NAACL 2025 Papers

Training conversational question-answering (QA) systems demands a substantial amount of in-domain data, which is often scarce in practice. A common solution to this challenge is to generate synthetic data. Traditional methods typically follow a top-down approach, where a large language model (LLM) generates multi-turn dialogues from a broad prompt. While this method produces coherent conversations, it offers limited fine-grained control over the content and is susceptible to hallucinations. We introduce a bottom-up conversation synthesis approach, where QA pairs are generated first and then combined into a coherent dialogue. This method offers greater control and precision by dividing the process into two distinct steps, enabling refined instructions and validations to be handled separately. Additionally, this structure allows the use of non-local models in stages that do not involve proprietary knowledge, enhancing the overall quality of the generated data. Both human and automated evaluations demonstrate that our approach produces more realistic and higher-quality dialogues compared to top-down methods.

This paper introduces BnTTS (Bangla Text-To-Speech), the first framework for Bangla speaker adaptation-based TTS, designed to bridge the gap in Bangla speech synthesis using minimal training data. Building upon the XTTS architecture, our approach integrates Bangla into a multilingual TTS pipeline, with modifications to account for the phonetic and linguistic characteristics of the language. We pretrain BnTTS on 3.85k hours of Bangla speech dataset with corresponding text labels and evaluate performance in both zero-shot and few-shot settings on our proposed test dataset. Empirical evaluations in few-shot settings show that BnTTS significantly improves the naturalness, intelligibility, and speaker fidelity of synthesized Bangla speech. Compared to state-of-the-art Bangla TTS systems, BnTTS exhibits superior performance in Subjective Mean Opinion Score (SMOS), Naturalness, and Clarity metrics.

Large Vision Language Models (LVLMs) often suffer from object hallucination, which undermines their reliability. Surprisingly, we find that simple object-based visual prompting—overlaying visual cues (e.g., bounding box, circle) on images—can significantly mitigate such hallucination; however, different visual prompts (VPs) vary in effectiveness. To address this, we propose Black-Box Visual Prompt Engineering (BBVPE), a framework to identify optimal VPs that enhance LVLM responses without needing access to model internals. Our approach employs a pool of candidate VPs and trains a router model to dynamically select the most effective VP for a given input image. This black-box approach is model-agnostic, making it applicable to both open-source and proprietary LVLMs. Evaluations on benchmarks such as POPE and CHAIR demonstrate that BBVPE effectively reduces object hallucination.

Phishing often targets victims through visually perturbed texts to bypass security systems. The noise contained in these texts functions as an adversarial attack, designed to deceive language models and hinder their ability to accurately interpret the content. However, since it is difficult to obtain sufficient phishing cases, previous studies have used synthetic datasets that do not contain real-world cases. In this study, we propose the BitAbuse dataset, which includes real-world phishing cases, to address the limitations of previous research. Our dataset comprises a total of 325,580 visually perturbed texts. The dataset inputs are drawn from the raw corpus, consisting of visually perturbed sentences and sentences generated through an artificial perturbation process. Each input sentence is labeled with its corresponding ground truth, representing the restored, non-perturbed version. Language models trained on our proposed dataset demonstrated significantly better performance compared to previous methods, achieving an accuracy of approximately 96%. Our analysis revealed a significant gap between real-world and synthetic examples, underscoring the value of our dataset for building reliable pre-trained models for restoration tasks. We release the BitAbuse dataset, which includes real-world phishing cases annotated with visual perturbations, to support future research in adversarial attack defense.

We study the evolution of opinions inside a population of interacting large language models (LLMs). Every LLM needs to decide how much funding to allocate to an item with three initial possibilities: full, partial, or no funding. We identify biases that drive the exchange of opinions based on the LLM’s tendency to find consensus with the other LLM’s opinion, display caution when specifying funding, and consider ethical concerns in its opinion. We find these biases are affected by the perceived absence of compelling reasons for opinion change, the perceived willingness to engage in discussion, and the distribution of allocation values. Moreover, tensions among biases can lead to the survival of funding for items with negative connotations. We also find that the final distribution of full, partial, and no funding opinions is more diverse when an LLM freely forms its opinion after an interaction than when its opinion is a multiple-choice selection among the three allocation options. In the latter case, consensus is mostly attained. When agents are aware of past opinions, they seek to maintain consistency with them, changing the opinion dynamics. Our study is performed using Llama 3 and Mistral LLMs.

