g., representing different interpretations of the query. We first quantify the limitations of existing retrievers. All retrievers we evaluate struggle more as the distance between target document embeddings grows. To address this limitation, we develop a new retriever architecture, \emph{A}utoregressive \emph{M}ulti-\emph{E}mbedding \emph{R}etriever (AMER). Our model autoregressively generates multiple query vectors, and all the predicted query vectors are used to retrieve documents from the corpus. We show that on the synthetic vectorized data, the proposed method could capture multiple target distributions perfectly, showing 4x better performance than single embedding model. We also fine-tune our model on real-world multi-answer retrieval datasets and evaluate in-domain. AMER presents 4 and 21\% relative gains over single-embedding baselines on two datasets we evaluate on. Furthermore, we consistently observe larger gains on the subset of dataset where the embeddings of the target documents are less similar to each other. We demonstrate the potential of using a multi-query vector retriever and open up a new direction for future work.

Shaswat Patel, Vishvesh Trivedi, Yue Han et al.

Recent work has identified a subset of attention heads in Transformer as retrieval heads, which are responsible for retrieving information from the context. In this work, we first investigate retrieval heads in multilingual contexts. In multilingual language models, we find that retrieval heads are often shared across multiple languages. Expanding the study to cross-lingual setting, we identify Retrieval-Transition heads(RTH), which govern the transition to specific target-language output. Our experiments reveal that RTHs are distinct from retrieval heads and more vital for Chain-of-Thought reasoning in multilingual LLMs. Across four multilingual benchmarks (MMLU-ProX, MGSM, MLQA, and XQuaD) and two model families (Qwen-2.5 and Llama-3.1), we demonstrate that masking RTH induces bigger performance drop than masking Retrieval Heads (RH). Our work advances understanding of multilingual LMs by isolating the attention heads responsible for mapping to target languages.

Thom Lake, Eunsol Choi, Greg Durrett

The alignment process changes several properties of a large language model's (LLM's) output distribution. We analyze two aspects of post-alignment distributional shift of LLM responses. First, we re-examine previously reported reductions in response diversity post-alignment. Our analysis suggests that an apparent drop in the diversity of responses is largely explained by quality control and information aggregation. Alignment suppresses irrelevant and unhelpful content while shifting the output distribution toward longer responses that cover information spanning several responses from the base LLM, essentially presenting diverse information in a single response. Finding little evidence that alignment suppresses useful information, it is natural to ask the opposite question: do aligned models surface information that cannot be recovered from base models? Our second investigation shows this is not the case and the behavior of aligned models is recoverable from base models without fine-tuning. A combination of in-context examples and lower-resolution semantic hints about response content can elicit responses from base LLMs that are as similar to alignment-tuned LLM responses as alignment-tuned LLM responses are to each other. Taken together, these results indicate that current alignment techniques capture but do not extend the useful subset of assistant-like base LLM behavior, providing further evidence for the Superficial Alignment Hypothesis. They also show that in-context alignment can go surprisingly far as a strategy for imitating aligned LLMs without fine-tuning. Our code and data is available at https://github.com/thomlake/investigating-alignment.

Michael J. Q. Zhang, Eunsol Choi

While large language models are able to retain vast amounts of world knowledge seen during pretraining, such knowledge is prone to going out of date and is nontrivial to update. Furthermore, these models are often used under temporal misalignment, tasked with answering questions about the present, despite having only been trained on data collected in the past. To mitigate the effects of temporal misalignment, we propose fact duration prediction: the task of predicting how long a given fact will remain true. In our experiments, we demonstrate that identifying which facts are prone to rapid change can help models avoid reciting outdated information and determine which predictions require seeking out up-to-date knowledge sources. We also show how modeling fact duration improves calibration for knowledge-intensive tasks, such as open-retrieval question answering, under temporal misalignment, by discarding volatile facts. Our data and code are released publicly at https://github.com/mikejqzhang/mitigating_misalignment.

Mandar Joshi, Eunsol Choi, Daniel S. Weld et al.

We present TriviaQA, a challenging reading comprehension dataset containing over 650K question-answer-evidence triples. TriviaQA includes 95K question-answer pairs authored by trivia enthusiasts and independently gathered evidence documents, six per question on average, that provide high quality distant supervision for answering the questions. We show that, in comparison to other recently introduced large-scale datasets, TriviaQA (1) has relatively complex, compositional questions, (2) has considerable syntactic and lexical variability between questions and corresponding answer-evidence sentences, and (3) requires more cross sentence reasoning to find answers. We also present two baseline algorithms: a feature-based classifier and a state-of-the-art neural network, that performs well on SQuAD reading comprehension. Neither approach comes close to human performance (23% and 40% vs. 80%), suggesting that TriviaQA is a challenging testbed that is worth significant future study. Data and code available at -- http://nlp.cs.washington.edu/triviaqa/

Zhisheng Zheng, Puyuan Peng, Ziyang Ma et al.

