Ameeta Agrawal

Francielle Vargas, João Robiatti, Diego Alves et al.

Ensuring factuality and interpretability in RAG remains an open and urgent problem. We introduce Contrastive Evidence Rationale Attention (CERA), the first retrieval framework to employ subjectivity-based hard negative selection and inject an evidential inductive bias into contrastive learning through an auxiliary attention alignment loss. CERA fine-tunes a dense retriever using two training objectives: triplet-based contrastive learning and interpretable attention alignment, which supervises CLS-to-token attention using a part-of-speech-weighted masking distribution over human-annotated factual rationales as evidence signals. Experiments on a large corpus of clinical trial reports demonstrate that the subjectivity-based hard negative selection substantially improves retrieval effectiveness compared to both Contriever and hard negative selection baselines. Furthermore, rationale alignment improves faithfulness while maintaining competitive retrieval performance, supporting the hypothesis that attention can serve as a more faithful explanation of model behavior when guided by human rationales. Moving beyond topical similarity, CERA enables the retriever to identify the specific tokens that constitute supporting evidence, promoting more interpretable evidence selection in RAG systems.

Yufei Tao, Adam Hiatt, Erik Haake et al.

Large language models (LLMs) have demonstrated remarkable progress in leveraging diverse knowledge sources. This study investigates how nine widely used LLMs allocate knowledge between local context and global parameters when answering open-ended questions in knowledge-consistent scenarios. We introduce a novel dataset, WikiAtomic, and systematically vary context sizes to analyze how LLMs prioritize and utilize the provided information and their parametric knowledge in knowledge-consistent scenarios. Additionally, we also study their tendency to hallucinate under varying context sizes. Our findings reveal consistent patterns across models, including a consistent reliance on both contextual (around 70%) and parametric (around 30%) knowledge, and a decrease in hallucinations with increasing context. These insights highlight the importance of more effective context organization and developing models that use input more deterministically for robust performance.

Philippe J. Giabbanelli, Jose J. Padilla, Ameeta Agrawal

Large Language Models (LLMs) are becoming ubiquitous to create intelligent virtual assistants that assist users in interacting with a system, as exemplified in marketing. Although LLMs have been discussed in Modeling & Simulation (M&S), the community has focused on generating code or explaining results. We examine the possibility of using LLMs to broaden access to simulations, by enabling non-simulation end-users to ask what-if questions in everyday language. Specifically, we discuss the opportunities and challenges in designing such an end-to-end system, divided into three broad phases. First, assuming the general case in which several simulation models are available, textual queries are mapped to the most relevant model. Second, if a mapping cannot be found, the query can be automatically reformulated and clarifying questions can be generated. Finally, simulation results are produced and contextualized for decision-making. Our vision for such system articulates long-term research opportunities spanning M&S, LLMs, information retrieval, and ethics.

Rhitabrat Pokharel, Ameeta Agrawal

The use of large language models (LLMs) for evaluating outputs is becoming an increasingly effective and scalable approach. However, it remains uncertain whether this capability extends beyond task-specific evaluations to more general assessments of text quality, particularly in multilingual contexts. In this study, we introduce, MTQ-Eval, a novel framework for multilingual text quality evaluation that learns from examples of both high- and low-quality texts, adjusting its internal representations. To develop MTQ-Eval, we first automatically generate text quality preference data and then use it to train open-source base LLMs to align with ratings of high- and low-quality text. Our comprehensive evaluation across 115 languages demonstrates the improved performance of the proposed model. Upon further analysis, we find that this enhanced evaluation capability also leads to notable improvements in downstream tasks.

Francielle Vargas, Jackson Trager, Diego Alves et al.

