NAACL 2025 论文列表

Large language models (LLMs) undergo extensive safety training to maximize both helpfulness and harmlessness in their responses. However, various jailbreak attacks jeopardize model safety, allowing malicious actors to bypass safety guidelines. Existing defense methods primarily focus on aligning the model’s output towards less harmful responses through post-processing or input perturbation. Consequently, these approaches are prone to general performance degradation and lack the ability to defend against a wide variety of attacks. In this paper, we propose goal-conditioned direct preference optimization (GC-DPO), which is trained to prioritize the system prompt over the user prompt through goal-conditioning, and thus enables a good balance between safety and performance. Empirically, we show that our approach significantly reduces the average Attack Success Rate (ASR) on a wide variety of jailbreak attacks. In particular, GC-DPO achieves a reduction of 67.1% to 5.0% in ASR for Vicuna-7B, a state-of-the-art result, without compromising the model’s general performance.

Uncovering hidden topics from short texts is challenging for traditional and neural models due to data sparsity, which limits word co-occurrence patterns, and label sparsity, stemming from incomplete reconstruction targets. Although data aggregation offers a potential solution, existing neural topic models often overlook it due to time complexity, poor aggregation quality, and difficulty in inferring topic proportions for individual documents. In this paper, we propose a novel model, **GloCOM** (**Glo**bal **C**lustering C**O**ntexts for Topic **M**odels), which addresses these challenges by constructing aggregated global clustering contexts for short documents, leveraging text embeddings from pre-trained language models. GloCOM can infer both global topic distributions for clustering contexts and local distributions for individual short texts. Additionally, the model incorporates these global contexts to augment the reconstruction loss, effectively handling the label sparsity issue. Extensive experiments on short text datasets show that our approach outperforms other state-of-the-art models in both topic quality and document representations.

Automatic Post-Editing (APE) systems often struggle with over-correction, where unnecessary modifications are made to a translation, diverging from the principle of minimal editing. In this paper, we propose a novel technique to mitigate over-correction by incorporating word-level Quality Estimation (QE) information during the decoding process. This method is architecture-agnostic, making it adaptable to any APE system, regardless of the underlying model or training approach. Our experiments on English-German, English-Hindi, and English-Marathi language pairs show the proposed approach yields significant improvements over their corresponding baseline APE systems, with TER gains of 0.65, 1.86, and 1.44 points, respectively. These results underscore the complementary relationship between QE and APE tasks and highlight the effectiveness of integrating QE information to reduce over-correction in APE systems.

Contrastive Language-Image Pre-training (CLIP) has become the standard for cross- modal image-text representation learning. Improving CLIP typically requires additional data and retraining with new loss functions, but these demands raise resource and time costs, limiting practical use. In this work, we introduce HELIP, a cost-effective strategy that improves CLIP models by exploiting challenging text-image pairs within existing datasets in continuous training. This eliminates the need for additional data or extensive retraining. Moreover, HELIP integrates effortlessly into current training pipelines with minimal code modifications, allowing for quick and seamless implementation. On comprehensive benchmarks, HELIP consistently boosts existing models. In particular, within just two epochs of training, it improves zero-shot classification accuracy on ImageNet for SLIP models pre-trained on CC3M, CC12M, and YFCC15M datasets by 3.05%, 4.47%, and 10.1% , respectively. In addition, on fine-grained classification datasets, HELIP improves the zero-shot performance of CLIP and SLIP by an average of 8.4% and 18.6%, and their linear probe performance by an average of 9.5% and 3.0%.

Geometry problem-solving demands advanced reasoning abilities to process multimodal inputs and employ mathematical knowledge effectively. Vision-language models (VLMs) have made significant progress in various multimodal tasks. Yet, they still struggle with geometry problems and are significantly limited by their inability to perform mathematical operations not seen during pre-training, such as calculating the cosine of an arbitrary angle, and by difficulties in correctly applying relevant geometry formulas. To overcome these challenges, we present GeoCoder, which leverages modular code-finetuning to generate and execute code using a predefined geometry function library. By executing the code, we achieve accurate and deterministic calculations, contrasting the stochastic nature of autoregressive token prediction, while the function library minimizes errors in formula usage. We also propose a multimodal retrieval-augmented variant of GeoCoder, named RAG-GeoCoder, which incorporates a non-parametric memory module for retrieving functions from the geometry library, thereby reducing reliance on parametric memory. Our modular code-finetuning approach enhances the geometric reasoning capabilities of VLMs, yielding an average improvement of over 16% across various question complexities on the GeomVerse dataset compared to other fine-tuning methods.

