Shujian Huang

Wenhao Zhu, Hongyi Liu, Qingxiu Dong et al. · cmu, pku

Large language models (LLMs) have demonstrated remarkable potential in handling multilingual machine translation (MMT). In this paper, we systematically investigate the advantages and challenges of LLMs for MMT by answering two questions: 1) How well do LLMs perform in translating massive languages? 2) Which factors affect LLMs' performance in translation? We thoroughly evaluate eight popular LLMs, including ChatGPT and GPT-4. Our empirical results show that translation capabilities of LLMs are continually involving. GPT-4 has beat the strong supervised baseline NLLB in 40.91% of translation directions but still faces a large gap towards the commercial translation system like Google Translate, especially on low-resource languages. Through further analysis, we discover that LLMs exhibit new working patterns when used for MMT. First, LLM can acquire translation ability in a resource-efficient way and generate moderate translation even on zero-resource languages. Second, instruction semantics can surprisingly be ignored when given in-context exemplars. Third, cross-lingual exemplars can provide better task guidance for low-resource translation than exemplars in the same language pairs. Code will be released at: https://github.com/NJUNLP/MMT-LLM.

Wenhao Zhu, Yunzhe Lv, Qingxiu Dong et al. · cmu, pku

Existing large language models show disparate capability across different languages, due to the imbalance in the training data. Their performances on English tasks are often stronger than on tasks of other languages. In this paper, we empower pre-trained LLMs on non-English languages by building semantic alignment across languages. We start from targeting individual languages by performing cross-lingual instruction-tuning (CoIT) on LLaMA, i.e. tuning it with translation task data and cross-lingual general task data to obtain cross-lingual models (x-LLaMAs), and formulate underlying scaling laws to investigate the advantages of using scalable translation data. Then we perform multilingual instruction-tuning (MuIT) with mixed resources to build multilingual m-LLaMA. We also illustrate how we leverage the scaling laws to optimize data allocation in a resource-constrained setting. Experiment results on cross-lingual benchmarks XQUAD and MLQA show that x-LLaMAs surpass the English instruction-tuned counterpart (Alpaca) by an average of 27.83% across six non-English languages. Evaluation results on translation dataset Flores-101 show that x-LLaMAs outperform previous LLaMA-based models by an average of 18.89%. Encouragingly, m-LLaMA achieves comparable performance to x-LLaMAs on individual languages and demonstrates the ability to follow multilingual instructions. Further analysis on response content and representation space reveals the alignment of the multilingual semantic space within the middle layers of m-LLaMA.

Peng Ding, Jun Kuang, Dan Ma et al.

Large Language Models (LLMs), such as ChatGPT and GPT-4, are designed to provide useful and safe responses. However, adversarial prompts known as 'jailbreaks' can circumvent safeguards, leading LLMs to generate potentially harmful content. Exploring jailbreak prompts can help to better reveal the weaknesses of LLMs and further steer us to secure them. Unfortunately, existing jailbreak methods either suffer from intricate manual design or require optimization on other white-box models, which compromises either generalization or efficiency. In this paper, we generalize jailbreak prompt attacks into two aspects: (1) Prompt Rewriting and (2) Scenario Nesting. Based on this, we propose ReNeLLM, an automatic framework that leverages LLMs themselves to generate effective jailbreak prompts. Extensive experiments demonstrate that ReNeLLM significantly improves the attack success rate while greatly reducing the time cost compared to existing baselines. Our study also reveals the inadequacy of current defense methods in safeguarding LLMs. Finally, we analyze the failure of LLMs defense from the perspective of prompt execution priority, and propose corresponding defense strategies. We hope that our research can catalyze both the academic community and LLMs developers towards the provision of safer and more regulated LLMs. The code is available at https://github.com/NJUNLP/ReNeLLM.

Yu Bao, Hao Zhou, Shujian Huang et al. · cmu

Recently, parallel text generation has received widespread attention due to its success in generation efficiency. Although many advanced techniques are proposed to improve its generation quality, they still need the help of an autoregressive model for training to overcome the one-to-many multi-modal phenomenon in the dataset, limiting their applications. In this paper, we propose $\textit{latent}$-GLAT, which employs the discrete latent variables to capture word categorical information and invoke an advanced curriculum learning technique, alleviating the multi-modality problem. Experiment results show that our method outperforms strong baselines without the help of an autoregressive model, which further broadens the application scenarios of the parallel decoding paradigm.

