The widespread adoption of reinforcement learning-based alignment highlights the growing importance of reward models. Various benchmarks have been built to evaluate reward models in various domains and scenarios. However, a significant gap remains in assessing reward models for long-form generation, despite its critical role in real-world applications. To bridge this, we introduce Long-form RewardBench, the first reward modeling testbed specifically designed for long-form generation. Our benchmark encompasses five key subtasks: QA, RAG, Chat, Writing, and Reasoning. We collected instruction and preference data through a meticulously designed multi-stage data collection process, and conducted extensive experiments on 20+ mainstream reward models, including both classifiers and generative models. Our findings reveal that current models still lack long-form reward modeling capabilities. Furthermore, we designed a novel Long-form Needle-in-a-Haystack Test, which revealed a correlation between reward modeling performance and the error's position within a response, as well as the overall response length, with distinct characteristics observed between classification and generative models. Finally, we demonstrate that classifier exhibit better generalizability compared to generative models trained on the same data. As the first benchmark for long-form reward modeling, this work aims to offer a robust platform for visualizing progress in this crucial area.

The evaluation of large language models (LLMs) via benchmarks is widespread, yet inconsistencies between different leaderboards and poor separability among top models raise concerns about their ability to accurately reflect authentic model capabilities. This paper provides a critical analysis of benchmark effectiveness, examining mainstream prominent LLM benchmarks using results from diverse models. We first propose Pseudo-Siamese Network for Item Response Theory (PSN-IRT), an enhanced Item Response Theory framework that incorporates a rich set of item parameters within an IRT-grounded architecture. PSN-IRT can be utilized for accurate and reliable estimations of item characteristics and model abilities. Based on PSN-IRT, we conduct extensive analysis on 11 LLM benchmarks comprising 41,871 items, revealing significant and varied shortcomings in their measurement quality. Furthermore, we demonstrate that leveraging PSN-IRT is able to construct smaller benchmarks while maintaining stronger alignment with human preference.

Kunpeng Kang, Shuaimin Li, Kaiyuan Zhang, Luyang Zhang, Jiasheng Si, Bing Xu, Kehai Chen, Muyun Yang, Wenpeng Lu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Gewen Liang, Mufan Xu, Kehai Chen, Wei Wang, Yuwei Wang, Muyun Yang, Tiejun Zhao, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Despite being empowered with alignment mechanisms, large language models (LLMs) are increasingly vulnerable to emerging jailbreak attacks that can compromise their alignment mechanisms. This vulnerability poses significant risks to real-world applications. Existing work faces challenges in both training efficiency and generalization capabilities (i.e., Reinforcement Learning from Human Feedback and Red-Teaming). Developing effective strategies to enable LLMs to resist continuously evolving jailbreak attempts represents a significant challenge. To address this challenge, we propose a novel defensive paradigm called GuidelineLLM, which assists LLMs in recognizing queries that may have harmful content. Before LLMs respond to a query, GuidelineLLM first identifies potential risks associated with the query, summarizes these risks into guideline suggestions, and then feeds these guidelines to the responding LLMs. Importantly, our approach eliminates the necessity for additional safety fine-tuning of the LLMs themselves; only the GuidelineLLM requires fine-tuning. This characteristic enhances the general applicability of GuidelineLLM across various LLMs. Experimental results demonstrate that GuidelineLLM can significantly reduce the attack success rate (ASR) against LLM (an average reduction of 34.17% ASR) while maintaining the usefulness of LLM in handling benign queries.

