NAACL 2025 论文列表

Unrestricted access to external Large Language Models (LLM) based services like ChatGPT and Gemini can lead to potential data leakages, especially for large enterprises providing products and services to clients that require legal confidentiality guarantees. However, a blanket restriction on such services is not ideal as these LLMs boost employee productivity. Our goal is to build a solution that enables enterprise employees to query such external LLMs, without leaking confidential internal and client information. In this paper, we propose QueryShield - a platform that enterprises can use to interact with external LLMs without leaking data through queries. It detects if a query leaks data, and rephrases it to minimize data leakage while limiting the impact to its semantics. We construct a dataset of 1500 queries and manually annotate them for their sensitivity labels and their low sensitivity rephrased versions. We fine-tune a set of lightweight model candidates using this dataset and evaluate them using multiple metrics including one we propose specific to this problem.

Dense passage retrieval enhances Information Retrieval (IR) by encoding queries and passages into representation space. However, passage representations often fail to be referenced by their gold queries under domain shifts, revealing a weakness in representation space. One desirable concept for representations is ”argmaxable”. Being argmaxable ensures that no representations are theoretically excluded from selection due to geometric constraints. To be argmaxable, a notable approach is to increase isotropy, where representations are evenly spread out in all directions. These findings, while desirable also for IR, focus on passage representation and not on query, making it challenging to directly apply their findings to IR. In contrast, we introduce a novel query-focused concept of ”referentiable” tailored for IR tasks, which ensures that passage representations are referenced by their gold queries. Building on this, we propose Learning Referentiable Representation (LRR), and two strategic metrics, Self-P and Self-Q, quantifying how the representations are referentiable. Our experiments compare three dense model versions: Naive, Isotropic, and Referentiable, demonstrating that LRR leads to enhanced zero-shot performance, surpassing existing naive and isotropic versions.

This paper addresses the challenge of detecting query variants—pairs of queries with identical intents. One application in commercial search engines is reformulating user queries with its variant online. While measuring pairwise query similarity has been an established standard, it often falls short of capturing semantic equivalence when word forms or order differ. We propose leveraging the retrieval as an environment feedback (EF), based on the premise that desirable retrieval outcomes from equivalent queries should be interchangeable. Experimental results on both proprietary and public datasets demonstrate the efficacy of the proposed method, both with and without LLM calls.

Online discussion forums provide crucial data to understand the concerns of a wide range of real-world communities. However, the typical qualitative and quantitative methodologies used to analyze those data, such as thematic analysis and topic modeling, are infeasible to scale or require significant human effort to translate outputs to human readable forms. This study introduces QuaLLM, a novel LLM-based framework to analyze and extract quantitative insights from text data on online forums. The framework consists of a novel prompting and human evaluation methodology. We applied this framework to analyze over one million comments from two of Reddit’s rideshare worker communities, marking the largest study of its type. We uncover significant worker concerns regarding AI and algorithmic platform decisions, responding to regulatory calls about worker insights. In short, our work sets a new precedent for AI-assisted quantitative data analysis to surface concerns from online forums.

Chinese Search Query Spell Correction is a task designed to autonomously identify and correct typographical errors within queries in the search engine. Despite the availability of comprehensive datasets like Microsoft Speller and Webis, their monolingual nature and limited scope pose significant challenges in evaluating modern pre-trained language models such as BERT and GPT. To address this, we introduce QSpell 250K, a large-scale benchmark specifically developed for Chinese Query Spelling Correction. QSpell 250K offers several advantages: 1) It contains over 250K samples, which is ten times more than previous datasets. 2) It covers a broad range of topics, from formal entities to everyday colloquialisms and idiomatic expressions. 3) It includes both Chinese and English, addressing the complexities of code-switching. Each query undergoes three rounds of high-fidelity annotation to ensure accuracy. Our extensive testing across three popular models demonstrates that QSpell 250K effectively evaluates the efficacy of representative spelling correctors. We believe that QSpell 250K will significantly advance spelling correction methodologies. The accompanying data and code will be made publicly available.