Prompts used in recent large language model based applications are often fixed and lengthy, leading to significant computational overhead. To address this challenge, we propose Generative Prompt Internalization (GenPI), a lightweight method that employs a joint training approach. GenPI not only replicates the behavior of models with prompt inputs but also generates the content of the prompt along with reasons for why the model’s behavior should change accordingly. We demonstrate that our approach effectively internalizes complex prompts across various agent-based application scenarios. For effective training without interactions with the dedicated environments, we introduce a data synthesis technique that autonomously collects conversational datasets by swapping the roles of the agent and environment. This method is especially useful in scenarios where only a predefined prompt is available without a corresponding training dataset. By internalizing complex prompts, Generative Prompt Internalization enables high performance and efficient inference without the need for explicit prompts.

Long-form story generation task aims to produce coherent and sufficiently lengthy text, essential for applications such as novel writingand interactive storytelling. However, existing methods, including LLMs, rely on rigid outlines or lack macro-level planning, making it difficult to achieve both contextual consistency and coherent plot development in long-form story generation. To address this issues, we propose Dynamic Hierarchical Outlining with Memory-Enhancement long-form story generation method, named DOME, to generate the long-form story with coherent content and plot. Specifically, the Dynamic Hierarchical Outline(DHO) mechanism incorporates the novel writing theory into outline planning and fuses the plan and writing stages together, improving the coherence of the plot by ensuring the plot completeness and adapting to the uncertainty during story generation. A Memory-Enhancement Module (MEM) based on temporal knowledge graphs is introduced to store and access the generated content, reducing contextual conflicts and improving story coherence. Finally, we propose a Temporal Conflict Analyzer leveraging temporal knowledge graphs to automatically evaluate the contextual consistency of long-form story. Experiments demonstrate that DOME significantly improves the fluency, coherence, and overall quality of generated long stories compared to state-of-the-art methods.

Inductive reasoning — the process of inferring general rules from a small number of observations — is a fundamental aspect of human intelligence. Recent works suggest that large language models (LLMs) can engage in inductive reasoning by sampling multiple hypotheses about the rules and selecting the one that best explains the observations. However, due to the IID sampling, semantically redundant hypotheses are frequently generated, leading to significant wastage of compute. In this paper, we 1) demonstrate that increasing the temperature to enhance the diversity is limited due to text degeneration issue, and 2) propose a novel method to improve the diversity while maintaining text quality. We first analyze the effect of increasing the temperature parameter, which is regarded as the LLM’s diversity control, on IID hypotheses. Our analysis shows that as temperature rises, diversity and accuracy of hypotheses increase up to a certain point, but this trend saturates due to text degeneration. To generate hypotheses that are more semantically diverse and of higher quality, we propose a novel approach inspired by human inductive reasoning, which we call Mixture of Concepts (MoC). When applied to several inductive reasoning benchmarks, MoC demonstrated significant performance improvements compared to standard IID sampling and other approaches.

Question decomposition has been found to help large language models’ (LLMs) performance on complex question answering (QA) by breaking these questions into simpler sub-questions for answering. Nonetheless, performance on the task remains dominated by supervised approaches, suggesting room for making LLMs better decomposers. One way of improving LLM training and fine-tuning is to leverage synthetic training data, but the superior performance of supervised approaches collapses in the face of distribution shifts, making them unsuitable for generating synthetic data across new domains and at scale. To address this, we propose an approach to generate synthetic decomposition data with only five annotated examples; we do this by (i) extending recent advancements in using LLM-as-judge and for reranking in novel ways, as well as (ii) using a panel of smaller-sized LLMs for data generation instead of resource-intensive larger models. Through careful validation of our approach over two benchmark datasets, we show that our data generation and modelling approaches bring consistent improvements over using few-shot prompting with LLMs for the task. Our code and models can be found at https://github.com/hankelvin/complex_question_decomposition.

Recent advancements in controllable expressive speech synthesis, especially in text-to-speech (TTS) models, have allowed for the generation of speech with specific styles guided by textual descriptions, known as style prompts. While this development enhances the flexibility and naturalness of synthesized speech, there remains a significant gap in understanding how these models handle vague or abstract style prompts. This study investigates the potential gender bias in how models interpret occupation-related prompts, specifically examining their responses to instructions like “Act like a nurse”. We explore whether these models exhibit tendencies to amplify gender stereotypes when interpreting such prompts. Our experimental results reveal the model’s tendency to exhibit gender bias for certain occupations. Moreover, models of different sizes show varying degrees of this bias across these occupations.