Hate speech detection models rely on surface-level lexical features, increasing vulnerability to spurious correlations and limiting robustness, cultural contextualization, and interpretability. We propose Supervised Moral Rationale Attention (SMRA), the first self-explaining hate speech detection framework to incorporate moral rationales as direct supervision for attention alignment. Based on Moral Foundations Theory, SMRA aligns token-level attention with expert-annotated moral rationales, guiding models to attend to morally salient spans rather than spurious lexical patterns. Unlike prior rationale-supervised or post-hoc approaches, SMRA integrates moral rationale supervision directly into the training objective, producing inherently interpretable and contextualized explanations. To support our framework, we also introduce HateBRMoralXplain, a Brazilian Portuguese benchmark dataset annotated with hate labels, moral categories, token-level moral rationales, and socio-political metadata. Across binary hate speech detection and multi-label moral sentiment classification, SMRA consistently improves performance (e.g., +0.9 and +1.5 F1, respectively) while substantially enhancing explanation faithfulness, increasing IoU F1 (+7.4 pp) and Token F1 (+5.0 pp). Although explanations become more concise, sufficiency improves (+2.3 pp) and fairness remains stable, indicating more faithful rationales without performance or bias trade-offs

Enas Altarawneh, Ameeta Agrawal, Michael Jenkin et al.

Online conversations are particularly susceptible to derailment, which can manifest itself in the form of toxic communication patterns including disrespectful comments and abuse. Forecasting conversation derailment predicts signs of derailment in advance enabling proactive moderation of conversations. State-of-the-art approaches to conversation derailment forecasting sequentially encode conversations and use graph neural networks to model dialogue user dynamics. However, existing graph models are not able to capture complex conversational characteristics such as context propagation and emotional shifts. The use of common sense knowledge enables a model to capture such characteristics, thus improving performance. Following this approach, here we derive commonsense statements from a knowledge base of dialogue contextual information to enrich a graph neural network classification architecture. We fuse the multi-source information on utterance into capsules, which are used by a transformer-based forecaster to predict conversation derailment. Our model captures conversation dynamics and context propagation, outperforming the state-of-the-art models on the CGA and CMV benchmark datasets

Justin J. Xie, Ameeta Agrawal

Paraphrase generation, a.k.a. paraphrasing, is a common and important task in natural language processing. Emotional paraphrasing, which changes the emotion embodied in a piece of text while preserving its meaning, has many potential applications, including moderating online dialogues and preventing cyberbullying. We introduce a new task of fine-grained emotional paraphrasing along emotion gradients, that is, altering the emotional intensities of the paraphrases in fine-grained settings following smooth variations in affective dimensions while preserving the meaning of the original text. We reconstruct several widely used paraphrasing datasets by augmenting the input and target texts with their fine-grained emotion labels. Then, we propose a framework for emotion and sentiment guided paraphrasing by leveraging pre-trained language models for conditioned text generation. Extensive evaluation of the fine-tuned models suggests that including fine-grained emotion labels in the paraphrase task significantly improves the likelihood of obtaining high-quality paraphrases that reflect the desired emotions while achieving consistently better scores in paraphrase metrics such as BLEU, ROUGE, and METEOR.

Rhitabrat Pokharel, Ameeta Agrawal

Prior work typically describes out-of-domain (OOD) or out-of-distribution (OODist) samples as those that originate from dataset(s) or source(s) different from the training set but for the same task. When compared to in-domain (ID) samples, the models have been known to usually perform poorer on OOD samples, although this observation is not consistent. Another thread of research has focused on OOD detection, albeit mostly using supervised approaches. In this work, we first consolidate and present a systematic analysis of multiple definitions of OOD and OODist as discussed in prior literature. Then, we analyze the performance of a model under ID and OOD/OODist settings in a principled way. Finally, we seek to identify an unsupervised method for reliably identifying OOD/OODist samples without using a trained model. The results of our extensive evaluation using 12 datasets from 4 different tasks suggest the promising potential of unsupervised metrics in this task.

Ameeta Agrawal, Andy Dang, Sina Bagheri Nezhad et al.