Yiming Yan, Tao Wang, Chengqi Zhao et al.

Automatic metrics play a crucial role in machine translation. Despite the widespread use of n-gram-based metrics, there has been a recent surge in the development of pre-trained model-based metrics that focus on measuring sentence semantics. However, these neural metrics, while achieving higher correlations with human evaluations, are often considered to be black boxes with potential biases that are difficult to detect. In this study, we systematically analyze and compare various mainstream and cutting-edge automatic metrics from the perspective of their guidance for training machine translation systems. Through Minimum Risk Training (MRT), we find that certain metrics exhibit robustness defects, such as the presence of universal adversarial translations in BLEURT and BARTScore. In-depth analysis suggests two main causes of these robustness deficits: distribution biases in the training datasets, and the tendency of the metric paradigm. By incorporating token-level constraints, we enhance the robustness of evaluation metrics, which in turn leads to an improvement in the performance of machine translation systems. Codes are available at \url{https://github.com/powerpuffpomelo/fairseq_mrt}.

Peng Ding, Jingyu Wu, Jun Kuang et al.

Multi-modal Large Language Models (MLLMs) have demonstrated remarkable performance on various visual-language understanding and generation tasks. However, MLLMs occasionally generate content inconsistent with the given images, which is known as "hallucination". Prior works primarily center on evaluating hallucination using standard, unperturbed benchmarks, which overlook the prevalent occurrence of perturbed inputs in real-world scenarios-such as image cropping or blurring-that are critical for a comprehensive assessment of MLLMs' hallucination. In this paper, to bridge this gap, we propose Hallu-PI, the first benchmark designed to evaluate Hallucination in MLLMs within Perturbed Inputs. Specifically, Hallu-PI consists of seven perturbed scenarios, containing 1,260 perturbed images from 11 object types. Each image is accompanied by detailed annotations, which include fine-grained hallucination types, such as existence, attribute, and relation. We equip these annotations with a rich set of questions, making Hallu-PI suitable for both discriminative and generative tasks. Extensive experiments on 12 mainstream MLLMs, such as GPT-4V and Gemini-Pro Vision, demonstrate that these models exhibit significant hallucinations on Hallu-PI, which is not observed in unperturbed scenarios. Furthermore, our research reveals a severe bias in MLLMs' ability to handle different types of hallucinations. We also design two baselines specifically for perturbed scenarios, namely Perturbed-Reminder and Perturbed-ICL. We hope that our study will bring researchers' attention to the limitations of MLLMs when dealing with perturbed inputs, and spur further investigations to address this issue. Our code and datasets are publicly available at https://github.com/NJUNLP/Hallu-PI.

Xu Huang, Zhirui Zhang, Ruize Gao et al. · tencent-ai

We present IMTLab, an open-source end-to-end interactive machine translation (IMT) system platform that enables researchers to quickly build IMT systems with state-of-the-art models, perform an end-to-end evaluation, and diagnose the weakness of systems. IMTLab treats the whole interactive translation process as a task-oriented dialogue with a human-in-the-loop setting, in which human interventions can be explicitly incorporated to produce high-quality, error-free translations. To this end, a general communication interface is designed to support the flexible IMT architectures and user policies. Based on the proposed design, we construct a simulated and real interactive environment to achieve end-to-end evaluation and leverage the framework to systematically evaluate previous IMT systems. Our simulated and manual experiments show that the prefix-constrained decoding approach still gains the lowest editing cost in the end-to-end evaluation, while BiTIIMT achieves comparable editing cost with a better interactive experience.

Hao Zhou, Zhijun Wang, Shujian Huang et al.