The rise of large language models (LLMs) has significantly advanced various natural language processing (NLP) tasks. However, the resource demands of these models pose substantial challenges. Structured pruning is an effective approach to reducing model size, but it often results in significant accuracy degradation, necessitating parameter updates to adapt. Unfortunately, such fine-tuning requires substantial memory, which limits its applicability. To address these challenges, we introduce quantization into the structured pruning framework to reduce memory consumption during both fine-tuning and inference. However, the combined errors from pruning and quantization increase the difficulty of fine-tuning, requiring a more refined quantization scheme. To this end, we propose QPruner, a novel framework that employs structured pruning to reduce model size, followed by a layer-wise mixed-precision quantization scheme. Quantization precisions are assigned to each layer based on their importance to the target task, and Bayesian optimization is employed to refine precision allocation strategies, ensuring a balance between model accuracy and memory efficiency. Extensive experiments on benchmark datasets demonstrate that QPruner significantly outperforms existing methods in memory savings while maintaining or improving model performance.

In typical multimodal tasks, such as Visual Question Answering (VQA), adversarial attacks targeting a specific image and question can lead large vision-language models (LVLMs) to provide incorrect answers. However, it is common for a single image to be associated with multiple questions, and LVLMs may still answer other questions correctly even for an adversarial image attacked by a specific question. To address this, we introduce the query-agnostic visual attack (QAVA), which aims to create robust adversarial examples that generate incorrect responses to unspecified and unknown questions. Compared to traditional adversarial attacks focused on specific images and questions, QAVA significantly enhances the effectiveness and efficiency of attacks on images when the question is unknown, achieving performance comparable to attacks on known target questions. Our research broadens the scope of visual adversarial attacks on LVLMs in practical settings, uncovering previously overlooked vulnerabilities, particularly in the context of visual adversarial threats. The code is available at https://github.com/btzyd/qava.

Many adversarial attack approaches are proposed to verify the vulnerability of language models. However, they require numerous queries and the information on the target model. Even black-box attack methods also require the target model’s output information. They are not applicable in real-world scenarios, as in hard black-box settings where the target model is closed and inaccessible. Even the recently proposed hard black-box attacks still require many queries and demand extremely high costs for training adversarial generators. To address these challenges, we propose Q-faker (Query-free Hard Black-box Attacker), a novel and efficient method that generates adversarial examples without accessing the target model. To avoid accessing the target model, we use a surrogate model instead. The surrogate model generates adversarial sentences for a target-agnostic attack. During this process, we leverage controlled generation techniques. We evaluate our proposed method on eight datasets. Experimental results demonstrate our method’s effectiveness including high transferability and the high quality of the generated adversarial examples, and prove its practical in hard black-box settings.

The task of predicting time and location from images is challenging and requires complex human-like puzzle-solving ability over different clues. In this work, we formalize this ability into core skills and implement them using different modules in an expert pipeline called PuzzleGPT. PuzzleGPT consists of a perceiver to identify visual clues, a reasoner to deduce prediction candidates, a combiner to combinatorially combine information from different clues, a web retriever to get external knowledge if the task can’t be solved locally, and a noise filter for robustness. This results in a zero-shot, interpretable, and robust approach that records state-of-the-art performance on two datasets – TARA and WikiTilo. PuzzleGPT outperforms large VLMs such as BLIP-2, InstructBLIP, LLaVA, and even GPT-4V, as well as automatically generated reasoning pipelines like VisProg, by at least 32% and 38%, respectively. It even rivals or surpasses finetuned models.

In this work we present Pula, a suite of bilingual language models proficient in both Setswana and English. Leveraging recent advancements in data availability and efficient fine-tuning, Pula 8B and Pula 14B outperform GPT-4o and Gemini 1.5 Pro on English-Setswana translation tasks and achieve state-of-the-art performance on Setswana reasoning tasks for their size. We release the weights for Pula 1B, 3B, 8B, and 14B as well as training logs and training and evaluation code. Alongside Pula, we release the largest-ever Setswana text corpus, Marothodi, and the first comprehensive Setswana instruction-tuning dataset, Medupi, consisting of reformatted datasets, translated corpora, and synthetic LLM-generated text. To accompany this data, we release the code used for dataset construction, formatting, filtering, and scraping. Last, we release two Setswana LLM-translated benchmarks, MMLU-tsn and GSM8K-tsn, to measure Setswana knowledge and reasoning capabilities.