Recent large language models (LLMs) demonstrate impressive capabilities in handling long contexts, some exhibiting near-perfect recall on synthetic retrieval tasks. However, these evaluations have mainly focused on English text and involved a single target sentence within lengthy contexts. Our work investigates how LLM performance generalizes to multilingual settings with multiple hidden target sentences. We create a new dataset -- mLongRR -- to comprehensively evaluate several multilingual long-context LLMs on retrieval and reasoning tasks across five languages: English, Vietnamese, Indonesian, Swahili, and Somali. These languages share the Latin script but belong to distinct language families and resource levels. Our analysis reveals a significant performance gap between languages. The best-performing models such as Gemini-1.5 and GPT-4o, achieve around 96% accuracy in English to around 36% in Somali with a single target sentence. However, this accuracy drops to 40% in English and 0% in Somali when dealing with three target sentences. Our findings highlight the challenges long-context LLMs face when processing longer contexts, an increase in the number of target sentences, or languages of lower resource levels.

Sina Bagheri Nezhad, Sayan Bandyapadhyay, Ameeta Agrawal

Fairness in multi-document summarization of user-generated content remains a critical challenge in natural language processing (NLP). Existing summarization methods often fail to ensure equitable representation across different social groups, leading to biased outputs. In this paper, we introduce two novel methods for fair extractive summarization: FairExtract, a clustering-based approach, and FairGPT, which leverages GPT-3.5-turbo with fairness constraints. We evaluate these methods using Divsumm summarization dataset of White-aligned, Hispanic, and African-American dialect tweets and compare them against relevant baselines. The results obtained using a comprehensive set of summarization quality metrics such as SUPERT, BLANC, SummaQA, BARTScore, and UniEval, as well as a fairness metric F, demonstrate that FairExtract and FairGPT achieve superior fairness while maintaining competitive summarization quality. Additionally, we introduce composite metrics (e.g., SUPERT+F, BLANC+F) that integrate quality and fairness into a single evaluation framework, offering a more nuanced understanding of the trade-offs between these objectives. Our code is available online.

Yufei Tao, Ameeta Agrawal

Large language models (LLMs) can answer questions and summarize documents when conditioned on external contexts (e.g., retrieved evidence), yet context use remains unreliable: models may overwrite an already-correct output (neutral regression) even when the context is non-informative. We formalize neutral regression as a do-no-harm requirement and quantify it by measuring accuracy drops on baseline-correct items under answer-consistent contexts. We propose No-Worse Context-Aware Decoding (NWCAD), a decode-time adapter built on a two-stream setup with a two-stage gate: it backs off to no-context decoding when the context is non-informative, and otherwise uses context-conditioned decoding with a CAD-style fallback under uncertainty. We evaluate NWCAD on benchmarks that separate do-no-harm reliability from context utilization (accuracy gains on genuinely helpful contexts). NWCAD prevents neutral regression on baseline-correct items while preserving strong context-driven accuracy on helpful contexts.

Rhitabrat Pokharel, Ameeta Agrawal

The ability to process idiomatic or literal multiword expressions is a crucial aspect of understanding and generating any language. The task of generating contextually relevant continuations for narratives containing idiomatic (or literal) expressions can allow us to test the ability of generative language models (LMs) in understanding nuanced language containing non-compositional figurative text. We conduct a series of experiments using datasets in two distinct languages (English and Portuguese) under three different training settings (zero-shot, few-shot, and fine-tuned). Our results suggest that the models are only slightly better at generating continuations for literal contexts than idiomatic contexts, with exceedingly small margins. Furthermore, the models studied in this work perform equally well across both languages, indicating the robustness of generative models in performing this task.

Sina Bagheri Nezhad, Ameeta Agrawal

Multilingual language models have gained significant attention in recent years, enabling the development of applications that meet diverse linguistic contexts. In this paper, we present a comprehensive evaluation of three popular multilingual language models: mBERT, XLM-R, and GPT-3. We assess their performance across a diverse set of languages, with a focus on understanding the impact of resource availability (general and model-specific), language family, script type, and word order on model performance, under two distinct tasks - text classification and text generation. Our findings reveal that while the amount of language-specific pretraining data plays a crucial role in model performance, we also identify other factors such as general resource availability, language family, and script type, as important features. We hope that our study contributes to a deeper understanding of multilingual language models to enhance their performance across languages and linguistic contexts.

