NAACL 2025 Papers

Despite the unprecedented success of applying Large Language Models (LLMs) to graph discriminative tasks such as node classification and link prediction, its potential for graph structure generation remains largely unexplored. To fill this crucial gap, this paper presents a systematic investigation into the capability of LLMs for graph structure generation. Specifically, we design prompts triggering LLMs to generate codes that optimize network properties by injecting domain expertise from network science. Since graphs in different domains exhibit unique structural properties captured by various metrics (e.g., clustering coefficient capturing triangles in social networks while squares reflecting road segments in transportation networks), we first evaluate the capability of LLMs to generate graphs satisfying each structural property in different domains. After that, we select the optimal property configurations and benchmark the graph structure generation performance of LLMs against established graph generative models across multiple domains. Our findings shed light on generating graph structures from an LLM perspective. Our code is publically available https://github.com/yuwvandy/LLM-GraphGen.

Large Language Models (LLMs) have emerged as a dominant approach for a wide range of NLP tasks, with their access to external information further enhancing their capabilities. However, this introduces new vulnerabilities, known as prompt injection attacks, where external content embeds malicious instructions that manipulate the LLM’s output. Recently, the Base64 defense has been recognized as one of the most effective methods for reducing success rate of prompt injection attacks. Despite its efficacy, this method can degrade LLM performance on certain NLP tasks. To address this challenge, we propose a novel defense mechanism: mixture of encodings, which utilizes multiple character encodings, including Base64. Extensive experimental results show that our method achieves one of the lowest attack success rates under prompt injection attacks, while maintaining high performance across all NLP tasks, outperforming existing character encoding-based defense methods. This underscores the effectiveness of our mixture of encodings strategy for both safety and task performance metrics.

Fact-checking pipelines increasingly adopt the Decompose-Then-Verify paradigm, where texts are broken down into smaller claims for individual verification and subsequently combined for a veracity decision. While decomposition is widely-adopted in such pipelines, its effects on final fact-checking performance remain underexplored. Some studies have reported improvements from decompostition, while others have observed performance declines, indicating its inconsistent impact. To date, no comprehensive analysis has been conducted to understand this variability. To address this gap, we present an in-depth analysis that explicitly examines the impact of decomposition on downstream verification performance. Through error case inspection and experiments, we introduce a categorization of decomposition errors and reveal a trade-off between accuracy gains and the noise introduced through decomposition. Our analysis provides new insights into understanding current system’s instability and offers guidance for future studies toward improving claim decomposition in fact-checking pipelines.

Speculative Decoding is a widely used technique to speed up inference for Large Language Models (LLMs) without sacrificing quality. When performing inference, speculative decoding uses a smaller draft model to generate speculative tokens and then uses the target LLM to verify those draft tokens. The speedup provided by speculative decoding heavily depends on the choice of the draft model. In this work, we perform a detailed study comprising over 350 experiments with LLaMA-65B and OPT-66B using speculative decoding and delineate the factors that affect the performance gain provided by speculative decoding. Our experiments indicate that the performance of speculative decoding depends heavily on the latency of the draft model, and the draft model’s capability in language modeling does not correlate strongly with its performance in speculative decoding. Based on these insights we explore a new design space for draft models and design hardware-efficient draft models for speculative decoding. Our newly designed draft model can provide 111% higher throughput than existing draft models and our approach generalizes further to all LLaMA models (1/2/3.1) and supervised fine-tuned models.

Hate speech is a harmful form of online expression, often manifesting as derogatory posts. It is a significant risk in digital environments. With the rise of Large Language Models (LLMs), there is concern about their potential to replicate hate speech patterns, given their training on vast amounts of unmoderated internet data. Understanding how LLMs respond to hate speech is crucial for their responsible deployment. However, the behaviour of LLMs towards hate speech has been limited compared. This paper investigates the reactions of seven state-of-the-art LLMs (LLaMA 2, Vicuna, LLaMA 3, Mistral, GPT-3.5, GPT-4, and Gemini Pro) to hate speech. Through qualitative analysis, we aim to reveal the spectrum of responses these models produce, highlighting their capacity to handle hate speech inputs. We also discuss strategies to mitigate hate speech generation by LLMs, particularly through fine-tuning and guideline guardrailing. Finally, we explore the models’ responses to hate speech framed in politically correct language.

