Question answering over visually rich documents (VRDs) requires reasoning not only over isolated content but also over documents’ structural organization and cross-page dependencies. However, conventional retrieval-augmented generation (RAG) methods encode content in isolated chunks during ingestion, losing structural and cross-page dependencies, and retrieve a fixed number of pages at inference, regardless of the specific demands of the question or context. This often results in incomplete evidence retrieval and degraded answer quality for multi-page reasoning tasks. To address these limitations, we propose LAD-RAG, a novel Layout-Aware Dynamic RAG framework. During ingestion, LAD-RAG constructs a symbolic document graph that captures layout structure and cross-page dependencies, adding it alongside standard neural embeddings to yield a more holistic representation of the document. During inference, an LLM agent dynamically interacts with the neural and symbolic indices to adaptively retrieve the necessary evidence based on the query. Experiments on MMLongBench-Doc, LongDocURL, DUDE, and MP-DocVQA demonstrate that LAD-RAG improves retrieval, achieving over 90% perfect recall on average without any top-k tuning, and outperforming baseline retrievers by up to 20% in recall at comparable noise levels, yielding higher QA accuracy with minimal latency.

Large Language Models (LLMs) often produce inconsistent answers when faced with different phrasings of the same prompt. In this paper, we propose Flip-Flop Consistency (F²C), an unsupervised training method that improves robustness to such perturbations. F²C is composed of two key components. The first, Consensus Cross-Entropy (CCE), uses a majority vote across prompt variations to create a hard pseudo-label. The second is a representation alignment loss that pulls lower-confidence and non-majority predictors toward the consensus established by high-confidence, majority-voting variations. We evaluate our method on 11 datasets spanning four NLP tasks, with 4–15 prompt variations per dataset. On average, F²C raises observed agreement by 11.62%, improves mean F₁ by 8.94%, and reduces performance variance across formats by 3.29%. In out-of-domain evaluations, F²C generalizes effectively, increasing  ̅F₁ and agreement while decreasing variance across most source-target pairs. Finally, when trained on only a subset of prompt perturbations and evaluated on held-out formats, F²C consistently improves both performance and agreement while reducing variance. These findings highlight F²C as an effective unsupervised method for enhancing LLM consistency, performance, and generalization under prompt perturbations.

Traffic stops are among the most frequent police–civilian interactions, and body-worn cameras (BWCs) provide a unique record of how these encounters unfold. Respect is a central dimension of these interactions, shaping public trust and perceived legitimacy, yet its interpretation is inherently subjective and shaped by lived experience, rendering community-specific perspectives a critical consideration. Leveraging unprecedented access to Los Angeles Police Department BWC footage, we introduce the first large-scale traffic-stop dataset annotated with respect ratings and free-text rationales from multiple perspectives. By sampling annotators from police-affiliated, justice-system-impacted, and non-affiliated Los Angeles residents, we enable the systematic study of perceptual differences across diverse communities. To this end, (i) we develop a domain-specific evaluation rubric grounded in procedural justice theory, LAPD training materials, and extensive fieldwork; (ii) we introduce a criterion-driven preference data construction framework for perspective-consistent alignment, and (ii) we propose a perspective-aware modeling framework that predicts personalized respect ratings and generates annotator-specific rationales for both officers and civilian drivers from traffic-stop transcripts. Across all three annotator groups, our approach improves both rating prediction performance and rationale alignment. Our perspective-aware framework enables law enforcement to better understand diverse community expectations, providing a vital tool for building public trust and procedural legitimacy.

Multilingual Large Language Models (LLMs) develop cross-lingual abilities despite being trained on limited parallel data. However, they often struggle to generate responses in the intended language, favoring high-resource languages such as English. In this work, we introduce CoCo-CoLa (Correct Concept - Correct Language), a novel metric to evaluate language adherence in multilingual LLMs. Using fine-tuning experiments on a closed-book QA task across seven languages, we analyze how training in one language affects others’ performance. Our findings reveal that multilingual models share task knowledge across languages but exhibit biases in the selection of output language. We identify language-specific layers, showing that final layers play a crucial role in determining output language. Accordingly, we propose a partial training strategy that selectively fine-tunes key layers, improving language adherence while reducing computational cost. Our method achieves comparable or superior performance to full fine-tuning, particularly for low-resource languages, offering a more efficient multilingual adaptation.