Olubusayo Olabisi, Ekata Mitra, Ameeta Agrawal

Summarizing deeply nested discussion threads requires handling interleaved replies, quotes, and overlapping topics, which standard LLM summarizers struggle to capture reliably. We introduce ThreadSumm, a multi-stage LLM framework that treats thread summarization as a hierarchical reasoning problem over explicit aspect and content unit representations. Our method first performs content planning via LLM-based extraction of discourse aspects and Atomic Content Units, then applies sentence ordering to construct thread-aware sequences that surface multiple viewpoints rather than a single linear strand. On top of these interpretable units, ThreadSumm employs a Tree of Thoughts search that generates and scores multiple paragraph candidates, jointly optimizing coherence and coverage within a unified search space. With this multi-proposal and iterative refinement design, we show improved performance in generating logically structured summaries compared to existing baselines, while achieving higher aspect retention and opinion coverage in nested discussions.

Rhitabrat Pokharel, Ameeta Agrawal, Tanay Nagar

Large language models exhibit strong multilingual capabilities, yet significant performance gaps persist between dominant and non-dominant languages. Prior work attributes this gap to imbalances between shared and language-specific neurons in multilingual representations. We propose Cross-Lingual Activation Steering (CLAS), a training-free inference-time intervention that selectively modulates neuron activations. We evaluate CLAS on classification and generation benchmarks, achieving average improvements of 2.3% (Acc.) and 3.4% (F1) respectively, while maintaining high-resource language performance. We discover that effective transfer operates through functional divergence rather than strict alignment; performance gains correlate with increased language cluster separation. Our results demonstrate that targeted activation steering can unlock latent multilingual capacity in existing models without modification to model weights.

Rhitabrat Pokharel, Yufei Tao, Ameeta Agrawal

Preference optimization is a critical post-training technique used to align large language models (LLMs) with human preferences, typically by fine-tuning on ranked response pairs. While methods like Direct Preference Optimization (DPO) have proven effective in English, they often fail to generalize robustly to multilingual settings. We propose a simple yet effective alternative, Confidence-Aware Preference Optimization (CAPO), which replaces DPO's fixed treatment of preference pairs with a dynamic loss scaling mechanism based on a relative reward. By modulating the learning signal according to the confidence in each preference pair, CAPO enhances robustness to noisy or low-margin comparisons, typically encountered in multilingual text. Empirically, CAPO outperforms existing preference optimization baselines by at least 16% in reward accuracy, and improves alignment by widening the gap between preferred and dispreferred responses across languages.

Sina Bagheri Nezhad, Yao Li, Ameeta Agrawal

Large Language Models (LLMs) often struggle with complex mathematical reasoning, where prose-based generation leads to unverified and arithmetically unsound solutions. Current prompting strategies like Chain of Thought still operate within this unreliable medium, lacking a mechanism for deterministic verification. To address these limitations, we introduce SymCode, a neurosymbolic framework that reframes mathematical problem-solving as a task of verifiable code generation using the SymPy library. We evaluate SymCode on challenging benchmarks, including MATH-500 and OlympiadBench, demonstrating significant accuracy improvements of up to 13.6 percentage points over baselines. Our analysis shows that SymCode is not only more token-efficient but also fundamentally shifts model failures from opaque logical fallacies towards transparent, programmatic errors. By grounding LLM reasoning in a deterministic symbolic engine, SymCode represents a key step towards more accurate and trustworthy AI in formal domains.

Sina Bagheri Nezhad, Ameeta Agrawal

This study investigates the factors influencing the performance of multilingual large language models (MLLMs) across diverse languages. We study 6 MLLMs, including masked language models, autoregressive models, and instruction-tuned LLMs, on the SIB-200 dataset, a topic classification dataset encompassing 204 languages. Our analysis considers three scenarios: ALL languages, SEEN languages (present in the model's pretraining data), and UNSEEN languages (not present or documented in the model's pretraining data in any meaningful way). We examine the impact of factors such as pretraining data size, general resource availability, language family, and script type on model performance. Decision tree analysis reveals that pretraining data size is the most influential factor for SEEN languages. However, interestingly, script type and language family are crucial for UNSEEN languages, highlighting the importance of cross-lingual transfer learning. Notably, model size and architecture do not significantly alter the most important features identified. Our findings provide valuable insights into the strengths and limitations of current MLLMs and hope to guide the development of more effective and equitable multilingual NLP systems.

