NAACL 2025 Papers

A language can have different varieties. These varieties can affect the performance of natural language processing (NLP) models, including large language models (LLMs), which are often trained on data from widely spoken varieties. This paper introduces a novel and cost-effective approach to benchmark model performance across language varieties. We argue that international online review platforms,such as Booking.com, can serve as effective data sources for constructing datasets that capture comments in different language varieties from similar real-world scenarios, like reviews for the same hotel with the same rating using the same language (e.g., Mandarin Chinese) but different language varieties (e.g., Taiwan Mandarin, Mainland Mandarin). To prove this concept, we constructed a contextually aligned dataset comprising reviews in Taiwan Mandarin and Mainland Mandarin and tested six LLMs in a sentiment analysis task. Our results show that LLMs consistently underperform in Taiwan Mandarin.

Data abundance across different domains exhibits a long-tailed distribution: few domains have abundant data, while most face data scarcity. Our work focuses on a multilingual setting, where available data is heavily skewed toward high-resource languages, creating significant imbalances in training data sizes across languages. This disparity challenges training language models to perform uniformly well in all languages. Two common strategies to address this issue are upsampling low-resource languages (Temperature Sampling) and upweighting their loss functions (Scalarization). These methods are often assumed to be equivalent, but this equivalence has not been rigorously established, prompting our investigation.Through theoretical and empirical analysis, we identify when these two methods are equivalent and when they diverge. We prove that they are equivalent under full gradient descent but differ under stochastic gradient descent due to differences in gradient variance. Specifically, Temperature Sampling exhibits lower variance in gradient estimation, leading to faster convergence but a higher risk of overfitting. Based on these insights, we propose Cooldown, a strategy that starts by heavily upsampling low-resource languages to accelerate convergence and gradually reduces the upsampling to prevent overfitting—achieving the best of both worlds. Our method competes effectively with existing data re-weighting techniques while offering computational efficiency.

Recent work has demonstrated that the unequal representation of cultures and socioeconomic groups in training data leads to biased Large Multi-modal (LMM) models. To improve LMM model performance on underrepresented data, we propose and evaluate several prompting strategies using non-English, geographic, and socioeconomic attributes. We show that these geographic and socioeconomic integrated prompts favor retrieving topic appearances commonly found in data from low-income households across different countries leading to improved LMM model performance on lower-income data. Our analyses identify and highlight contexts where these strategies yield the most improvements.

Hate speech relies heavily on cultural influences, leading to varying individual interpretations. For that reason, we propose a Semantic Componential Analysis (SCA) framework for a cross-cultural and cross-domain analysis of hate speech definitions. We create the first dataset of hate speech definitions encompassing 493 definitions from more than 100 cultures, drawn from five key domains: online dictionaries, academic research, Wikipedia, legal texts, and online platforms. By decomposing these definitions into semantic components,our analysis reveals significant variation across definitions, yet many domains borrow definitions from one another without taking into account the target culture. We conduct zero-shot model experiments using our proposed dataset, employing three popular open-sourced LLMs to understand the impact of different definitions on hate speech detection. Our findings indicate that LLMs are sensitive to definitions: responses for hate speech detection change according to the complexity of definitions used in the prompt.

Sentence representation learning benefits from data augmentation strategies to improve model performance and generalization, yet existing approaches often encounter issues such as semantic inconsistencies and feature suppression. To address these limitations, we propose a method for generating Syntactically Aligned Negative (SAN) samples through a semantic importance-aware Masked Language Model (MLM) approach. Our method quantifies semantic contributions of individual words to produce negative samples that have substantial textual overlap with the original sentences while conveying different meanings. We further introduce Hierarchical-InfoNCE (HiNCE), a novel contrastive learning objective employing differential temperature weighting to optimize the utilization of both in-batch and syntactically aligned negative samples. Extensive evaluations across seven semantic textual similarity benchmarks demonstrate consistent improvements over state-of-the-art models.