In subjective NLP tasks, where a single ground truth does not exist, the inclusion of diverse annotators becomes crucial as their unique perspectives significantly influence the annotations. In realistic scenarios, the annotation budget often becomes the main determinant of the number of perspectives (i.e., annotators) included in the data and subsequent modeling. We introduce a novel framework for annotation collection and modeling in subjective tasks that aims to minimize the annotation budget while maximizing the predictive performance for each annotator. Our framework has a two-stage design: first, we rely on a small set of annotators to build a multitask model, and second, we augment the model for a new perspective by strategically annotating a few samples per annotator. To test our framework at scale, we introduce and release a unique dataset, Moral Foundations Subjective Corpus, of 2000 Reddit posts annotated by 24 annotators for moral sentiment. We demonstrate that our framework surpasses the previous SOTA in capturing the annotators’ individual perspectives with as little as 25% of the original annotation budget on two datasets. Furthermore, our framework results in more equitable models, reducing the performance disparity among annotators.

Detecting problematic content, such as hate speech, is a multifaceted and ever-changing task, influenced by social dynamics, user populations, diversity of sources, and evolving language. There has been significant efforts, both in academia and in industry, to develop annotated resources that capture various aspects of problematic content. Due to researchers’ diverse objectives, these annotations are often inconsistent and hence, reports of progress on the detection of problematic content are fragmented. This pattern is expected to persist unless we pool these resources, taking into account the dynamic nature of this issue. In this paper, we propose integrating the available resources, leveraging their dynamic nature to break this pattern, and introduce a continual learning framework and benchmark for problematic content detection. Our benchmark, comprising 84 related tasks, creates a novel measure of progress: prioritizing the adaptability of classifiers to evolving tasks over excelling in specific tasks. To ensure continuous relevance, our benchmark is designed for seamless integration of new tasks. Our results demonstrate that continual learning methods outperform static approaches by up to 17% and 4% AUC in capturing the evolving content and adapting to novel forms of problematic content

Language usage is related to speaker age, gender, moral concerns, political ideology, and other attributes. Current state-of-the-art methods for predicting these attributes take a speaker’s utterances as input and provide a prediction per speaker attribute. Most of these approaches struggle to handle a large number of utterances per speaker. This difficulty is primarily due to the computational constraints of the models. Additionally, only a subset of speaker utterances may be relevant to specific attributes. In this paper, we formulate speaker attribute prediction as a Multiple Instance Learning (MIL) problem and propose RL-MIL, a novel approach based on Reinforcement Learning (RL) that effectively addresses both of these challenges. Our experiments demonstrate that our RL-based methodology consistently outperforms previous approaches across a range of related tasks: predicting speakers’ psychographics and demographics from social media posts, and political ideologies from transcribed speeches. We create synthetic datasets and investigate the behavior of RL-MIL systematically. Our results show the success of RL-MIL in improving speaker attribute prediction by learning to select relevant speaker utterances.

Social stereotypes negatively impact individuals’ judgments about different groups and may have a critical role in understanding language directed toward marginalized groups. Here, we assess the role of social stereotypes in the automated detection of hate speech in the English language by examining the impact of social stereotypes on annotation behaviors, annotated datasets, and hate speech classifiers. Specifically, we first investigate the impact of novice annotators’ stereotypes on their hate-speech-annotation behavior. Then, we examine the effect of normative stereotypes in language on the aggregated annotators’ judgments in a large annotated corpus. Finally, we demonstrate how normative stereotypes embedded in language resources are associated with systematic prediction errors in a hate-speech classifier. The results demonstrate that hate-speech classifiers reflect social stereotypes against marginalized groups, which can perpetuate social inequalities when propagated at scale. This framework, combining social-psychological and computational-linguistic methods, provides insights into sources of bias in hate-speech moderation, informing ongoing debates regarding machine learning fairness.