Existing studies have shown that AI-generated images tend to reinforce social biases, including those related to race and gender. However, no studies have investigated weight bias, or fatphobia, in AI-generated images. This study utilizes DALL-E 3 to determine the extent to which anti-fat and pro-thin biases are present in AI-generated images, and examines stereotypical associations between moral character and body weight. Four-thousand images are generated using twenty pairs of positive and negative textual prompts. These images are then manually labeled with weight information and analyzed to determine the extent to which they reflect fatphobia. The findings and their impact are discussed and related to existing research on weight bias.

Understanding and mitigating the potential risks associated with foundation models (FMs) hinges on developing effective interpretability methods. Sparse Autoencoders (SAEs) have emerged as a promising tool for disentangling FM representations, but they struggle to capture rare, yet crucial concepts in the data. We introduce Specialized Sparse Autoencoders (SSAEs), designed to illuminate these elusive dark matter features by focusing on specific subdomains. We present a practical recipe for training SSAEs, demonstrating the efficacy of dense retrieval for data selection and the benefits of Tilted Empirical Risk Minimization as a training objective to improve concept recall. Our evaluation of SSAEs on standard metrics, such as downstream perplexity and L0 sparsity, show that they effectively capture subdomain tail concepts, exceeding the capabilities of general-purpose SAEs. We showcase the practical utility of SSAEs in a case study on the Bias in Bios dataset, where SSAEs achieve a 12.5% increase in worst-group classification accuracy over the pretrained general-purpose SAE when applied to remove spurious gender information. SSAEs provide a powerful new lens for peering into the inner workings of FMs in subdomains.

The use of AI in legal analysis and prediction (LegalAI) has gained attention, with past research focusing on retrieval-based methods and fine-tuning large models. However, these approaches often require large datasets and underutilize the capabilities of modern large language models (LLMs). In this paper, inspired by the debate phase of real courtroom trials, we propose a novel legal judgment prediction model based on the Debate-Feedback architecture, which integrates LLM multi-agent debate and reliability evaluation models. Unlike traditional methods, our model achieves significant improvements in efficiency by minimizing the need for large historical datasets, thus offering a lightweight yet robust solution. Comparative experiments show that it outperforms several general-purpose and domain-specific legal models, offering a dynamic reasoning process and a promising direction for future LegalAI research.

While Large Language Models (LLMs) excel in zero-shot Question Answering (QA), they tend to expose biases in their internal knowledge when faced with socially sensitive questions, leading to a degradation in performance. Existing zero-shot methods are efficient but failto consider context and prevent bias propagation in the answers. To address this, we propose *DeCAP*, a method for debiasing LLMs usingContext-Adaptive Prompt Generation. *DeCAP* leverages a *Question Ambiguity Detection* to take appropriate debiasing actions based on the context and a *Neutral Answer Guidance Generation* to suppress the LLMs make objective judgments about the context, minimizing thepropagation of bias from their internal knowledge. Our various experiments across eight LLMs show that *DeCAP* achieves state-of-the-art zero-shot debiased QA performance. This demonstrates *DeCAP*’s efficacy in enhancing the fairness and accuracy of LLMs in diverseQA settings.

To address the challenges associated with data processing at scale, we propose Dataverse, a unified open-source Extract-Transform-Load (ETL) pipeline for large language models (LLMs) with a user-friendly design at its core. Easy addition of custom processors with block-based interface in Dataverse allows users to readily and efficiently use Dataverse to build their own ETL pipeline. We hope that Dataverse will serve as a vital tool for LLM development and open source the entire library to welcome community contribution. Additionally, we provide a concise, two-minute video demonstration of our system, illustrating its capabilities and implementation.