Performance prediction is a method to estimate the performance of Language Models (LMs) on various Natural Language Processing (NLP) tasks, mitigating computational costs associated with model capacity and data for fine-tuning. Our paper presents ProxyLM, a scalable task- and language-agnostic framework designed to predict the performance of LMs using proxy models. These proxy models act as surrogates, approximating the performance of the LM of interest. By leveraging these proxy models, ProxyLM significantly reduces computational overhead in task evaluations, achieving up to a 37.08x speedup over traditional methods, even with our smallest proxy models. Our results across multiple multilingual NLP tasks and various robustness tests demonstrate that ProxyLM not only adapts well to previously unseen languages in pre-trained LMs, but also generalizes effectively across different datasets, outperforming the state-of-the-art by at least 1.78x in terms of root-mean-square error (RMSE).

Extreme Multi-label Classification (XMC) methods predict relevant labels for a given query in an extremely large label space. Recent works in XMC address this problem using deep encoders that project text descriptions to an embedding space suitable for recovering the closest labels. However, learning deep models can be computationally expensive in large output spaces, resulting in a trade-off between high performing brute-force approaches and efficient solutions. In this paper, we propose PRIME, a XMC method that employs a novel prototypical contrastive learning technique to reconcile efficiency and performance surpassing brute-force approaches. We frame XMC as a data-to-prototype prediction task where label prototypes aggregate information from related queries. More precisely, we use a shallow transformer encoder that we coin as Label Prototype Network, which enriches label representations by aggregating text-based embeddings, label centroids and learnable free vectors. We jointly train a deep encoder and the Label Prototype Network using an adaptive triplet loss objective that better adapts to the high granularity and ambiguity of extreme label spaces. PRIME achieves state-of-the-art results in several public benchmarks of different sizes and domains, while keeping the model efficient.

Few-Shot Document-Level Relation Extraction (FSDLRE) aims to develop models capable of generalizing to new categories with minimal support examples. Although Large Language Models (LLMs) demonstrate exceptional In-Context Learning (ICL) capabilities on many few-shot tasks, their performance on FSDLRE tasks remains suboptimal due to the significant gap between the task format and the intrinsic capabilities of language models, coupled with the complexity of ICL prompts for document-level text. To address these challenges, we introduce a novel meta-training approach for LLMs termed Prototype Tuning. We construct simulated episodes using data with relation types that do not overlap with the test corpus, fundamentally enhancing the ICL capabilities of LLMs in FSDLRE through meta-learning. To further enhance the effects of meta-learning, we innovatively integrate the concept of prototype into the fine-tuning process of LLMs. This involves aggregating entity pairs from support documents into prototypes within the prompts and altering the way of determining relation categories to identifying the closest prototype. Experimental results demonstrate that our LLMs trained with this approach outperform all baselines. Our proposed approach markedly improves the ICL capabilities of LLMs in FSDLRE and mitigates the impact of relation semantic discrepancies between the training corpus and the test corpus on model performance.

Generative replay has proven effective in addressing the catastrophic forgetting issue of continual learning (CL) in natural language processing (NLP). However, relying on a single task-specific token or prompt often falls short in generating pseudo-samples that accurately reflect the true data distribution. This leads to issues of semantic inconsistency and scale inconsistency.To tackle these challenges, we propose a Prototype Conditioned Generative Replay (PCGR) method, which enhances generative reply by incorporating task-level statistics through a Prototype Conditioned Variational Autoencoder (PCVAE).Specifically, task-level embedding statistics are stored as prototypes for each old task. When a new task is introduced, PCVAE draws samples from task-specific prototype-based distributions to generate pseudo-samples.By incorporating the prototype, the generated pseudo-samples are both more representative and sufficiently diverse to reflect the real data distribution.Furthermore, as previously stored prototypes may become outdated due to evolving model parameters, we propose a Prototype Shift Estimation (PSE) to adjust for these changes.Experiments on NLP tasks across two different scenarios show that PCGR outperforms previous state-of-the-art (SOTA) methods.