With the rapid advancement of large language models (LLMs), recent years have witnessed many promising studies on leveraging LLM-based agents to simulate human social behavior. While prior work has demonstrated significant potential across various domains, much of it has focused on specific scenarios involving a limited number of agents and has lacked the ability to adapt when errors occur during simulation. To overcome these limitations, we propose a novel LLM-agent-based simulation platform called GenSim, which: (1) Abstracts a set of general functions to simplify the simulation of customized social scenarios; (2) Supports one hundred thousand agents to better simulate large-scale populations in real-world contexts; (3) Incorporates error-correction mechanisms to ensure more reliable and long-term simulations. To evaluate our platform, we assess both the efficiency of large-scale agent simulations and the effectiveness of the error-correction mechanisms. To our knowledge, GenSim represents an initial step toward a general, large-scale, and correctable social simulation platform based on LLM agents, promising to further advance the field of social science.

Training-free embedding methods directly leverage pretrained large language models (LLMs) to embed text, bypassing the costly and complex procedure of contrastive learning. Previous training-free embedding methods have mainly focused on optimizing embedding prompts and have overlooked the benefits of utilizing the generative abilities of LLMs. We propose a novel method, GenEOL, which uses LLMs to generate diverse transformations of a sentence that preserve its meaning, and aggregates the resulting embeddings of these transformations to enhance the overall sentence embedding. GenEOL significantly outperforms the existing training-free embedding methods by an average of 2.85 points across several LLMs on the sentence semantic text similarity (STS) benchmark. GenEOL also achieves notable gains in clustering, reranking, and pair-classification tasks from the MTEB benchmark. Additionally, GenEOL stabilizes representation quality across LLM layers and remains robust to perturbations of embedding prompts.

The prevalence of toxic behavior in online gaming communities necessitates robust detection methods to ensure user safety. We introduce GameTox, a novel dataset comprising 53K game chat utterances annotated for toxicity detection through intent classification and slot filling. This dataset captures the complex relationship between user intent and specific linguistic features that contribute to toxic interactions. We extensively analyze the dataset to uncover key insights into the nature of toxic speech in gaming environments. Furthermore, we establish baseline performance metrics using state-of-the-art natural language processing and large language models, demonstrating the dataset’s contribution towards enhancing the detection of toxic behavior and revealing the limitations of contemporary models. Our results indicate that leveraging both intent detection and slot filling provides a significantly more granular and context-aware understanding of harmful messages. This dataset serves as a valuable resource to train advanced models that can effectively mitigate toxicity in online gaming and foster healthier digital spaces. Our dataset is publicly available at: https://github.com/shucoll/GameTox.

Refusal-Aware Instruction Tuning (RAIT) aims to enhance Large Language Models (LLMs) by improving their ability to refuse responses to questions beyond their knowledge, thereby reducing hallucinations and improving reliability. Effective RAIT must address two key challenges: firstly, effectively reject unknown questions to minimize hallucinations; secondly, avoid over-refusal to ensure questions that can be correctly answered are not rejected, thereby maintain the helpfulness of LLM outputs. In this paper, we address the two challenges by deriving insightful observations from the gradient-based perspective, and proposing the Gradient-driven Refusal Aware Instruction Tuning Framework GRAIT: (1) employs gradient-driven sample selection to effectively minimize hallucinations and (2) introduces an adaptive weighting mechanism during fine-tuning to reduce the risk of over-refusal, achieving the balance between accurate refusals and maintaining useful responses. Experimental evaluations on open-ended and multiple-choice question answering tasks demonstrate that GRAIT significantly outperforms existing RAIT methods in the overall performance. The source code and data will be available at https://github.com/opendatalab/GRAIT .

Naive Retrieval-Augmented Generation (RAG) focuses on individual documents during retrieval and, as a result, falls short in handling networked documents which are very popular in many applications such as citation graphs, social media, and knowledge graphs. To overcome this limitation, we introduce Graph Retrieval-Augmented Generation (GRAG), which tackles the fundamental challenges in retrieving textual subgraphs and integrating the joint textual and topological information into Large Language Models (LLMs) to enhance its generation. To enable efficient textual subgraph retrieval, we propose a novel divide-and-conquer strategy that retrieves the optimal subgraph structure in linear time. To achieve graph context-aware generation, incorporate textual graphs into LLMs through two complementary views—the text view and the graph view—enabling LLMs to more effectively comprehend and utilize the graph context. Extensive experiments on graph reasoning benchmarks demonstrate that in scenarios requiring multi-hop reasoning on textual graphs, our GRAG approach significantly outperforms current state-of-the-art RAG methods. Our datasets as well as codes of GRAG are available at https://github.com/HuieL/GRAG.

