NAACL 2025 论文列表

Books, while often rich in cultural insights, can also mirror societal biases of their eras—biases that Large Language Models (LLMs) may learn and perpetuate during training. We introduce a novel method to trace and quantify these biases using fine-tuned LLMs. We develop BookPAGE, a corpus comprising 593 fictional books across seven decades (1950-2019), to track bias evolution. By fine-tuning LLMs on books from each decade and using targeted prompts, we examine shifts in biases related to gender, sexual orientation, race, and religion. Our findings indicate that LLMs trained on decade-specific books manifest biases reflective of their times, with both gradual trends and notable shifts. For example, model responses showed a progressive increase in the portrayal of women in leadership roles (from 8% to 22%) from the 1950s to 2010s, with a significant uptick in the 1990s (from 4% to 12%), possibly aligning with third-wave feminism. Same-sex relationship references increased markedly from the 1980s to 2000s (from 0% to 10%), mirroring growing LGBTQ+ visibility. Concerningly, negative portrayals of Islam rose sharply in the 2000s (26% to 38%), likely reflecting post-9/11 sentiments. Importantly, we demonstrate that these biases stem mainly from the books’ content and not the models’ architecture or initial training. Our study offers a new perspective on societal bias trends by bridging AI, literary studies, and social science research.

The spread of harmful content online is a dynamic issue evolving over time. Existing detection models, reliant on static data, are becoming less effective and generalizable. Developing new models requires sufficient up-to-date data, which is challenging. A potential solution is to combine existing datasets with minimal new data. However, detection tasks vary—some focus on hate speech, offensive, or abusive content, which differ in the intent to harm, while others focus on identifying targets of harmful speech such as racism, sexism, etc.—raising the challenge of handling nuanced class differences. To address these issues, we introduce a novel transfer learning method that leverages class-specific knowledge to enhance harmful content detection. In our approach, we first present label-specific soft prompt tuning, which captures and represents class-level information. Secondly, we propose two approaches to transfer this fine-grained knowledge from source (existing tasks) to target (unseen and new tasks): initializing the target task prompts from source prompts and using an attention mechanism that learns and adjusts attention scores to utilize the most relevant information from source prompts. Experiments demonstrate significant improvements in harmful content detection across English and German datasets, highlighting the effectiveness of label-specific representations and knowledge transfer.

This study addresses two key challenges in structuring radiology reports: the lack of a practical structuring schema and datasets to evaluate model generalizability. To address these challenges, we propose a “Finding-Centric Structuring,” which organizes reports around individual findings, facilitating secondary use. We also construct JRadFCS, a large-scale dataset with annotated named entities (NEs) and relations, comprising 8,428 Japanese Computed Tomography (CT) reports from seven facilities, providing a comprehensive resource for evaluating model generalizability. Our experiments reveal performance gaps when applying models trained on single-facility reports to those from other facilities. We further analyze factors contributing to these gaps and demonstrate that augmenting the training set based on these performance-correlated factors can efficiently enhance model generalizability.

Through numerous endeavors, large language models (LLMs) have witnessed significant advancements in their instruction-following capability. However, we discern that LLMs are prone to generate responses to instruction-formatted statements in an instinctive manner, rather than comprehending the underlying user intention reside within the given instructions. We also recognize that the significance of instruction understanding capability is largely overlooked in most of LLM evaluation benchmarks. To ensure more comprehensive evaluation on the instruction understanding capability of LLM, we propose Intention of Instruction (IntInst) benchmark, which primary objective is to distinguish the appropriate instruction that accurately instruct to generate a given context. IntInst presents four instruction candidates and requires LLMs to select one among them. Through extensive experiments with several instruction-tuned LLMs, we reveal that most LLMs struggle to grasp the actual intention concealed in the instruction and thoroughly analyze the factors influencing instruction understanding.

We introduce FinNLI, a benchmark dataset for Financial Natural Language Inference (FinNLI) across diverse financial texts like SEC Filings, Annual Reports, and Earnings Call transcripts. Our dataset framework ensures diverse premise-hypothesis pairs while minimizing spurious correlations. FinNLI comprises 21,304 pairs, including a high-quality test set of 3,304 instances annotated by finance experts. Evaluations show that domain shift significantly degrades general-domain NLI performance. The highest Macro F1 scores for pre-trained (PLMs) and large language models (LLMs) baselines are 74.57% and 78.62%, respectively, highlighting the dataset’s difficulty. Surprisingly, instruction-tuned financial LLMs perform poorly, suggesting limited generalizability. FinNLI exposes weaknesses in current LLMs for financial reasoning, indicating room for improvement.