Existing bias mitigation methods require social-group-specific word pairs (e.g., “man” – “woman”) for each social attribute (e.g., gender), restricting the bias mitigation to only one specified social attribute. Further, this constraint renders such methods impractical and costly for mitigating bias in understudied and/or unmarked social groups. We propose that the Stereotype Content Model (SCM) — a theoretical framework developed in social psychology for understanding the content of stereotyping — can help debiasing efforts to become social-group-agnostic by capturing the underlying connection between bias and stereotypes. SCM proposes that the content of stereotypes map to two psychological dimensions of warmth and competence. Using only pairs of terms for these two dimensions (e.g., warmth: “genuine” – “fake”; competence: “smart” – “stupid”), we perform debiasing with established methods on both pre-trained word embeddings and large language models. We demonstrate that our social-group-agnostic, SCM-based debiasing technique performs comparably to group-specific debiasing on multiple bias benchmarks, but has theoretical and practical advantages over existing approaches.

Bias mitigation approaches reduce models’ dependence on sensitive features of data, such as social group tokens (SGTs), resulting in equal predictions across the sensitive features. In hate speech detection, however, equalizing model predictions may ignore important differences among targeted social groups, as hate speech can contain stereotypical language specific to each SGT. Here, to take the specific language about each SGT into account, we rely on counterfactual fairness and equalize predictions among counterfactuals, generated by changing the SGTs. Our method evaluates the similarity in sentence likelihoods (via pre-trained language models) among counterfactuals, to treat SGTs equally only within interchangeable contexts. By applying logit pairing to equalize outcomes on the restricted set of counterfactuals for each instance, we improve fairness metrics while preserving model performance on hate speech detection.

Hate speech classifiers trained on imbalanced datasets struggle to determine if group identifiers like “gay” or “black” are used in offensive or prejudiced ways. Such biases manifest in false positives when these identifiers are present, due to models’ inability to learn the contexts which constitute a hateful usage of identifiers. We extract post-hoc explanations from fine-tuned BERT classifiers to detect bias towards identity terms. Then, we propose a novel regularization technique based on these explanations that encourages models to learn from the context of group identifiers in addition to the identifiers themselves. Our approach improved over baselines in limiting false positives on out-of-domain data while maintaining and in cases improving in-domain performance.

This paper demonstrates a state-of-the-art end-to-end multilingual (English, Russian, and Ukrainian) knowledge extraction system that can perform entity discovery and linking, relation extraction, event extraction, and coreference. It extracts and aggregates knowledge elements across multiple languages and documents as well as provides visualizations of the results along three dimensions: temporal (as displayed in an event timeline), spatial (as displayed in an event heatmap), and relational (as displayed in entity-relation networks). For our system to further support users’ analyses of causal sequences of events in complex situations, we also integrate a wide range of human moral value measures, independently derived from region-based survey, into the event heatmap. This system is publicly available as a docker container and a live demo.

Cognitive tests have traditionally resorted to standardizing testing materials in the name of equality and because of the onerous nature of creating test items. This approach ignores participants’ diverse language experiences that potentially significantly affect testing outcomes. Here, we seek to explain our prior finding of significant performance differences on two cognitive tests (reading span and SPiN) between clusters of participants based on their media consumption. Here, we model the language contained in these media sources using an LSTM trained on corpora of each cluster’s media sources to predict target words. We also model semantic similarity of test items with each cluster’s corpus using skip-thought vectors. We find robust, significant correlations between performance on the SPiN test and the LSTMs and skip-thought models we present here, but not the reading span test.

Official reports of hate crimes in the US are under-reported relative to the actual number of such incidents. Further, despite statistical approximations, there are no official reports from a large number of US cities regarding incidents of hate. Here, we first demonstrate that event extraction and multi-instance learning, applied to a corpus of local news articles, can be used to predict instances of hate crime. We then use the trained model to detect incidents of hate in cities for which the FBI lacks statistics. Lastly, we train models on predicting homicide and kidnapping, compare the predictions to FBI reports, and establish that incidents of hate are indeed under-reported, compared to other types of crimes, in local press.