Spatial sound reasoning is a fundamental human skill, enabling us to navigate and interpret our surroundings based on sound. In this paper we present BAT, which combines the spatial sound perception ability of a binaural acoustic scene analysis model with the natural language reasoning capabilities of a large language model (LLM) to replicate this innate ability. To address the lack of existing datasets of in-the-wild spatial sounds, we synthesized a binaural audio dataset using AudioSet and SoundSpaces 2.0. Next, we developed SpatialSoundQA, a spatial sound-based question-answering dataset, offering a range of QA tasks that train BAT in various aspects of spatial sound perception and reasoning. The acoustic front end encoder of BAT is a novel spatial audio encoder named Spatial Audio Spectrogram Transformer, or Spatial-AST, which by itself achieves strong performance across sound event detection, spatial localization, and distance estimation. By integrating Spatial-AST with LLaMA-2 7B model, BAT transcends standard Sound Event Localization and Detection (SELD) tasks, enabling the model to reason about the relationships between the sounds in its environment. Our experiments demonstrate BAT's superior performance on both spatial sound perception and reasoning, showcasing the immense potential of LLMs in navigating and interpreting complex spatial audio environments.

Michael J. Q. Zhang, W. Bradley Knox, Eunsol Choi

Large language models (LLMs) must often respond to highly ambiguous user requests. In such cases, the LLM's best response may be to ask a clarifying question to elicit more information. Existing LLMs often respond by presupposing a single interpretation of such ambiguous requests, frustrating users who intended a different interpretation. We speculate this is caused by current preference data labeling practice, where LLM responses are evaluated only on their prior contexts. To address this, we assign preference labels by simulating their expected outcomes in future turns. This allows LLMs to learn to ask clarifying questions when it can generate responses that are tailored to each user interpretation in future turns. On open-domain QA datasets with multiple annotations, we evaluate systems based on their ability to ask clarifying questions to recover each user's interpretation and expected answer. We compare systems trained using our proposed preference labeling methods against standard methods, which assign preferences based on only prior context. Our method achieves a 5% improvement in F1 measured against the answer set from different interpretations of each query, showing the value of modeling future conversation turns. We further demonstrate that our method can be used to train models to judiciously determine when to ask clarifying questions, directly answering the question when clarification is unnecessary. In our experiments, we find that our method achieves a 3% improvement in accuracy of such judgments over existing methods.

Michael JQ Zhang, Zhilin Wang, Jena D. Hwang et al. · allen-ai

We examine diverging preferences in human-labeled preference datasets. We develop a taxonomy of disagreement sources spanning 10 categories across four high-level classes -- task underspecification, response style, refusals, and annotation errors. We find that the majority of disagreements are in opposition with standard reward modeling approaches, which are designed with the assumption that annotator disagreement is noise. We then explore how these findings impact two areas of LLM development: reward modeling and evaluation. In our experiments, we demonstrate how standard reward modeling methods, like the Bradley-Terry model, fail to differentiate whether a given preference judgment is the result of unanimous agreement among annotators or the majority opinion among diverging user preferences. We also find that these tendencies are also echoed by popular LLM-as-Judge evaluation methods, which consistently identify a winning response in cases of diverging preferences. These findings highlight remaining challenges in LLM evaluations, which are greatly influenced by divisive features like response style, and in developing pluralistically aligned LLMs. To address these issues, we develop methods for identifying diverging preferences to mitigate their influence on evaluation and training.

Yoonsang Lee, Xi Ye, Eunsol Choi · princeton

Different entities with the same name can be difficult to distinguish. Handling confusing entity mentions is a crucial skill for language models (LMs). For example, given the question "Where was Michael Jordan educated?" and a set of documents discussing different people named Michael Jordan, can LMs distinguish entity mentions to generate a cohesive answer to the question? To test this ability, we introduce a new benchmark, AmbigDocs. By leveraging Wikipedia's disambiguation pages, we identify a set of documents, belonging to different entities who share an ambiguous name. From these documents, we generate questions containing an ambiguous name and their corresponding sets of answers. Our analysis reveals that current state-of-the-art models often yield ambiguous answers or incorrectly merge information belonging to different entities. We establish an ontology categorizing four types of incomplete answers and automatic evaluation metrics to identify such categories. We lay the foundation for future work on reasoning across multiple documents with ambiguous entities.