Large language models (LLMs) have demonstrated remarkable capabilities in simulating human behaviour and social intelligence. However, they risk perpetuating societal biases, especially when demographic information is involved. We introduce a novel framework using cosine distance to measure semantic shifts in responses and an LLM-judged Preference Win Rate (WR) to assess how demographic prompts affect response quality across power-disparate social scenarios. Evaluating five LLMs over 100 diverse social scenarios and nine demographic axes, our findings suggest a “default persona” bias toward middle-aged, able-bodied, native-born, Caucasian, atheistic males with centrist views. Moreover, interactions involving specific demographics are associated with lower-quality responses. Lastly, the presence of power disparities increases variability in response semantics and quality across demographic groups, suggesting that implicit biases may be heightened under power-imbalanced conditions. These insights expose the demographic biases inherent in LLMs and offer potential paths toward future bias mitigation efforts in LLMs.

Text-to-SQL systems empower users to interact with databases using natural language, automatically translating queries into executable SQL code. However, their reliance on database schema information for SQL generation exposes them to significant security vulnerabilities, particularly schema inference attacks that can lead to unauthorized data access or manipulation. In this paper, we introduce a novel zero-knowledge framework for reconstructing the underlying database schema of text-to-SQL models without any prior knowledge of the database. Our approach systematically probes text-to-SQL models with specially crafted questions and leverages a surrogate GPT-4 model to interpret the outputs, effectively uncovering hidden schema elements—including tables, columns, and data types. We demonstrate that our method achieves high accuracy in reconstructing table names, with F1 scores of up to .99 for generative models and .78 for fine-tuned models, underscoring the severity of schema leakage risks. We also show that our attack can steal prompt information in non-text-to-SQL models. Furthermore, we propose a simple protection mechanism for generative models and empirically show its limitations in mitigating these attacks.

Large language models (LLMs) have demonstrated impressive task-solving capabilities through prompting techniques and system designs, including solving planning tasks (e.g., math proofs, basic travel planning) when sufficient data is available online and used during pre-training. However, for planning tasks with limited prior data (e.g., blocks world, advanced travel planning), the performance of LLMs, including proprietary models like GPT and Gemini, is poor. This paper investigates the impact of fine-tuning on the planning capabilities of LLMs, revealing that LLMs can achieve strong performance in planning through substantial (tens of thousands of specific examples) fine-tuning. Yet, this process incurs high economic, time, and computational costs for each planning problem variation. To address this, we propose Clustering-Based Maximum Diversity Sampling (CMDS), which selects diverse and representative data to enhance sample efficiency and the model’s generalization capability. Extensive evaluations demonstrate that CMDS-l, a baseline method combining CMDS with language embeddings, outperforms random sampling. Furthermore, we introduce a novel algorithm, CMDS-g, which encodes planning task instances with their graph representations into the embedding space. Empirical results show that CMDS-g consistently outperforms baseline methods across various scales and multiple benchmark domains.

The Sejong dictionary dataset offers a valuable resource, providing extensive coverage of morphology, syntax, and semantic representation. This dataset can be utilized to explore linguistic information in greater depth.The labeled linguistic structures within this dataset form the basis for uncovering relationships between words and phrases and their associations with target verbs. This paper introduces a user-friendly web interface designed for the collection and consolidation of verb-related information, with a particular focus on subcategorization frames. Additionally, it outlines our efforts in mapping this information by aligning subcategorization frames with corresponding illustrative sentence examples.Furthermore, we provide a Python library that would simplify syntactic parsing and semantic role labeling. These tools are intended to assist individuals interested in harnessing the Sejong dictionary dataset to develop applications for Korean language processing.

The task of multi-objective alignment aims at balancing and controlling the different alignment objectives, e.g., helpfulness, harmlessness and honesty) of large language models to meet the personalized requirements of different users. However, previous methods tend to train multiple models to deal with various user preferences, with the number of trained models growing linearly with the number of alignment objectives and the number of different preferences. Meanwhile, existing methods are generally poor in extensibility and require significant re-training for each new alignment objective considered. Considering the limitation of previous approaches, we propose MCA, which constructs an expert prompt and an adversarial prompt for each objective to contrast at the decoding time and balances the objectives through combining the contrast. Our approach is verified to be superior to previous methods in obtaining a well-distributed Pareto front among different alignment objectives.