Yufei Tao, Ameeta Agrawal, Judit Dombi et al.

Recent advances in interactive large language models like ChatGPT have revolutionized various domains; however, their behavior in natural and role-play conversation settings remains underexplored. In our study, we address this gap by deeply investigating how ChatGPT behaves during conversations in different settings by analyzing its interactions in both a normal way and a role-play setting. We introduce a novel dataset of broad range of human-AI conversations annotated with user motives and model naturalness to examine (i) how humans engage with the conversational AI model, and (ii) how natural are AI model responses. Our study highlights the diversity of user motives when interacting with ChatGPT and variable AI naturalness, showing not only the nuanced dynamics of natural conversations between humans and AI, but also providing new avenues for improving the effectiveness of human-AI communication.

Sina Bagheri Nezhad, Ameeta Agrawal, Rhitabrat Pokharel

Multilingual language models (MLLMs) are crucial for handling text across various languages, yet they often show performance disparities due to differences in resource availability and linguistic characteristics. While the impact of pre-train data percentage and model size on performance is well-known, our study reveals additional critical factors that significantly influence MLLM effectiveness. Analyzing a wide range of features, including geographical, linguistic, and resource-related aspects, we focus on the SIB-200 dataset for classification and the Flores-200 dataset for machine translation, using regression models and SHAP values across 204 languages. Our findings identify token similarity and country similarity as pivotal factors, alongside pre-train data and model size, in enhancing model performance. Token similarity facilitates cross-lingual transfer, while country similarity highlights the importance of shared cultural and linguistic contexts. These insights offer valuable guidance for developing more equitable and effective multilingual language models, particularly for underrepresented languages.

So Young Lee, Russell Scheinberg, Amber Shore et al.

This study explores how recent large language models (LLMs) navigate relative clause attachment {ambiguity} and use world knowledge biases for disambiguation in six typologically diverse languages: English, Chinese, Japanese, Korean, Russian, and Spanish. We describe the process of creating a novel dataset -- MultiWho -- for fine-grained evaluation of relative clause attachment preferences in ambiguous and unambiguous contexts. Our experiments with three LLMs indicate that, contrary to humans, LLMs consistently exhibit a preference for local attachment, displaying limited responsiveness to syntactic variations or language-specific attachment patterns. Although LLMs performed well in unambiguous cases, they rigidly prioritized world knowledge biases, lacking the flexibility of human language processing. These findings highlight the need for more diverse, pragmatically nuanced multilingual training to improve LLMs' handling of complex structures and human-like comprehension.

Atharva Phatak, Vijay K. Mago, Ameeta Agrawal et al.

The use of generative AI to create text descriptions from graphs has mostly focused on knowledge graphs, which connect concepts using facts. In this work we explore the capability of large pretrained language models to generate text from causal graphs, where salient concepts are represented as nodes and causality is represented via directed, typed edges. The causal reasoning encoded in these graphs can support applications as diverse as healthcare or marketing. Using two publicly available causal graph datasets, we empirically investigate the performance of four GPT-3 models under various settings. Our results indicate that while causal text descriptions improve with training data, compared to fact-based graphs, they are harder to generate under zero-shot settings. Results further suggest that users of generative AI can deploy future applications faster since similar performances are obtained when training a model with only a few examples as compared to fine-tuning via a large curated dataset.

Yufei Tao, Adam Hiatt, Rahul Seetharaman et al.

Large language models are capable of leveraging both contextual and parametric knowledge but how they prioritize and integrate these sources remains underexplored. We introduce CoPE, a novel evaluation framework that systematically measures contextual knowledge (CK) and parametric knowledge (PK) across models and languages. Using our MultiWikiAtomic dataset in English, Spanish, and Danish, we analyze how large language models (LLMs) integrate context, prioritize information, and incorporate PK in open-ended question answering. Our analysis uncovers a phenomenon we call lost-in-the-later, where LLMs tend to overlook or deprioritize information that appears later in a given context, revealing a strong positional bias that affects contextual grounding. We further find that reasoning models, as well as non-reasoning models prompted with chain-of-thought (CoT), use context even less than non-reasoning models without CoT and fail to mitigate the lost-in-the-later effect. CoT prompting, in particular, results in lower recall and shorter responses, leading to degraded contextual grounding. Based on these insights, we design prompt-based methods to effectively leverage input context. A case study applying CoPE to summarization demonstrates that CK-informed prompting improves factual grounding and reduces hallucination.