Fangyuan Xu, Kyle Lo, Luca Soldaini et al. · allen-ai

Large language models (LLMs) adapted to follow user instructions are now widely deployed as conversational agents. In this work, we examine one increasingly common instruction-following task: providing writing assistance to compose a long-form answer. To evaluate the capabilities of current LLMs on this task, we construct KIWI, a dataset of knowledge-intensive writing instructions in the scientific domain. Given a research question, an initial model-generated answer and a set of relevant papers, an expert annotator iteratively issues instructions for the model to revise and improve its answer. We collect 1,260 interaction turns from 234 interaction sessions with three state-of-the-art LLMs. Each turn includes a user instruction, a model response, and a human evaluation of the model response. Through a detailed analysis of the collected responses, we find that all models struggle to incorporate new information into an existing answer, and to perform precise and unambiguous edits. Further, we find that models struggle to judge whether their outputs successfully followed user instructions, with accuracy at least 10 points short of human agreement. Our findings indicate that KIWI will be a valuable resource to measure progress and improve LLMs' instruction-following capabilities for knowledge intensive writing tasks.

Jianing Qi, Xi Ye, Hao Tang et al.

Scaling test-time compute brings substantial performance gains for large language models (LLMs). By sampling multiple answers and heuristically aggregate their answers (e.g., either through majority voting or using verifiers to rank the answers), one can achieve consistent performance gains in math domains. In this paper, we propose a new way to leverage such multiple sample set. We train a compact LLM, called Sample Set Aggregator (SSA), that takes a concatenated sequence of multiple samples and output the final answer, optimizing it for the answer accuracy with reinforcement learning. Experiments on five reasoning datasets demonstrate both the efficacy and efficiency of SSA. Notably, SSA improves over naive majority voting by 8% pass@5 on MATH. Furthermore, our 3B SSA surpasses model-based re-ranking with a much larger 72B process reward model. Our analysis also shows promising generalization ability of SSA, across sample set sizes, base model families and scales, and tasks. By separating LLMs to generate answers and LLMs to analyze and aggregate sampled answers, our approach can work with the outputs from premier black box models easily and efficiently.

James Allan, Eunsol Choi, Daniel P. Lopresti et al.

In the fast-evolving field of information retrieval (IR), the integration of generative AI technologies such as large language models (LLMs) is transforming how users search for and interact with information. Recognizing this paradigm shift at the intersection of IR and generative AI (IR-GenAI), a visioning workshop supported by the Computing Community Consortium (CCC) was held in July 2024 to discuss the future of IR in the age of generative AI. This workshop convened 44 experts in information retrieval, natural language processing, human-computer interaction, and artificial intelligence from academia, industry, and government to explore how generative AI can enhance IR and vice versa, and to identify the major challenges and opportunities in this rapidly advancing field. This report contains a summary of discussions as potentially important research topics and contains a list of recommendations for academics, industry practitioners, institutions, evaluation campaigns, and funding agencies.

Yuhan Liu, Michael J. Q. Zhang, Eunsol Choi

Once language models (LMs) are deployed, they can interact with users long-term, ideally evolving based on their feedback. Asking for direct user feedback can be disruptive; thus, we study harvesting implicit user feedback from user-LM interaction logs. We study two user-LM interaction datasets (WildChat and LMSYS). First, we analyze user feedback in the user-LLM conversation logs, providing insights into when and why such feedback occurs. Second, we study harvesting learning signals from such implicit user feedback. Specifically, we study whether incorporating the contents of user feedback (e.g., user wanted clarification), in addition to the polarity of the feedback, can improve the model performance. We observe mixed results, showing this helps in short human-designed questions (MTBench) but not on longer and more complex questions (WildBench). Together, we provide an in-depth study of implicit user feedback, showing its potential and limitations.

Fangyuan Xu, Tanya Goyal, Eunsol Choi

Generating long sequences of tokens given a long-context input is a very compute-intensive inference scenario for large language models (LLMs). One prominent inference speed-up approach is to construct a smaller key-value (KV) cache, relieving LLMs from computing attention over a long sequence of tokens. While such methods work well to generate short sequences, their performance degrades rapidly for long-form generation. Most KV compression happens once, prematurely removing tokens that can be useful later in the generation. We propose a new inference method, RefreshKV, that flexibly alternates between full context attention and attention over a subset of input tokens during generation. After each full attention step, we update the smaller KV cache based on the attention pattern over the entire input. Applying our method to off-the-shelf LLMs achieves comparable speedup to eviction-based methods while improving performance for various long-form generation tasks. Lastly, we show that continued pretraining with our inference setting brings further gains in performance.