Multi-hop reasoning, which requires multi-step reasoning based on the supporting documents within a given context, remains challenging for large language models (LLMs). LLMs often struggle to filter out irrelevant documents within the context, and their performance is sensitive to the absolute position of supporting documents within that context. In this paper, we identify an additional challenge: LLMs’ performance is also sensitive to the order, relative position, in which the supporting documents are presented. We refer to this as the misordered context problem. To address this issue, based on the theoretical approach, we propose a simple yet effective method called context repetition (CoRe), which involves prompting the model by repeatedly presenting the context. This ensures that certain contiguous reasoning segments within supporting documents are presented in the optimal order, effectively guiding the model’s reasoning in the appropriate direction. Applying CoRe, we improve the F1 score by up to 30%p on multi-hop QA tasks and increase accuracy by up to 70%p on a synthetic task. Additionally, CoRe helps mitigate the well-known “lost-in-the-middle” problem in LLMs and can be effectively combined with retrieval-based approaches utilizing Chain-of-Thought (CoT) reasoning.

Machine unlearning has been used to remove unwanted knowledge acquired by large language models (LLMs). In this paper, we examine machine unlearning from an optimization perspective, framing it as a regularized multi-task optimization problem, where one task optimizes a forgetting objective and another optimizes the model performance. In particular, we introduce a normalized gradient difference algorithm, enabling us to have better control over the trade-off between the objectives, while integrating a new, automatic learning rate scheduler. We provide a theoretical analysis and empirically demonstrate the superior performance of among state-of-the-art unlearning methods on the TOFU and MUSE datasets while exhibiting stable training.

In this paper, we investigate whether current state-of-the-art large language models (LLMs) are effective as AI tutors and whether they demonstrate pedagogical abilities necessary for good AI tutoring in educational dialogues. Previous efforts towards evaluation have beenlimited to subjective protocols and benchmarks. To bridge this gap, we propose a unified evaluation taxonomy with eight pedagogical dimensions based on key learning sciences principles, which is designed to assess the pedagogical value of LLM-powered AI tutor responses grounded in student mistakes or confusions in the mathematical domain. We release MRBench – a new evaluation benchmark containing 192 conversations and 1,596 responses from seven state-of-the-art LLM-based and human tutors, providing gold annotations for eight pedagogical dimensions. We assess reliability of the popular Prometheus2 and Llama-3.1-8B LLMs as evaluators and analyze each tutor’s pedagogical abilities, highlighting which LLMs are good tutors and which ones are more suitable as question-answering systems. We believe that the presented taxonomy, benchmark, and human-annotated labels will streamline the evaluation process and help track the progress in AI tutors’ development.

Significant advancements has recently been achieved in the field of multi-modal large language models (MLLMs), demonstrating their remarkable capabilities in understanding and reasoning across diverse tasks. However, these models are often trained for specific tasks and rely on task-specific input-output formats, limiting their applicability to a broader range of tasks. This raises a fundamental question: Can we develop a unified approach to represent and handle different multi-modal tasks to maximize the generalizability of MLLMs? In this paper, we propose UnifiedMLLM, a comprehensive model designed to represent various tasks using a unified representation. Our model exhibits strong capabilities in comprehending the implicit intent of user instructions and preforming reasoning. In addition to generating textual responses, our model also outputs task tokens and grounding tokens, serving as indicators of task types and task granularity. These outputs are subsequently routed through the task router and directed to specific expert models for task completion. To train our model, we construct a task-specific dataset and an 100k multi-task dataset encompassing complex scenarios. Employing a three-stage training strategy, we equip our model with robust reasoning and task processing capabilities while preserving its generalization capacity and knowledge reservoir. Extensive experiments showcase the impressive performance of our unified representation approach across various tasks, surpassing existing methodologies. Furthermore, our approach exhibits exceptional scalability and generality.

This study explores the integration of automated writing evaluation (AWE) and grammatical error correction (GEC) through multitask learning, demonstrating how combining these distinct tasks can enhance performance in both areas. By leveraging a shared learning framework, we show that models trained jointly on AWE and GEC outperform those trained on each task individually. To support this effort, we introduce a dataset specifically designed for multitask learning using AWE and GEC. Our experiments reveal significant synergies between tasks, leading to improvements in both writing assessment accuracy and error correction precision. This research represents a novel approach for optimizing language learning tools by unifying writing evaluation and correction tasks, offering insights into the potential of multitask learning in educational applications.