In an effort to mitigate the harms of large language models (LLMs), learning from human feedback (LHF) has been used to steer LLMs towards outputs that are intended to be both less harmful and more helpful. Despite the widespread adoption of LHF in practice, the quality of this feedback and its effectiveness as a safety mitigation technique remain unclear. This study addresses these issues by auditing the widely-used Helpful and Harmless (HH) dataset by Anthropic. Our work includes: (1) a thorough investigation of the dataset’s content through both manual and automated evaluation; (2) experiments demonstrating the dataset’s impact on models’ safety; and (3) an analysis of the 100 most influential papers citing this dataset. Through our audit, we showcase how conceptualization failures and quality issues identified in the HH dataset can create additional harms by leading to disparate safety behaviors across demographic groups. Our findings highlight the need for more nuanced, context-sensitive approaches to safety mitigation in LLMs.

This paper delves into sequence-level knowledge distillation (KD) of multilingual pre-trained translation models. We posit that, beyond the approximated mode obtained via beam search, the whole output distribution of the teacher contains valuable insights for students. We explore the potential of n-best lists from beam search to guide student’s learning and then investigate alternative decoding methods to address observed issues like low variability and under-representation of infrequent tokens. Our research in data-limited scenarios reveals that although sampling methods can slightly compromise the translation quality of the teacher output compared to beam search based methods, they enrich the generated corpora with increased variability and lexical richness, ultimately enhancing student model performance and reducing the gender bias amplification commonly associated with KD.

Investigating value alignment in Large Language Models (LLMs) based on cultural context has become a critical area of research. However, similar biases have not been extensively explored in large vision-language models (VLMs). As the scale of multimodal models continues to grow, it becomes increasingly important to assess whether images can serve as reliable proxies for culture and how these values are embedded through the integration of both visual and textual data. In this paper, we conduct a thorough evaluation of multimodal model at different scales, focusing on their alignment with cultural values. Our findings reveal that, much like LLMs, VLMs exhibit sensitivity to cultural values, but their performance in aligning with these values is highly context-dependent. While VLMs show potential in improving value understanding through the use of images, this alignment varies significantly across contexts highlighting the complexities and underexplored challenges in the alignment of multimodal models.

Large language models (LLMs) have achieved remarkable success but still tend to generate factually erroneous responses, a phenomenon known as hallucination. A recent trend is to use preference learning to fine-tune models to align with factuality. However, existing work primarily evaluates fine-tuned models on in-domain (ID) datasets and the factuality on out-of-domain (OOD) datasets remains underexplored. In this paper, we conduct a comprehensive evaluation of the factuality of different models tuned by various preference learning algorithms and demonstrate that their performance on OOD datasets either increases minimally or decreases. Subsequently, we reveal that the main cause of model’s failure to uphold factuality under a distribution shift is under-alignment, rather than over-alignment, by analyzing the token distribution shift of the models before and after tuning. Finally, we propose APEFT (Atomic Preference Enhanced Factuality Tuning), a framework that enhances model’s awareness of factuality at the granularity of individual facts. Extensive experiments demonstrate that APEFT improves model performance by an average of on both ID and OOD datasets, which is highly effective.

The rapid development of Large Multimodal Models (LMMs) has significantly advanced multimodal understanding by harnessing the language abilities of Large Language Models (LLMs) and integrating modality-specific encoders. However, LMMs are plagued by hallucinations that limit their reliability and adoption. While traditional methods to detect and mitigate these hallucinations often involve costly training or rely heavily on external models, recent approaches utilizing internal model features present a promising alternative. In this paper, we critically assess the limitations of the state-of-the-art training-free technique, the logit lens, in handling generalized visual hallucinations. We introduce *ContextualLens*, a refined method that leverages contextual token embeddings from middle layers of LMMs. This approach significantly improves hallucination detection and grounding across diverse categories, including actions and OCR, while also excelling in tasks requiring contextual understanding, such as spatial relations and attribute comparison. Our novel grounding technique yields highly precise bounding boxes, facilitating a transition from Zero-Shot Object Segmentation to Grounded Visual Question Answering. Our contributions pave the way for more reliable and interpretable multimodal models.