Sina Bagheri Nezhad, Ameeta Agrawal

Large language models (LLMs) often struggle to perform multi-target reasoning in long-context scenarios where relevant information is scattered across extensive documents. To address this challenge, we introduce NeuroSymbolic Augmented Reasoning (NSAR), which combines the benefits of neural and symbolic reasoning during inference. NSAR explicitly extracts symbolic facts from text and generates executable Python code to handle complex reasoning steps. Through extensive experiments across seven languages and diverse context lengths, we demonstrate that NSAR significantly outperforms both a vanilla RAG baseline and advanced prompting strategies in accurately identifying and synthesizing multiple pieces of information. Our results highlight the effectiveness of combining explicit symbolic operations with neural inference for robust, interpretable, and scalable reasoning in multilingual settings.

Enas Altarawneh, Ameeta Agrawal, Michael Jenkin et al.

Understanding and predicting the emotional trajectory in multi-party multi-turn conversations is of great significance. Such information can be used, for example, to generate empathetic response in human-machine interaction or to inform models of pre-emptive toxicity detection. In this work, we introduce the novel problem of Predicting Emotions in Conversations (PEC) for the next turn (n+1), given combinations of textual and/or emotion input up to turn n. We systematically approach the problem by modeling three dimensions inherently connected to evoked emotions in dialogues, including (i) sequence modeling, (ii) self-dependency modeling, and (iii) recency modeling. These modeling dimensions are then incorporated into two deep neural network architectures, a sequence model and a graph convolutional network model. The former is designed to capture the sequence of utterances in a dialogue, while the latter captures the sequence of utterances and the network formation of multi-party dialogues. We perform a comprehensive empirical evaluation of the various proposed models for addressing the PEC problem. The results indicate (i) the importance of the self-dependency and recency model dimensions for the prediction task, (ii) the quality of simpler sequence models in short dialogues, (iii) the importance of the graph neural models in improving the predictions in long dialogues.

Russell Scheinberg, Ameeta Agrawal, Amber Shore et al.

Large language models (LLMs) can explain grammatical rules, yet they often fail to apply those rules when judging sentence acceptability. We present "grammar prompting", an explain-then-process paradigm: a large LLM first produces a concise explanation of the relevant syntactic phenomenon, then that explanation is fed back as additional context to the target model -- either an LLM or a smaller language model (SLM) -- before deciding which sentence of a minimal pair is grammatical. On the English BLiMP, Chinese SLING, and Russian RuBLiMP benchmarks, this simple prompt design yields substantial improvements over strong baselines across many syntactic phenomena. Feeding an LLM's metalinguistic explanation back to the target model bridges the gap between knowing a rule and using it. On SLMs, grammar prompting alone trims the average LLM-SLM accuracy gap by about 20%, and when paired with chain-of-thought, by 56% (13.0 pp -> 5.8 pp), all at negligible cost. The lightweight, language-agnostic cue lets low-cost SLMs approach frontier-LLM performance in multilingual settings.

Olubusayo Olabisi, Ameeta Agrawal

Text summarization models have typically focused on optimizing aspects of quality such as fluency, relevance, and coherence, particularly in the context of news articles. However, summarization models are increasingly being used to summarize diverse sources of text, such as social media data, that encompass a wide demographic user base. It is thus crucial to assess not only the quality of the generated summaries, but also the extent to which they can fairly represent the opinions of diverse social groups. Position bias, a long-known issue in news summarization, has received limited attention in the context of social multi-document summarization. We deeply investigate this phenomenon by analyzing the effect of group ordering in input documents when summarizing tweets from three distinct linguistic communities: African-American English, Hispanic-aligned Language, and White-aligned Language. Our empirical analysis shows that although the textual quality of the summaries remains consistent regardless of the input document order, in terms of fairness, the results vary significantly depending on how the dialect groups are presented in the input data. Our results suggest that position bias manifests differently in social multi-document summarization, severely impacting the fairness of summarization models.