Backdoor attack is a severe threat to the trustworthiness of DNN-based language models. In this paper, we first extend the definition of memorization of language models from sample-wise to more fine-grained sentence element-wise (e.g., word, phrase, structure, and style), and then point out that language model backdoors are a type of element-wise memorization. Through further analysis, we find that the strength of such memorization is positively correlated to the frequency of duplicated elements in the training dataset. In conclusion, duplicated sentence elements are necessary for successful backdoor attacks. Based on this, we propose a data-centric defense. We first detect trigger candidates in training data by finding memorizable elements, i.e., duplicated elements, and then confirm real triggers by testing if the candidates can activate backdoor behaviors (i.e., malicious elements). Results show that our method outperforms state-of-the-art defenses in defending against different types of NLP backdoors.

3D scene understanding has gained significant attention due to its wide range of applications. However, existing methods for 3D scene understanding are limited to specific downstream tasks, which hinders their practicality in real-world applications. This paper presents Chat-3D, which combines the 3D visual perceptual ability of pre-trained 3D representations and the impressive reasoning and conversation capabilities of advanced LLMs to achieve the first universal dialogue systems for 3D scenes. Specifically, we align 3D representations into the feature space of LLMs, thus enabling LLMs to perceive the 3D world. Given the scarcity of 3D scene-text data, we propose a three-stage training strategy to efficiently utilize the available data for better alignment. To enhance the reasoning ability and develop a user-friendly interaction scheme, we further construct a high-quality object-centric 3D instruction dataset and design an associated object-centric prompt. With limited data, Chat-3D achieves a 82.2% relative score compared with GPT-4 on the constructed instruction dataset, and comparable performance to state-of-the-art LLM-based methods.

In-context learning (ICL) has emerged as a capability of large language models (LLMs), enabling them to adapt to new tasks using provided examples. While ICL has demonstrated its strong effectiveness, there is limited understanding of its vulnerability against potential threats. This paper examines ICL’s vulnerability to data poisoning attacks. We introduce ICLPoison, an attacking method specially designed to exploit ICL’s unique learning mechanisms by identifying discrete text perturbations that influence LLM hidden states. We propose three representative attack strategies, evaluated across various models and tasks. Our experiments, including those on GPT-4, show that ICL performance can be significantly compromised by these attacks, highlighting the urgent need for improved defense mechanisms to protect LLMs’ integrity and reliability.

Current intent detection work experiments with minor intent categories. However, in real-world scenarios of data analysis dialogue systems, intents are composed of combinations of numerous metrics and dimensions, resulting in countless intents and posing challenges for the language model. The retrieval-augmented generation (RAG) method efficiently retrieves key intents. However, the single retrieval route sometimes fails to recall target intents and causes incorrect results. To alleviate the above challenges, we introduce the DSRAG framework combining query-to-query (Q2Q) and query-to-metadata (Q2M) double-stream RAG approaches. Specifically, we build a repository of query statements for Q2Q using the query templates with the key intents. When a user’s query comes, it rapidly matches repository statements. Once the relevant query is retrieved, the results can be quickly returned. In contrast, Q2M retrieves the relevant intents from the metadata and utilizes large language models to choose the answer. Experimental results show that DSRAG achieves significant improvements compared with merely using prompt engineering and a single retrieval route.

Synthetic data has recently proven effective in enhancing the accuracy of Text-to-SQL parsers. However, existing methods generate SQL queries first by randomly sampling tables and columns based on probability and then synthesize natural language questions (NLQs). This approach often produces a large number of NLQ-SQL pairs that are irrelevant to the target domain and inconsistent in query intent, significantly diminishing the fine-tuning effectiveness of LLMs. In this paper, we introduce DSQG-Syn, a novel text-to-SQL data synthesis framework that based on domain-specific question generation. Specifically, we design a question generation method that creates domain-relevant questions based on predefined question types, ensuring coverage of major SQL operations. Guided by these questions, we synthesize NLQ-SQL pairs that are both domain-relevant and intent-consistent. To further enhance data quality, we filter out noisy samples from the generated pairs. When popular open-source LLMs are fine-tuned on our high-quality synthesized dataset, they achieve significant accuracy improvements, surpassing the performance of closed-source LLM-based approaches. Moreover, we demonstrate that our method outperforms existing state-of-the-art (SOTA) data synthesis techniques.