Previous work has considered token overlap, or even similarity of token distributions, as predictors for multilinguality and cross-lingual knowledge transfer in language models. However, these very literal metrics assign large distances to language pairs with different scripts, which can nevertheless show good cross-linguality. This limits the explanatory strength of token overlap for knowledge transfer between language pairs that use distinct scripts or follow different orthographic conventions. In this paper, we propose subword token alignability as a new way to understand the impact and quality of multilingual tokenisation. In particular, this metric predicts multilinguality much better when scripts are disparate and the overlap of literal tokens is low. We analyse this metric in the context of both encoder and decoder models, look at data size as a potential distractor, and discuss how this insight may be applied to multilingual tokenisation in future work. We recommend our subword token alignability metric for identifying optimal language pairs for cross-lingual transfer, as well as to guide the construction of better multilingual tokenisers in the future. We publish our code and reproducibility details.

Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning large language models (LLMs) with human values. However, it has been noted that reward models in RLHF often exhibit unintended biases, such as an overemphasis on response length based on the erroneous assumption that longer responses are universally preferred. This “length bias” can lead to excessively verbose responses that compromise the quality of LLMs alignment. Previous efforts to mitigate length bias in reward models have inadvertently decreased their accuracy by neglecting the legitimate influence of response length on human preferences. In this work, we argue that response length is a context-specific factor in human evaluations, with different queries naturally eliciting varying preferences for response length. We propose an adaptive approach to modeling length preference that dynamically adjusts the influence of response length in reward evaluations according to the context of the query. Experimental results demonstrate that our adaptive approach effectively balances the mitigation of undesired length hacking and alignment accuracy, reducing unnecessary verbosity while improving overall response quality.

Despite advances in the multilingual capabilities of Large Language Models (LLMs) across diverse tasks, English remains the dominant language for LLM research and development. So, when working with a different language, this has led to the widespread practice of pre-translation, i.e., translating the task prompt into English before inference. Selective pre-translation, a more surgical approach, focuses on translating specific prompt components. However, its current use is sporagic and lacks a systematic research foundation. Consequently, the optimal pre-translation strategy for various multilingual settings and tasks remains unclear. In this work, we aim to uncover the optimal setup for pre-translation by systematically assessing its use. Specifically, we view the prompt as a modular entity, composed of four functional parts: instruction, context, examples, and output, either of which could be translated or not. We evaluate pre-translation strategies across 35 languages covering both low and high-resource languages, on various tasks including Question Answering (QA), Natural Language Inference (NLI), Named Entity Recognition (NER), and Abstractive Summarization. Our experiments show the impact of factors as similarity to English, translation quality and the size of pre-trained data, on the model performance with pre-translation. We suggest practical guidelines for choosing optimal strategies in various multilingual settings

Flowcharts are typically presented as images, driving the trend of using vision-language models (VLMs) for end-to-end flowchart understanding. However, two key challenges arise: (i) Limited controllability—users have minimal influence over the downstream task, as they can only modify input images, while the training of VLMs is often out of reach for most researchers. (ii) Lack of explainability—it is difficult to trace VLM errors to specific causes, such as failures in visual encoding or reasoning. We propose TextFlow, addressing aforementioned issues with two stages: (i) Vision Textualizer—which generates textual representations from flowchart images; and (ii) Textual Reasoner—which performs question-answering based on the text representations. TextFlow offers three key advantages: (i) users can select the type of text representations (e.g., Graphviz, Mermaid, PlantUML), or further convert them into executable graph object to call tools, enhancing performance and controllability; (ii) it improves explainability by helping to attribute errors more clearly to visual or textual processing components; and (iii) it promotes the modularization of the solution, such as allowing advanced LLMs to be used in the reasoner stage when VLMs underperform in end-to-end fashion. Experiments on the FlowVQA and FlowLearn benchmarks demonstrate TextFlow’s state-of-the-art performance as well as its robustness. All code and data are publicly available.