Andong Chen, Yuchen Song, Kehai Chen, Xuefeng Bai, Muyun Yang, Liqiang Nie, Jie Liu, Tiejun Zhao, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Hui Huang, Jiaheng Liu, Yancheng He, Shilong Li, Bing Xu, Conghui Zhu, Muyun Yang, Tiejun Zhao. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Over-correction is a critical issue for large language models (LLMs) to address Grammatical Error Correction (GEC) task, esp. for Chinese. This paper proposes a Chain-of-Task (CoTask) framework to reduce over-correction. The CoTask framework is applied as multi-task instruction tuning of LLMs by decomposing the process of grammatical error analysis to design auxiliary tasks and adjusting the types and combinations of training tasks. A supervised fine-tuning (SFT) strategy is also presented to enhance the performance of LLMs, together with an algorithm for automatic dataset annotation to avoid additional manual costs. Experimental results demonstrate that our method achieves new state-of-the-art results on both FCGEC (in-domain) and NaCGEC (out-of-domain) test sets.

Low-Rank Adaptation (LoRA) is currently the most commonly used Parameter-efficient fine-tuning (PEFT) method. However, it still faces high computational and storage costs to models with billions of parameters. Most previous studies have tackled this issue by using pruning techniques. Nonetheless, these efforts only analyze LoRA parameter features to evaluate their importance, such as parameter count, size, and gradient. In fact, the output of LoRA directly impacts the fine-tuned model. Preliminary experiments indicate that a fraction of LoRA possesses significantly high output values, substantially influencing the layer output. Motivated by the observation, we propose LoRA-drop. Concretely, LoRA-drop evaluates the importance of LoRA based on the LoRA output. Then we retain LoRA for important layers and the other layers share the same LoRA. We conduct abundant experiments with models of different scales on NLU and NLG tasks. Results demonstrate that LoRA-drop can achieve performance comparable to full fine-tuning and LoRA while retaining 50% of the LoRA parameters on average.

Large language models (LLMs) have shown remarkable performance in general translation tasks.However, the increasing demand for high-quality translations that are not only adequate but also fluent and elegant.To assess the extent to which current LLMs can meet these demands, we introduce a suitable benchmark (PoetMT) for translating classical Chinese poetry into English.This task requires not only adequacy in translating culturally and historically significant content but also a strict adherence to linguistic fluency and poetic elegance.Our study reveals that existing LLMs fall short of this task.To address these issues, we propose RAT, a Retrieval-Augmented machine Translation method that enhances the translation process by incorporating knowledge related to classical poetry.Additionally, we propose an automatic evaluation metric based on GPT-4, which better assesses translation quality in terms of adequacy, fluency, and elegance, overcoming the limitations of traditional metrics.Our dataset and code will be made available 1 .

Junkai Liu, Yujie Tong, Hui Huang, Bowen Zheng, Yiran Hu, Peicheng Wu, Chuan Xiao, Makoto Onizuka, Muyun Yang, Shuyuan Zheng. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Kaiyuan Zhang, Qian Liu, Luyang Zhang, Chaoqun Zheng, Shuaimin Li, Bing Xu, Muyun Yang, Xinxiao Qiao, Wenpeng Lu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

The International Classification of Diseases (ICD) coding assigns standardized codes to diseases. Automating this process enhances the efficiency and accuracy of clinical records processing. However, current methods struggle with noisy and lengthy clinical texts, making it difficult to ensure the reliability of feature extraction. Furthermore, they typically make separate binary predictions for each code, overlooking the dependencies between them. To address these issues, we propose a novel model called Hierarchical Feature Optimization and Rescorer (HiRes). We employ a cascaded convolution architecture to mitigate noise and enhance feature representation. Additionally, a masked autoencoder-based rescorer is introduced to capture ICD code interdependencies, refining initial predictions for improved accuracy. The model also considers the inconsistencies in code representations. Experiments on the MIMIC datasets demonstrate the effectiveness of our model.