Trading range breakout is a key method in the technical analysis of financial trading, widely employed by traders in financial markets such as stocks, futures, and foreign exchange. However, distinguishing between true and false breakout and providing the correct rationale cause significant challenges to investors. Traditional quantitative methods require large amounts of data and cannot directly present the reasoning process to users, making them less than perfect in this field. Recently, large language models have achieved success in various downstream applications, but their effectiveness in the domain of financial breakout detection has been subpar. The reason is that the unique data and specific knowledge are required in breakout detection. To address these issues, we created the first financial breakout dataset and introduce FinLLM-B, the premier large language model for financial breakout detection, which enhances the effectiveness of breakout trading strategies. Furthermore, we have developed a novel framework for large language models, namely multi-stage structure, effectively reducing mistakes in downstream applications. Experimental results indicate that compared to GPT-3.5, FinanceGPT-B improves the average accuracy of answers and rational by 49.97%, with the multi-stage structure contributing 9.72% to the improvement. Additionally, it outperforms ChatGPT-4 by 42.38%.

Large language models have demonstrated outstanding performance in various natural language processing tasks, but their security capabilities in the financial domain have not been explored, and their performance on complex tasks like financial agent remains unknown. This paper presents FinEval, a benchmark designed to evaluate LLMs’ financial domain knowledge and practical abilities. The dataset contains 8,351 questions categorized into four different key areas: Financial Academic Knowledge, Financial Industry Knowledge, Financial Security Knowledge, and Financial Agent. Financial Academic Knowledge comprises 4,661 multiple-choice questions spanning 34 subjects such as finance and economics. Financial Industry Knowledge contains 1,434 questions covering practical scenarios like investment research. Financial Security Knowledge assesses models through 1,640 questions on topics like application security and cryptography. Financial Agent evaluates tool usage and complex reasoning with 616 questions. FinEval has multiple evaluation settings, including zero-shot, five-shot with chain-of-thought, and assesses model performance using objective and subjective criteria. Our results show that Claude 3.5-Sonnet achieves the highest weighted average score of 72.9 across all financial domain categories under zero-shot setting. Our work provides a comprehensive benchmark closely aligned with Chinese financial domain. The data and the code are available at https://github.com/SUFE-AIFLMLab/FinEval.

In text classification tasks, models often rely on spurious correlations for predictions, incorrectly associating irrelevant features with the target labels. This issue limits the robustness and generalization of models, especially when faced with out-of-distribution data where such spurious correlations no longer hold. To address this challenge, we propose the Causally Calibrated Robust Classifier (CCR), which aims to reduce models’ reliance on spurious correlations and improve model robustness. Our approach integrates a causal feature selection method based on counterfactual reasoning, along with an unbiased inverse propensity weighting (IPW) loss function. By focusing on selecting causal features, we ensure that the model relies less on spurious features during prediction. We theoretically justify our approach and empirically show that CCR achieves state-of-the-art performance among methods without group labels, and in some cases, it can compete with the models that utilize group labels. Our code can be found at: https://github.com/yuqing-zhou/Causal-Learning-For-Robust-Classifier.

Data is the lifeblood of large language models (LLMs). While the quantity of open-source data available for training LLMs is substantial, its integrity often falls short. For instance, the open-source chat version of Yi-1.5-9B scores 5.20 on AlignBench, while the Chinese Alpaca-GPT4 version scores 4.12. This discrepancy makes it challenging for developers to create models that excel in downstream tasks and instruction following. Therefore, it is essential to improve data integrity. Currently, there are two mainstream methods for enhancing data integrity: data filtering and data augmentation. Due to the labor-intensive and time-consuming nature of performing these tasks manually, some of these efforts are now being undertaken by LLMs, owing to their high alignment with human preferences. However, we have found that performing data filtering or data augmentation with LLMs has limited effectiveness in improving data integrity. In this work, we propose FiNE (Filtering and Improving Noisy data Elaborately), a method that performs refined filtering and improvement of training data with LLMs. Using the data obtained through our method to train Yi-1.5-9B, the performance gap on AlignBench between our model and the open-source chat version is reduced from 1.08 to 0.35. Additionally, on HalluQA, our model surpasses the open-source chat version by 8.45.