Yuhan Liu, Fangyuan Xu, Vishakh Padmakumar et al.

When posed with prompts that permit a large number of valid answers, comprehensively generating them is the first step towards satisfying a wide range of users. In this paper, we study methods to elicit a comprehensive set of valid responses. To evaluate this, we introduce \textbf{diversity coverage}, a metric that measures the total quality scores assigned to each \textbf{unique} answer in the predicted answer set relative to the best possible answer set with the same number of answers. Using this metric, we evaluate 18 LLMs, finding no single model dominates at generating diverse responses to a wide range of open-ended prompts. Yet, per each prompt, there exists a model that outperforms all other models significantly at generating a diverse answer set. Motivated by this finding, we introduce a router that predicts the best model for each query. On NB-Wildchat, our trained router outperforms the single best model baseline (26.3% vs $23.8%). We further show generalization to an out-of-domain dataset (NB-Curated) as well as different answer-generation prompting strategies. Our work lays foundation for studying generating comprehensive answers when we have access to a suite of models.

Atula Tejaswi, Yoonsang Lee, Sujay Sanghavi et al. · princeton

We investigate whether in-context examples, widely used in decoder-only language models (LLMs), can improve embedding model performance in retrieval tasks. Unlike in LLMs, naively prepending in-context examples (query-document pairs) to the target query at inference time does not work out of the box. We introduce a simple approach to enable retrievers to use in-context examples. Our approach, RARe, finetunes a pre-trained model with in-context examples whose query is semantically similar to the target query. This can be applied to adapt various base architectures (i.e., decoder-only language models, retriever models) and consistently achieves performance gains of up to +2.72% nDCG across various open-domain retrieval datasets (BeIR, RAR-b). In particular, we find RARe exhibits stronger out-of-domain generalization compared to models using queries without in-context examples, similar to what is seen for in-context learning in LLMs. We further provide analysis on the design choices of in-context example augmentation and lay the foundation for future work in this space.

Deniz Qian, Hung-Ting Chen, Eunsol Choi

Comprehensively retrieving diverse documents is crucial to address queries that admit a wide range of valid answers. We introduce retrieve-verify-retrieve (RVR), a multi-round retrieval framework designed to maximize answer coverage. Initially, a retriever takes the original query and returns a candidate document set, followed by a verifier that identifies a high-quality subset. For subsequent rounds, the query is augmented with previously verified documents to uncover answers that are not yet covered in previous rounds. RVR is effective even with off-the-shelf retrievers, and fine-tuning retrievers for our inference procedure brings further gains. Our method outperforms baselines, including agentic search approaches, achieving at least 10% relative and 3% absolute gain in complete recall percentage on a multi-answer retrieval dataset (QAMPARI). We also see consistent gains on two out-of-domain datasets (QUEST and WebQuestionsSP) across different base retrievers. Our work presents a promising iterative approach for comprehensive answer recall leveraging a verifier and adapting retrievers to a new inference scenario.

Xiang Liu, Xuming Hu, Xiaowen Chu et al.

Recent reasoning Large Language Models (LLMs) demonstrate remarkable problem-solving abilities but often generate long thinking traces whose utility is unclear. Our work aims to improve their efficiency, enabling them to reach high performance without overthinking. First, we analyze the entropy of token probabilities in reasoning traces. Across three models, we observe a consistent U-shaped entropy pattern: high entropy on easy problems despite high accuracy, low entropy on problems with medium difficulty, and high entropy on hard problems reflecting uncertainty. Specifically, we notice 22--25\% entropy reduction from easy to medium difficulty regions, suggesting an {overthinking} phenomenon on easy instances. Building on these insights, we introduce \textbf{DiffAdapt}, a lightweight framework that selects Easy/Normal/Hard inference strategies per question based on their difficulty and reasoning trace entropy. Each inference strategy consists of a fixed prompt, temperature and maximum token length. In contrast to existing efficiency optimization methods, our approach does not fine-tune base LLM but a small probe that classifies LLM's final hidden state, allowing inexpensive adaptation. We comprehensively evaluate our method on five models and eight benchmarks. Our method achieves comparable or improved accuracy while reducing token usage by up to 22.4\%, establishing a practical path toward compute-efficient reasoning.