Recently, Multimodal Named Entity Recognition (MNER) has attracted significant attention. Although MNER utilizing in-context learning has shown improved performance, modality retrieval bias often diminishes the relevance of in-context examples. To address this issue, we propose a self-relevance-based multimodal in-context learning method to mitigate modality retrieval bias by dynamically adjusting the weight of each modality. Specifically, we first measure the self-relevance of the query by calculating the similarity between textual and visual modalities, which helps to assess how much visual information contributes to the textual context. Then, we rank the similarity of different modalities, adjust the image rankings based on self-relevance to reduce modality retrieval bias, and integrate them to select the k most relevant examples. Finally, we use task definition and retrieved examples as effective guidance provided to the Multimodal Large Language Models to obtain feedback. Experimental results demonstrate that our method achieves SOTA performance on two benchmark datasets.

The advancement of Large Language Models (LLMs) has spurred significant interest in Role-Playing Agents (RPAs) for applications such as emotional companionship and virtual interaction. However, recent RPAs are often built on explicit dialogue data, lacking deep, human-like internal thought processes, resulting in superficial knowledge and style expression. While Large Reasoning Models (LRMs) can be employed to simulate character thought, their direct application is hindered by attention diversion (i.e., RPAs forget their role) and style drift (i.e., overly formal and rigid reasoning rather than character-consistent reasoning). To address these challenges, this paper introduces a novel Role-Aware Reasoning (RAR) method, which consists of two important stages: Role Identity Activation (RIA) and Reasoning Style Optimization (RSO). RIA explicitly guides the model with character profiles during reasoning to counteract attention diversion, and then RSO aligns reasoning style with the character and scene via LRM distillation to mitigate style drift. Extensive experiments demonstrate that the proposed RAR significantly enhances the performance of RPAs by effectively addressing attention diversion and style drift.

State-of-the-art translation Quality Estimation (QE) models are proven to be biased.More specifically, they over-rely on monolingual features while ignoring the bilingual semantic alignment.In this work, we propose a novel method to mitigate the bias of the QE model and improve estimation performance.Our method is based on the contrastive learning between clean and noisy sentence pairs.We first introduce noise to the target side of the parallel sentence pair, forming the negative samples.With the original parallel pairs as the positive sample, the QE model is contrastively trained to distinguish the positive samples from the negative ones.This objective is jointly trained with the regression-style quality estimation, so as to prevent the QE model from overfitting to monolingual features.Experiments on WMT QE evaluation datasets demonstrate that our method improves the estimation performance by a large margin while mitigating the bias 1 .

The word embedding models such as Word2vec and FastText simultaneously learn dual representations of input vectors and output vectors. In contrast, almost all existing unsupervised bilingual lexicon induction (UBLI) methods use only input vectors without utilizing output vectors. In this paper, we propose a novel approach to making full use of both input and output vectors for more robust and strong UBLI. We discover the Common Difference Property that one orthogonal transformation can connect not only the input vectors of two languages but also the output vectors. Therefore, we can learn just one transformation to induce two different dictionaries from the input and output vectors, respectively. Between these two quite different dictionaries, a more accurate lexicon with less noise can be induced by taking the intersection of them in UBLI procedure. Extensive experiments show that our method achieves much more robust and strong results than state-of-the-art methods in distant language pairs, while reserving comparable performances in similar language pairs.

Document-level neural machine translation (DocNMT) universally encodes several local sentences or the entire document. Thus, DocNMT does not consider the relevance of document-level contextual information, for example, some context (i.e., content words, logical order, and co-occurrence relation) is more effective than another auxiliary context (i.e., functional and auxiliary words). To address this issue, we first utilize the word frequency information to recognize content words in the input document, and then use heuristical relations to summarize content words and sentences as a graph structure without relying on external syntactic knowledge. Furthermore, we apply graph attention networks to this graph structure to learn its feature representation, which allows DocNMT to more effectively capture the document-level context. Experimental results on several widely-used document-level benchmarks demonstrated the effectiveness of the proposed approach.