Cross-Document Event Coreference (CDEC) annotation is challenging and difficult to scale, resulting in existing datasets being small and lacking diversity. We introduce a new approach leveraging large language models (LLMs) to decontextualize event mentions, by simplifying the document-level annotation task to sentence pairs with enriched context, enabling the creation of Richer EventCorefBank (RECB), a denser and more expressive dataset annotated at faster speed. Decontextualization has been shown to improve annotation speed without compromising quality and to enhance model performance. Our baseline experiment indicates that systems trained on RECB achieve comparable results on the EventCorefBank(ECB+) test set, showing the high quality of our dataset and its generalizability on other CDEC datasets. In addition, our evaluation shows that the strong baseline models are still struggling with RECB comparing to other CDEC datasets, suggesting that the richness and diversity of RECB present significant challenges to current CDEC systems.

Given the large number of Hindi speakers worldwide, there is a pressing need for robust and efficient information retrieval systems for Hindi. Despite ongoing research, comprehensive benchmarks for evaluating retrieval models in Hindi are lacking. To address this gap, we introduce the Hindi-BEIR benchmark, comprising 15 datasets across seven distinct tasks. We evaluate state-of-the-art multilingual retrieval models on the Hindi-BEIR benchmark, identifying task and domain-specific challenges that impact Hindi retrieval performance. Building on the insights from these results, we introduce NLLB-E5, a multilingual retrieval model that leverages a zero-shot approach to support Hindi without the need for Hindi training data. We believe our contributions, including the release of the Hindi-BEIR benchmark and the NLLB-E5 model, will be a valuable resource for researchers and promote advancements in multilingual retrieval models.

LLMs have demonstrated impressive performance in answering medical questions, such as achieving passing scores on medical licensing examinations. However, medical board exams or general clinical questions do not capture the complexity of realistic clinical cases. Moreover, the lack of reference explanations means we cannot easily evaluate the reasoning of model decisions, a crucial component of supporting doctors in making complex medical decisions. To address these challenges, we construct two new datasets: JAMA Clinical Challenge and Medbullets. JAMA Clinical Challenge consists of questions based on challenging clinical cases, while Medbullets comprises simulated clinical questions. Both datasets are structured as multiple-choice question-answering tasks, accompanied by expert-written explanations. We evaluate seven LLMs on the two datasets using various prompts. Experiments demonstrate that our datasets are harder than previous benchmarks. In-depth automatic and human evaluations of model-generated explanations provide insights into the promise and deficiency of LLMs for explainable medical QA.

As LLMs become increasingly prevalent, it is interesting to consider how “creative” these models can be. From cognitive science, creativity consists of at least two key characteristics: convergent thinking (purposefulness to achieve a given goal) and divergent thinking (adaptability to explore new environments or constraints) (CITATION). In this work, we introduce a framework for quantifying LLM creativity that incorporates the two design ingredients: (1) We introduce DENIAL PROMPTING which pushes LLMs to develop more creative solutions to a given problem by incrementally imposing new constraints on the previous solution, compelling LLMs to adopt new strategies. (2) We define NEOGAUGE, a metric that quantifies both convergent and divergent thinking in the generated creative responses by LLMs. We test the proposed framework on Codeforces problems, which serve as both a natural dataset for coding tasks and a collection of prior human solutions. We quantify NEOGAUGE for various proprietary and open-source models and find that even the most creative model, GPT-4, still falls short of demonstrating human-like creativity. We also experiment with advanced reasoning strategies (MCTS, self-correction, etc.) and observe no significant improvement in creativity. As a by-product of our analysis, we release NEOCODER dataset for reproducing our results on future models.

The integration of tools has extended the capabilities of language models (LMs) beyond vanilla text generation to versatile scenarios. However, tool-augmented language models (TaLMs) often assume ‘perfect’ information access and tool availability, which may not hold in the real world. To systematically study TaLMs imperfections, we introduce the FAIL-TaLMs benchmark, featuring two major failures: under-specified user queries and non-available tools. FAIL-TaLMS contains 1,749 examples using 906 tools across 21 categories, including single- and multi-tool usage. We evaluate top-performing proprietary and open-source models, and find all current models except for Claude struggle to recognize missing tools or information. Further, to study possible mitigation of the failures, we enable real-time human interaction, named the Ask-and-Help method, to provide missing information or replace non-functional tools. While Ask-and-Help can help models solve tasks more correctly when queries are under-specified, it brings minimal benefit when complex tools are broken.