Olga Kolesnikova, Moein Shahiki Tash, Zahra Ahani et al.

The widespread use of social media highlights the need to understand its impact, particularly the role of online social support. This study uses a dataset focused on online social support, which includes binary and multiclass classifications of social support content on social media. The classification of social support is divided into three tasks. The first task focuses on distinguishing between supportive and non-supportive. The second task aims to identify whether the support is directed toward an individual or a group. The third task categorizes the specific type of social support, grouping it into categories such as Nation, LGBTQ, Black people, Women, Religion, and Other (if it does not fit into the previously mentioned categories). To address data imbalances in these tasks, we employed K-means clustering for balancing the dataset and compared the results with the original unbalanced data. Using advanced machine learning techniques, including transformers and zero-shot learning approaches with GPT3, GPT4, and GPT4-o, we predict social support levels in various contexts. The effectiveness of the dataset is evaluated using baseline models across different learning approaches, with transformer-based methods demonstrating superior performance. Additionally, we achieved a 0.4\% increase in the macro F1 score for the second task and a 0.7\% increase for the third task, compared to previous work utilizing traditional machine learning with psycholinguistic and unigram-based TF-IDF values.

Amber Shore, Russell Scheinberg, Ameeta Agrawal et al.

Large Language Models (LLMs) are intended to reflect human linguistic competencies. But humans have access to a broad and embodied context, which is key in detecting and resolving linguistic ambiguities, even in isolated text spans. A foundational case of semantic ambiguity is found in the task of coreference resolution: how is a pronoun related to an earlier person mention? This capability is implicit in nearly every downstream task, and the presence of ambiguity at this level can alter performance significantly. We show that LLMs can achieve good performance with minimal prompting in both coreference disambiguation and the detection of ambiguity in coreference, however, they cannot do both at the same time. We present the CORRECT-DETECT trade-off: though models have both capabilities and deploy them implicitly, successful performance balancing these two abilities remains elusive.

Philippe J. Giabbanelli, Ameeta Agrawal

Modeling & Simulation (M&S) approaches such as agent-based models hold significant potential to support decision-making activities in health, with recent examples including the adoption of vaccines, and a vast literature on healthy eating behaviors and physical activity behaviors. These models are potentially usable by different stakeholder groups, as they support policy-makers to estimate the consequences of potential interventions and they can guide individuals in making healthy choices in complex environments. However, this potential may not be fully realized because of the models' complexity, which makes them inaccessible to the stakeholders who could benefit the most. While Large Language Models (LLMs) can translate simulation outputs and the design of models into text, current approaches typically rely on one-size-fits-all summaries that fail to reflect the varied informational needs and stylistic preferences of clinicians, policymakers, patients, caregivers, and health advocates. This limitation stems from a fundamental gap: we lack a systematic understanding of what these stakeholders need from explanations and how to tailor them accordingly. To address this gap, we present a step-by-step framework to identify stakeholder needs and guide LLMs in generating tailored explanations of health simulations. Our procedure uses a mixed-methods design by first eliciting the explanation needs and stylistic preferences of diverse health stakeholders, then optimizing the ability of LLMs to generate tailored outputs (e.g., via controllable attribute tuning), and then evaluating through a comprehensive range of metrics to further improve the tailored generation of summaries.

Martha O. Dimgba, Sharon Oba, Ameeta Agrawal et al.

Language models have been shown to propagate social bias through their output, particularly in the representation of gender and ethnicity. This paper investigates gender and ethnicity biases in AI-generated occupational stories. Representation biases are measured before and after applying our proposed mitigation strategy, Bias Analysis and Mitigation through Explanation (BAME), revealing improvements in demographic representation ranging from 2% to 20%. BAME leverages model-generated explanations to inform targeted prompt engineering, effectively reducing biases without modifying model parameters. By analyzing stories generated across 25 occupational groups, three large language models (Claude 3.5 Sonnet, Llama 3.1 70B Instruct, and GPT-4 Turbo), and multiple demographic dimensions, we identify persistent patterns of overrepresentation and underrepresentation linked to training data stereotypes. Our findings demonstrate that guiding models with their own internal reasoning mechanisms can significantly enhance demographic parity, thereby contributing to the development of more transparent generative AI systems.