The encoder–decoder framework has been widely used in various practical artificial intelligence cyber-physical systems, including intelligent translation systems. The decoding process in such a framework usually demands the target-side representation, which is often learned by an autoaggressive decoder to simulate the target context information at the current time-step. However, the autoaggressive decoder only captures the previously generated partial target fragment and fails in simulating the global contextual information. In this article, we propose a new data-driven fuzzy context representation strategy to simulate the global target information. Specifically, we design two fuzzy methods to the global target contextual information, which are bag-of-words of target language generated via a softmax layer from the source-side representation and whole target sentence retrieved from the translation memory according to the source-side representation. Both methods facilitate the autoaggressive decoder to handle the global target context at the current time-step, thereby learning a more effective context vector for the generation of target translation. Extensive experiments on two machine translation tasks demonstrated that the proposed method achieved 3% improvement of BLEU score over a strong baseline.

Attribution extraction refers to find the attributes for the instances of a given semantic class, which is essential to enhance the schema of a knowledge graph. To facilitate the attribution extraction from the query log, this article proposes a pattern driven graph ranking approach to jointly employ the pattern and context distribution information. First, a simple pattern on query text is applied to automatically acquire seed attributes. Then, a graph-based weight propagation is designed to rank the patterns by context distribution algorithm information. Experimental results show that, on a Chinese query log collected by Baidu, the automatically acquired seeds are more representative than the classical manually assembled seeds, achieving an improvement of 11.6% in MAP as compared to the baseline approach. And the graph-based ranking algorithm manipulates the two types of evidence more effectively, outperforming both the distributional similarity based baseline and the HITS algorithm by 29.2% and 11.3%, respectively.

Neural models have achieved great success on the task of machine reading comprehension (MRC), which are typically trained on hard labels.We argue that hard labels limit the model capability on generalization due to the label sparseness problem.In this paper, we propose a robust training method for MRC models to address this problem.Our method consists of three strategies, 1) label smoothing, 2) word overlapping, 3) distribution prediction.All of them help to train models on soft labels.We validate our approach on the representative architecture -ALBERT.Experimental results show that our method can greatly boost the baseline with 1% improvement in average, and achieve state-of-the-art performance on NewsQA and QUOREF.

Source input information plays a very important role in the Transformer-based translation system. In practice, word embedding and positional embedding of each word are added as the input representation. Then self-attention networks are used to encode the global dependencies in the input representation to generate a source representation. However, this processing on the source representation only adopts a single source feature and excludes richer and more diverse features such as recurrence features, local features, and syntactic features, which results in tedious representation and thereby hinders the further translation performance improvement. In this paper, we introduce a simple and efficient method to encode more diverse source features into the input representation simultaneously, and thereby learning an effective source representation by self-attention networks. In particular, the proposed grouped strategy is only applied to the input representation layer, to keep the diversity of translation information and the efficiency of the self-attention networks at the same time. Experimental results show that our approach improves the translation performance over the state-of-the-art baselines of Transformer in regard to WMT14 English-to-German and NIST Chinese-to-English machine translation tasks.

Rare words are usually replaced with a singletoken in the current encoder-decoder style of neural machine translation, challenging the translation modeling by an obscured context. In this article, we propose to build a fuzzy semantic representation (FSR) method for rare words through a hierarchical clustering method to group rare words together, and integrate it into the encoder-decoder framework. This hierarchical structure can compensate for the semantic information in both source and target sides, and providing fuzzy context information to capture the semantic of rare words. The introduced FSR can also alleviate the data sparseness, which is the bottleneck in dealing with rare words in neural machine translation. In particular, our method is easily extended to the transformer-based neural machine translation model and learns the FSRs of all in-vocabulary words to enhance the sentence representations in addition to rare words. Our experiments on Chinese-to-English translation tasks confirm a significant improvement in the translation quality brought by the proposed method.