Proteins play critical roles in biological systems, yet 99.7% of over 227 million known protein sequences remain uncharacterized due to the limitations of experimental methods. To assist experimentalists in narrowing down hypotheses and accelerating protein characterization, we present Protein2Text, a multimodal large language model that interprets protein sequences and generates informative text to address open-ended questions about protein functions and attributes. By integrating a resampling mechanism within an adapted LLaVA framework, our model effectively maps protein sequences into a language-compatible space, enhancing its capability to handle diverse and complex queries. Trained on a newly curated dataset derived from PubMed articles and rigorously evaluated using four comprehensive benchmarks—including in-domain and cross-domain evaluations—Protein2Text outperforms several existing models in open-ended question-answering tasks. Our work also highlights the limitations of current evaluation metrics applied to template-based approaches, which may lead to misleading results, emphasizing the need for unbiased assessment methods. Our model weights, evaluation datasets, and evaluation scripts are publicly available at https://github.com/alaaj27/Protein2Text.git.

Generative models such as Large Language Models (LLM) and Multimodal Large Language models (MLLMs) trained on massive web corpora can memorize and disclose individuals’ confidential and private data, raising legal and ethical concerns. While many previous works have addressed this issue in LLM via machine unlearning, it remains largely unexplored for MLLMs. To tackle this challenge, we introduce Multimodal Large Language Model Unlearning Benchmark (MLLMU-Bench), a novel benchmark aimed at advancing the understanding of multimodal machine unlearning. MLLMU-Bench consists of 500 fictitious profiles and 153 profiles for public celebrities, each profile feature over 14 customized question-answer pairs, evaluated from both multimodal (image+text) and unimodal (text) perspectives. The benchmark is divided into four sets to assess unlearning algorithms in terms of efficacy, generalizability, and model utility. Finally, we provide baseline results using existing generative model unlearning algorithms. Surprisingly, our experiments show that unimodal unlearning algorithms excel in generation tasks, while multimodal unlearning approaches perform better in classification with multimodal inputs.

Recent surge in Large Language Model (LLM) availability has opened exciting avenues for research. However, efficiently interacting with these models presents a significant hurdle since LLMs often reside on proprietary or self-hosted API endpoints, each requiring custom code for interaction. Conducting comparative studies between different models can therefore be time-consuming and necessitate significant engineering effort, hindering research efficiency and reproducibility. To address these challenges, we present prompto, an open source Python library which facilitates asynchronous querying of LLM endpoints enabling researchers to interact with multiple LLMs concurrently, while maximising efficiency and utilising individual rate limits. Our library empowers researchers and developers to interact with LLMs more effectively and allowing faster experimentation, data generation and evaluation. prompto is released with an introductory video (https://youtu.be/lWN9hXBOLyQ) under MIT License and is available via GitHub (https://github.com/alan-turing-institute/prompto).

Multilingual LLMs have achieved remarkable benchmark performance, but we find they continue to underperform on non-Latin script languages across contemporary LLM families. This discrepancy arises from the fact that LLMs are pretrained with orthographic scripts, which are dominated by Latin characters that obscure their shared phonology with non-Latin scripts. We propose leveraging phonemic transcriptions as complementary signals to induce script-invariant representations. Our study demonstrates that integrating phonemic signals improves performance across both non-Latin and Latin languages, with a particularly significant impact on closing the performance gap between the two. Through detailed experiments, we show that phonemic and orthographic scripts retrieve distinct examples for in-context learning (ICL). This motivates our proposed Mixed-ICL retrieval strategy, where further aggregation leads to our significant performance improvements for both Latin script languages (up to 12.6%) and non-Latin script languages (up to 15.1%) compared to randomized ICL retrieval.

Large Language Models (LLMs) have recently demonstrated impressive few-shot learning capabilities through in-context learning (ICL). However, ICL performance is highly dependent on the choice of few-shot demonstrations, making the selection of the most optimal examples a persistent research challenge. This issue is further amplified in low-resource Indic languages, where the scarcity of ground-truth data complicates the selection process. In this work, we propose PromptRefine, a novel Alternating Minimization approach for example selection that improves ICL performance on low-resource Indic languages. PromptRefine leverages auxiliary example banks from related high-resource Indic languages and employs multi-task learning techniques to align language-specific retrievers, enabling effective cross-language retrieval. Additionally, we incorporate diversity in the selected examples to enhance generalization and reduce bias. Through comprehensive evaluations on four text generation tasks—Cross-Lingual Question Answering, Multilingual Question Answering, Machine Translation, and Cross-Lingual Summarization using state-of-the-art LLMs such as LLAMA-3.1-8B, LLAMA-2-7B, Qwen-2-7B, and Qwen-2.5-7B, we demonstrate that PromptRefine significantly outperforms existing frameworks for retrieving examples.