Jackson Trager, Francielle Vargas, Diego Alves et al.

Ensuring the moral reasoning capabilities of Large Language Models (LLMs) is a growing concern as these systems are used in socially sensitive tasks. Nevertheless, current evaluation benchmarks present two major shortcomings: a lack of annotations that justify moral classifications, which limits transparency and interpretability; and a predominant focus on English, which constrains the assessment of moral reasoning across diverse cultural settings. In this paper, we introduce MFTCXplain, a multilingual benchmark dataset for evaluating the moral reasoning of LLMs via multi-hop hate speech explanation using the Moral Foundations Theory. MFTCXplain comprises 3,000 tweets across Portuguese, Italian, Persian, and English, annotated with binary hate speech labels, moral categories, and text span-level rationales. Our results show a misalignment between LLM outputs and human annotations in moral reasoning tasks. While LLMs perform well in hate speech detection (F1 up to 0.836), their ability to predict moral sentiments is notably weak (F1 < 0.35). Furthermore, rationale alignment remains limited mainly in underrepresented languages. Our findings show the limited capacity of current LLMs to internalize and reflect human moral reasoning

So Young Lee, Russell Scheinberg, Amber Shore et al.

This study examines how large language models (LLMs) resolve relative clause (RC) attachment ambiguities and compares their performance to human sentence processing. Focusing on two linguistic factors, namely the length of RCs and the syntactic position of complex determiner phrases (DPs), we assess whether LLMs can achieve human-like interpretations amid the complexities of language. In this study, we evaluated several LLMs, including Claude, Gemini and Llama, in multiple languages: English, Spanish, French, German, Japanese, and Korean. While these models performed well in Indo-European languages (English, Spanish, French, and German), they encountered difficulties in Asian languages (Japanese and Korean), often defaulting to incorrect English translations. The findings underscore the variability in LLMs' handling of linguistic ambiguities and highlight the need for model improvements, particularly for non-European languages. This research informs future enhancements in LLM design to improve accuracy and human-like processing in diverse linguistic environments.

Rhitabrat Pokharel, Sina Bagheri Nezhad, Ameeta Agrawal et al.

The rapid advancement of Large Language Models (LLMs), particularly those trained on multilingual corpora, has intensified the need for a deeper understanding of their performance across a diverse range of languages and model sizes. Our research addresses this critical need by studying the performance and scaling behavior of multilingual LLMs in text classification and machine translation tasks across 204 languages. We systematically examine both seen and unseen languages across three model families of varying sizes in zero-shot and few-shot settings. Our findings show significant differences in scaling behavior between zero-shot and two-shot scenarios, with striking disparities in performance between seen and unseen languages. Model scale has little effect on zero-shot performance, which remains mostly flat. However, in two-shot settings, larger models show clear linear improvements in multilingual text classification. For translation tasks, however, only the instruction-tuned model showed clear benefits from scaling. Our analysis also suggests that overall resource levels, not just the proportions of pretraining languages, are better predictors of model performance, shedding light on what drives multilingual LLM effectiveness.

Yufei Tao, Tiernan Mines, Ameeta Agrawal

Conversation systems accommodate diverse users with unique personalities and distinct writing styles. Within the domain of multi-turn dialogue modeling, this work studies the impact of varied utterance lengths on the quality of subsequent responses generated by conversation models. Using GPT-3 as the base model, multiple dialogue datasets, and several metrics, we conduct a thorough exploration of this aspect of conversational models. Our analysis sheds light on the complex relationship between utterance lengths and the quality of follow-up responses generated by dialogue systems. Empirical findings suggests that, for certain types of conversations, utterance lengths can be reduced by up to 72% without any noticeable difference in the quality of follow-up responses.