Recently, Large Vision Language Models (LVLMs) have unlocked many complex use cases that require Multi-Modal (MM) understanding (e.g., image captioning or visual question answering) and MM generation (e.g., text-guided image generation or editing) capabilities. To further improve the output fidelity of LVLMs we introduce UniRAG, a plug-and-play technique that adds relevant retrieved information to prompts as few-shot examples during inference. Unlike the common belief that Retrieval Augmentation (RA) mainly improves generation or understanding of uncommon entities, our evaluation results on the MSCOCO dataset with common entities show that both proprietary models like GPT-4o and Gemini-Pro and smaller open-source models like LLaVA, LaVIT, and Emu2 significantly enhance their generation quality when their input prompts are augmented with relevant information retrieved by Vision-Language (VL) retrievers like UniIR models. All the necessary code to reproduce our results is available at https://github.com/castorini/UniRAG.

Existing information retrieval (IR) models often assume a homogeneous structure for knowledge sources and user queries, limiting their applicability in real-world settings where retrieval is inherently heterogeneous and diverse. In this paper, we introduce UniHGKR, a unified instruction-aware heterogeneous knowledge retriever that (1) builds a unified retrieval space for heterogeneous knowledge and (2) follows diverse user instructions to retrieve knowledge in specified types. UniHGKR consists of three principal stages, including heterogeneous self-supervised pretraining, text-anchored embedding alignment, and instruction-aware retriever fine-tuning, enabling it to generalize across varied retrieval contexts. This framework is highly scalable, with a BERT-based version and a UniHGKR-7B version trained on large language models. Also, we introduce CompMix-IR, the first native heterogeneous knowledge retrieval benchmark. It includes two retrieval scenarios with various instructions, over 9,400 question answer (QA) pairs, and a corpus of 10 million entries, covering four different types of data. Extensive experiments show that UniHGKR consistently outperform state-of-the-art methods on CompMix-IR, achieving up to 6.36% and 54.23% relative improvements in two scenarios, respectively. Finally, by equipping our retriever for open-domain heterogeneous QA systems, we achieve a new state-of-the-art result on the popular ConvMix task, with an absolute improvement of up to 5.90 points.

In the fast-changing realm of information, the capacity to construct coherent timelines from extensive event-related content has become increasingly significant and challenging. The complexity arises in aggregating related documents to build a meaningful event graph around a central topic. This paper proposes CHRONOS - Causal Headline Retrieval for Open-domain News Timeline SummarizatiOn via Iterative Self-Questioning, which offers a fresh perspective on the integration of Large Language Models (LLMs) to tackle the task of Timeline Summarization (TLS). By iteratively reflecting on how events are linked and posing new questions regarding a specific news topic to gather information online or from an offline knowledge base, LLMs produce and refresh chronological summaries based on documents retrieved in each round. Furthermore, we curate Open-TLS, a novel dataset of timelines on recent news topics authored by professional journalists to evaluate open-domain TLS where information overload makes it impossible to find comprehensive relevant documents from the web. Our experiments indicate that CHRONOS is not only adept at open-domain timeline summarization but also rivals the performance of existing state-of-the-art systems designed for closed-domain applications, where a related news corpus is provided for summarization.

Large language models are known to hallucinate, but the underlying mechanism that govern how models hallucinate are not yet fully understood. In this work, we find that unfamiliar examples in the models’ finetuning data – those that introduce concepts beyond the base model’s scope of knowledge – are crucial in shaping these errors. In particular, we find that an LLM’s hallucinated predictions tend to mirror the responses associated with its unfamiliar finetuning examples. This suggests that by modifying how unfamiliar finetuning examples are supervised, we can influence a model’s responses to unfamiliar queries (e.g., say “I don’t know”). We empirically validate this observation in a series of controlled experiments involving SFT, RL, and reward model finetuning on TriviaQA and MMLU. Our work further investigates RL finetuning strategies for improving the factuality of long-form model generations. We find that, while hallucinations from the reward model can significantly undermine the effectiveness of RL factuality finetuning, strategically controlling how reward models hallucinate can minimize these negative effects. Leveraging our previous observations on controlling hallucinations, we propose an approach for learning more reliable reward models, and show that they improve the efficacy of RL factuality finetuning in long-form biography and book/movie plot generation tasks.