In statistical machine translation, translation prediction considers not only the aligned source word itself but also its source contextual information. Learning context representation is a promising method for improving translation results, particularly through neural networks. Most of the existing methods process context words sequentially and neglect source long-distance dependencies. In this paper, we propose a novel neural approach to source dependence-based context representation for translation prediction. The proposed model is capable of not only encoding source long-distance dependencies but also capturing functional similarities to better predict translations (i.e., word form translations and ambiguous word translations). To verify our method, the proposed mode is incorporated into phrase-based and hierarchical phrase-based translation models, respectively. Experiments on large-scale Chinese-to-English and English-to-German translation tasks show that the proposed approach achieves significant improvement over the baseline systems and outperforms several existing context-enhanced methods.

Learning context representations is very promising to improve translation results, particularly through neural networks. Previous efforts process the context words sequentially and neglect their internal syntactic structure. In this paper, we propose a novel neural network based on bi-convolutional architecture to represent the source dependency-based context for translation prediction. The proposed model is able to not only encode the long-distance dependencies but also capture the functional similarities for better translation prediction (i.e., ambiguous words translation and word forms translation). Examined by a large-scale Chinese-English translation task, the proposed approach achieves a significant improvement (of up to +1.9 BLEU points) over the baseline system, and meanwhile outperforms a number of context-enhanced comparison system.

East Asian languages are thought to handle reference differently from English, particularly in terms of the marking of definiteness and number.We present the first Data-Text corpus for Referring Expressions in Mandarin, and we use this corpus to test some initial hypotheses inspired by the theoretical linguistics literature.Our findings suggest that function words deserve more attention in Referring Expression Generation than they have so far received, and they have a bearing on the debate about whether different languages make different trade-offs between clarity and brevity.

This paper proposes a novel hierarchical recurrent neural network language model (HRNNLM) for document modeling.After establishing a RNN to capture the coherence between sentences in a document, HRNNLM integrates it as the sentence history information into the word level RNN to predict the word sequence with cross-sentence contextual information.A two-step training approach is designed, in which sentence-level and word-level language models are approximated for the convergence in a pipeline style.Examined by the standard sentence reordering scenario, HRNNLM is proved for its better accuracy in modeling the sentence coherence.And at the word level, experimental results also indicate a significant lower model perplexity, followed by a practical better translation result when applied to a Chinese-English document translation reranking task.

This paper introduces a maximum entropy method to Discourse Coherence Modeling (DCM). Different from the state-of-art supervised entity-grid model and unsupervised cohesion-driven model, the model we proposed only takes as input lexicon features, which increases the training speed and decoding speed significantly. We conduct an evaluation on two publicly available benchmark data sets via sentence ordering tasks, and the results confirm the effectiveness of our maximum entropy based approach in DCM.

Statistical Machine Translation (SMT) usually utilizes contextual information to disambiguate translation candidates. However, it is often limited to contexts within sentence boundaries, hence broader topical information cannot be leveraged. In this paper, we propose a novel approach to learning topic representation for paral-lel data using a neural network architec-ture, where abundant topical contexts are embedded via topic relevant monolingual data. By associating each translation rule with the topic representation, topic rele-vant rules are selected according to the dis-tributional similarity with the source text during SMT decoding. Experimental re-sults show that our method significantly improves translation accuracy in the NIST Chinese-to-English translation task com-pared to a state-of-the-art baseline. 1

The goal of predicting query potential for personalization is to determine which queries can benefit from personalization. In this paper, we investigate which kind of strategy is better for this task: classification or regression. We quantify the potential benefits of personalizing search results using two implicit click-based measures: Click entropy and Potential@N. Meanwhile, queries are characterized by query features and history features. Then we build C-SVM classification model and epsilon-SVM regression model respectively according to these two measures. The experimental results show that the classification model is a better choice for predicting query potential for personalization.

Logistic average misclassification percentage (lam%) is a key measure for the spam filtering performance. This paper demonstrates that a spam filter can achieve a perfect 0.00% in lam%, the minimal value in theory, by simply setting a biased threshold during the classifier modeling. At the same time, the overall classification performance reaches only a low accuracy. The result suggests that the role of lam% for spam filtering evaluation should be re-examined.