Despite the fact that large-scale Language Models (LLM) have achieved SOTA performances on a variety of NLP tasks, its performance on NER is still significantly below supervised baselines. This is due to the gap between the two tasks the NER and LLMs: the former is a sequence labeling task in nature while the latter is a text-generation model.In this paper, we propose GPT-NER to resolve this issue. GPT-NER bridges the gap by transforming the sequence labeling task to a generation task that can be easily adapted by LLMs e.g., the task of finding location entities in the input text “Columbus is a city” is transformed to generate the text sequence "@@Columbus## is a city”, where special tokens @@## marks the entity to extract. To efficiently address the hallucination issue of LLMs, where LLMs have a strong inclination to over-confidently label NULL inputs as entities, we propose a self-verification strategy by prompting LLMs to ask itself whether the extracted entities belong to a labeled entity tag.We conduct experiments on five widely adopted NER datasets, and GPT-NER achieves comparable performances to fully supervised baselines, which is the first time as far as we are concerned. More importantly, we find that GPT-NER exhibits a greater ability in the low-resource and few-shot setups, when the amount of training data is extremely scarce, GPT-NER performs significantly better than supervised models. This demonstrates the capabilities of GPT-NER in real-world NER applications where the number of labeled examples is limited.

GPT-4’s purported strong multimodal abilities raise interests in using it to automate radiology report writing, but there lacks thorough evaluations. In this work, we perform a systematic evaluation of GPT-4 (4o and vision-preview) in generating radiology reports across three chest X-ray report benchmarks: MIMIC-CXR, CheXpert Plus, and IU X-Ray. We attempt to directly generate reports with different prompting strategies and find that the models fail terribly in both lexical metrics and clinical efficacy metrics. To understand the low performance, we decompose the task into two steps: 1) the **medical image reasoning** step of predicting medical condition labels from images; and 2) the **report synthesis** step of generating reports from (groundtruth) conditions. We show that GPT-4’s performance in image reasoning is consistently low across different prompts. In fact, the distributions of model-predicted labels remain constant regardless of which groundtruth conditions are present on the image, suggesting that the model is not interpreting chest X-rays meaningfully. Even when given groundtruth conditions in report synthesis, its generated reports are less correct and less natural-sounding than a finetuned Llama. Altogether, our findings cast doubt on the viability of using GPT-4 in a radiology workflow.

We introduce GLiREL, an efficient architecture and training paradigm for zero-shot relation classification. Identifying relationships between entities is a key task in information extraction pipelines. The zero-shot setting for relation extraction, where a taxonomy of relations is not pre-specified, has proven to be particularly challenging because of the computational complexity of inference, and because of the lack of labeled training data with sufficient coverage. Existing approaches rely upon distant supervision using auxiliary models to generate training data for unseen labels, upon very large general-purpose large language models (LLMs), or upon complex pipelines models with multiple inference stages. Inspired by the recent advancements in zero-shot named entity recognition, this paper introduces an approach to efficiently and accurately predict zero-shot relationship labels between multiple entities in a single forward pass. Experiments using the FewRel and WikiZSL benchmarks demonstrate that our approach achieves state-of-the-art results on the zero-shot relation classification task. In addition, we contribute a protocol for synthetically-generating datasets with diverse relation labels.

Illiteracy is a predictor of many negative social and personal outcomes. Illiteracy rates are particularly high in countries with underresourced languages, where few books exist that are suitable for children to learn to read from. We present GAIfE (Generative AI for Education), a toolchain and workflow developed through empirical methods, that demonstrates how existing tools can be adapted to address low literacy for an underresourced language. We used GAIfE (a play on the Bambara word for “book”) to construct materials for developing children’s reading competence in Bambara, the vehicular language of Mali. Our approach to the generation and post-generation editing of content skewed by the Global-North-centric bias of available LLMs, enabled us to rapidly multiply the content in Bambara available online by 10 times while maintaining high standards of attractiveness of the material to maintain high engagement, accurate representation of the Malian culture and physical and social environment and language quality. Using our materials, pilot reading programs achieved a 67% reduction in the number of children unable to read Bambara. Our approach demonstrated the power of bias-aware application of generative AI to the problem domain as well as the potential impact the application of this technology could have on reducing illiteracy and improving learning outcomes through native language education.