As the diversity of users increases, the capability of providing personalized responses by large language models (LLMs) has become increasingly important. Existing approaches have only limited successes in LLM personalization, due to the absence of personalized learning or the reliance on shared personal data. This paper proposes a new approach for a few-shot personalization of LLMs with their mis-aligned responses (Fermi). Our key idea is to learn a set of personalized prompts for each user by progressively improving the prompts using LLMs, based on user profile (e.g., demographic information) and a few examples of previous opinions. During an iterative process of prompt improvement, we incorporate the contexts of mis-aligned responses by LLMs, which are especially crucial for the effective personalization of LLMs. In addition, we develop an effective inference method to further leverage the context of the test query and the personalized prompts. Our experimental results demonstrate that Fermi significantly improves performance across various benchmarks, compared to best-performing baselines.

Translation of natural language to first-order logical formula (NL-FOL) has recently gained significant attention for its critical role in logic-based NLP applications. Some studies attempt to utilize pretrained language models in a sequence-to-sequence manner for the NL-FOL task. However, these methods encounter challenges such as (1) inconsistency between the training and inference phases and (2) the data-intensive and resource-intensive finetuning process. This paper introduces a novel NL-FOL translation method, dubbed Code4Logic, which is based on in-context learning and employs code snippets to bridge the gap between natural language and first-order logic. By converting the translation task into a progressive code generation task, Code4Logic demonstrates strong generalization within a training-free manner, and enhances the performance of large language models (LLMs) to generate complex first-order logical formulas. Experimental results on NL-FOL task and downstream task datasets indicate that Code4Logic surpasses prominent training-free baselines and is comparable to supervised models trained on the full training data.

Large Language Models (LLMs) achieve state-of-the-art performance but are challenging to deploy due to their high computational and storage demands. Pruning can reduce model size, yet existing methods assume public access to calibration data, which is impractical for privacy-sensitive applications. To address the challenge of pruning LLMs in privacy-preserving settings, we propose FedSpaLLM, the first federated learning framework designed specifically for pruning LLMs. FedSpaLLM enables clients to locally prune their models based on private data while accounting for system heterogeneity and maintaining communication efficiency. Our framework introduces several key innovations: (1) a novel ℓ0-norm aggregation function that ensures only non-zero weights are averaged across clients, preserving important model parameters; (2) an adaptive mask expansion technique that meets global sparsity targets while accommodating client-specific pruning decisions; and (3) a layer sampling strategy that reduces communication overhead and personalizes the pruning process based on client resources. Extensive experiments show that FedSpaLLM improves pruning performance in diverse federated settings.

Sparse Autoencoders (SAEs) are a promising approach for extracting neural network representations by learning a sparse and overcomplete decomposition of the network’s internal activations. However, SAEs are traditionally trained considering only activation values and not the effect those activations have on downstream computations. This limits the information available to learn features, and biases the autoencoder towards neglecting features which are represented with small activation values but strongly influence model outputs.To address this, we introduce Gradient SAEs (g-SAEs), which modify the k-sparse autoencoder architecture by augmenting the TopK activation function to rely on the gradients of the input activation when selecting the k elements. For a given sparsity level, g-SAEs produce reconstructions that are more faithful to original network performance when propagated through the network.Additionally, we find evidence that g-SAEs learn latents that are on average more effective at steering models in arbitrary contexts.By considering the downstream effects of activations, our approach leverages the dual nature of neural network features as both representations, retrospectively, and actions, prospectively. While previous methods have approached the problem of feature discovery primarily focused on the former aspect, g-SAEs represent a step towards accounting for the latter as well.

Although the multilingual capability of LLMs offers new opportunities to overcome the language barrier, do these capabilities translate into real-life scenarios where linguistic divide and knowledge conflicts between multilingual sources are known occurrences? In this paper, we studied LLM’s linguistic preference in a cross-language RAG-based information search setting. We found that LLMs displayed systemic bias towards information in the same language as the query language in both document retrieval and answer generation. Furthermore, in scenarios where no information is in the language of the query, LLMs prefer documents in high-resource languages during generation, potentially reinforcing the dominant views. Such bias exists for both factual and opinion-based queries. Our results highlight the linguistic divide within multilingual LLMs in information search systems. The seemingly beneficial multilingual capability of LLMs may backfire on information parity by reinforcing language-specific filter bubbles further marginalizing low-resource views.