Recent progress in video-text retrieval has been driven largely by advancements in model architectures and training strategies. However, the representation learning capabilities of video-text retrieval models remain constrained by low-quality and limited training data annotations. To address this issue, we present a novel Video-Text Retrieval Paradigm with Relevance-based Augmentation, namely dReAm, which enhances video and text data using large foundation models to learn more generalized features. Specifically, we first adopt a simple augmentation method, which generates self-similar data by randomly duplicating or dropping subwords and frames. In addition, inspired by the recent advancement in visual and language generative models, we propose a more robust augmentation method through textual paraphrasing and video stylization using large language models (LLMs) and visual generative models (VGMs). To further enrich video and text information, we propose a relevance-based augmentation method, where LLMs and VGMs generate and integrate new relevant information into the original data. Leveraging this enriched data, extensive experiments on several video-text retrieval benchmarks demonstrate the superiority of dReAm over existing methods. Code will be available upon acceptance.

Multimodal Large Language Models (MLLMs) pose unique safety challenges due to their integration of visual and textual data, thereby introducing new dimensions of potential attacks and complex risk combinations. In this paper, we begin with a detailed analysis aimed at disentangling risks through step-by-step reasoning within multimodal inputs. We find that systematic multimodal risk disentanglement substantially enhances the risk awareness of MLLMs. Via leveraging the strong discriminative abilities of multimodal risk disentanglement, we further introduce DREAM ( Disentangling Risks to Enhance Safety Alignment in MLLMs), a novel approach that enhances safety alignment in MLLMs through supervised fine-tuning and iterative Reinforcement Learning from AI Feedback (RLAIF). Experimental results show that DREAM significantly boosts safety during both inference and training phases without compromising performance on normal tasks (namely oversafety), achieving a 16.17% improvement in the SIUO safe&effective score compared to GPT-4V.

In direct preference alignment in LLMs, most existing methods seek to retrieve the reward function directly from preference data. However, real-world preference data often contains diversity in preference annotations reflective of true human preferences. Existing algorithms, including KTO, do not directly utilize such nuances in the annotations which limits their applicability. In this work, we propose Diverse Preference Learning (DPL), a reference model-free method that simultaneously learns a baseline desirability in LLM responses while being robust to the diversity of preference annotations. Our experiments for instruction-following on Ultrafeedback and AlpacaEval 2.0 and for text-summarization on Reddit TL;DR suggest that DPL is consistently better at learning the diversity of preferences compared to existing methods, including those that require a reference model in memory. Apart from overall quality, we find that DPL’s completions, on average, are more honest, helpful, truthful and safe compared to existing methods.

Despite the strong research interest in document-level Machine Translation (MT), the test-sets dedicated to this task are still scarce. The existing test-sets mainly cover topics from the general domain and fall short on specialised domains, such as legal and financial. Also, despite their document-level aspect, they still follow a sentence-level logic that doesn’t allow for including certain linguistic phenomena such as information reorganisation. In this work, we aim to fill this gap by proposing a novel test-set : DOLFIN. The dataset is built from specialised financial documents and it makes a step towards true document-level MT by abandoning the paradigm of perfectly aligned sentences, presenting data in units of sections rather than sentences. The test-set consists of an average of 1950 aligned sections for five language pairs. We present the detailed data collection pipeline that can serve as inspiration for aligning new document-level datasets. We demonstrate the usefulness and the quality of this test-set with the evaluation of a series of models. Our results show that the test-set is able to discriminate between context-sensitive and context-agnostic models and shows the weaknesses when models fail to accurately translate financial texts. The test-set will be made public for the community.