Ensembling neural machine translation (NMT) models to produce higher-quality translations than the L individual models has been extensively studied. Recent methods typically employ a candidate selection block (CSB) and an encoder-decoder fusion block (FB), requiring inference across all candidate models, leading to significant computational overhead, generally 𝛺(L). This paper introduces SmartGen, a reinforcement learning (RL)-based strategy that improves the CSB by selecting a small, fixed number of candidates and identifying optimal groups to pass to the fusion block for each input sentence. Furthermore, previously, the CSB and FB were trained independently, leading to suboptimal NMT performance. Our DQN-based SmartGen addresses this by using feedback from the FB block as a reward during training. We also resolve a key issue in earlier methods, where candidates were passed to the FB without modification, by introducing a Competitive Correction Block (CCB). Finally, we validate our approach with extensive experiments on English-Hindi translation tasks in both directions as well as English to Chinese and English to German.

Zero-shot named entity recognition (NER) is the task of detecting named entities of specific types (such as Person or Medicine) without any training examples. Current research increasingly relies on large synthetic datasets, automatically generated to cover tens of thousands of distinct entity types, to train zero-shot NER models. However, in this paper, we find that these synthetic datasets often contain entity types that are semantically highly similar to (or even the same as) those in standard evaluation benchmarks. Because of this overlap, we argue that reported F1 scores for zero-shot NER overestimate the true capabilities of these approaches. Further, we argue that current evaluation setups provide an incomplete picture of zero-shot abilities since they do not quantify the label shift (i.e., the similarity of labels) between training and evaluation datasets. To address these issues, we propose Familarity, a novel metric that captures both the semantic similarity between entity types in training and evaluation, as well as their frequency in the training data, to provide an estimate of label shift. It allows researchers to contextualize reported zero-shot NER scores when using custom synthetic training datasets. Further, it enables researchers to generate evaluation setups of various transfer difficulties for fine-grained analysis of zero-shot NER.

Code explanations are crucial in real-world life, from educating students to aligning technical projects with business goals. However, existing approaches face challenges balancing faithfulness to the original code and personalization for diverse user needs. This paper addresses these challenges by introducing a novel benchmark and method for generating faithful personalized code explanations. Our benchmark, FaithfulPersonaCodeX, incorporates code samples and user profiles, employing various evaluation metrics to evaluate both faithfulness and personalization. We propose DISCO, a new method that uses a self-critique mechanism and two-stage optimization to balance faithfulness and personalization in code explanations, addressing the limitations of current large language model approaches. Our proposed model, DISCO, achieves a notable 3.7% improvement in Pass@5 compared to the strong baseline method, Self-Consistency, while maintaining high personalization with a 61.08% win rate in the LLM-as-a-Judge evaluation, effectively balancing faithfulness and user-specific needs in code explanations.

Faithfulness evaluators based on Large Language Models (LLMs) are often fooled by the fluency of the text and struggle with identifying errors in the summaries, usually leading to high false negative rate. We propose an approach to summary faithfulness evaluation in which multiple LLM-based agents are assigned initial stances (regardless of what their belief might be) and forced to come up with a reason to justify the imposed belief, thus engaging in a multi-round debate to reach an agreement. The uniformly distributed initial assignments here result in a greater diversity of stances leading to more meaningful debates and ultimately more errors identified. Furthermore, by analyzing the recent faithfulness evaluation datasets, we observe that naturally, it is not always the case for a summary to be either faithful to the source document or not. We therefore introduce a new dimension ambiguity and a detailed taxonomy to identify such special cases. Experiments demonstrate our approach can help identify ambiguities, and have even a stronger performance on non-ambiguous summaries.

Summarization is one of the most common tasks performed by large language models (LLMs), especially in applications like Retrieval-Augmented Generation (RAG). However, existing evaluations of hallucinations in LLM-generated summaries, and evaluations of hallucination detection models both suffer from a lack of diversity and recency in the LLM and LLM families considered. This paper introduces FaithBench, a summarization hallucination benchmark comprising challenging hallucinations made by 10 modern LLMs from 8 different families, with ground truth annotations by human experts. “Challenging” here means summaries on which popular, state-of-the-art hallucination detection models, including GPT-4o-as-a-judge, disagreed on. Our results show GPT-4o and GPT-3.5-Turbo produce the least hallucinations. However, most state-of-the-art hallucination detection models have near 50% accuracies on FaithBench, indicating lots of room for future improvement.