Large Language Models (LLMs) are often English-centric due to the disproportionate distribution of languages in their pre-training data. Enhancing non-English language capabilities through post-pretraining often results in catastrophic forgetting of the ability of original languages. Previous methods either achieve good expansion with severe forgetting or slight forgetting with poor expansion, indicating the challenge of balancing language expansion while preventing forgetting. In this paper, we propose a method called MoE-LPR (Mixture-of-Experts with Language Priors Routing) to alleviate this problem. MoE-LPR employs a two-stage training approach to enhance the multilingual capability. First, the model is post-pretrained into a Mixture-of-Experts (MoE) architecture by upcycling, where all the original parameters are frozen and new experts are added. In this stage, we focus improving the ability on expanded languages, without using any original language data. Then, the model reviews the knowledge of the original languages with replay data amounting to less than 1% of post-pretraining, where we incorporate language priors routing to better recover the abilities of the original languages. Evaluations on multiple benchmarks show that MoE-LPR outperforms other post-pretraining methods. Freezing original parameters preserves original language knowledge while adding new experts preserves the learning ability. Reviewing with LPR enables effective utilization of multilingual knowledge within the parameters. Additionally, the MoE architecture maintains the same inference overhead while increasing total model parameters. Extensive experiments demonstrate MoE-LPR's effectiveness in improving expanded languages and preserving original language proficiency with superior scalability. Code and scripts are freely available at https://github.com/zjwang21/MoE-LPR.git.

Wenhao Zhu, Qianfeng Zhao, Yunzhe Lv et al.

Augmenting the base neural model with a token-level symbolic datastore is a novel generation paradigm and has achieved promising results in machine translation (MT). In this paper, we introduce a unified framework kNN-BOX, which enables quick development and interactive analysis for this novel paradigm. kNN-BOX decomposes the datastore-augmentation approach into three modules: datastore, retriever and combiner, thus putting diverse kNN generation methods into a unified way. Currently, kNN-BOX has provided implementation of seven popular kNN-MT variants, covering research from performance enhancement to efficiency optimization. It is easy for users to reproduce these existing works or customize their own models. Besides, users can interact with their kNN generation systems with kNN-BOX to better understand the underlying inference process in a visualized way. In the experiment section, we apply kNN-BOX for machine translation and three other seq2seq generation tasks, namely, text simplification, paraphrase generation and question generation. Experiment results show that augmenting the base neural model with kNN-BOX leads to a large performance improvement in all these tasks. The code and document of kNN-BOX is available at https://github.com/NJUNLP/knn-box.

Wenhao Zhu, Sizhe Liu, Shujian Huang et al.

Decoding by contrasting layers (DoLa), is designed to improve the generation quality of large language models (LLMs) by contrasting the prediction probabilities between an early exit output (amateur logits) and the final output (expert logits). However, we find that this approach does not work well on non-English tasks. Inspired by previous interpretability work on language transition during the model's forward pass, we discover that this issue arises from a language mismatch between early exit output and final output. In this work, we propose an improved contrastive decoding algorithm that is effective for diverse languages beyond English. To obtain more helpful amateur logits, we devise two strategies to skip a set of bottom, language-agnostic layers based on our preliminary analysis. Experimental results on multilingual reasoning benchmarks demonstrate that our proposed method outperforms previous contrastive decoding baselines and substantially improves LLM's chain-of-thought reasoning accuracy across 11 languages. The project will be available at: https://github.com/NJUNLP/SkipLayerCD.

Xiang Geng, Zhejian Lai, Yu Zhang et al.

We introduce the submissions of the NJUNLP team to the WMT 2023 Quality Estimation (QE) shared task. Our team submitted predictions for the English-German language pair on all two sub-tasks: (i) sentence- and word-level quality prediction; and (ii) fine-grained error span detection. This year, we further explore pseudo data methods for QE based on NJUQE framework (https://github.com/NJUNLP/njuqe). We generate pseudo MQM data using parallel data from the WMT translation task. We pre-train the XLMR large model on pseudo QE data, then fine-tune it on real QE data. At both stages, we jointly learn sentence-level scores and word-level tags. Empirically, we conduct experiments to find the key hyper-parameters that improve the performance. Technically, we propose a simple method that covert the word-level outputs to fine-grained error span results. Overall, our models achieved the best results in English-German for both word-level and fine-grained error span detection sub-tasks by a considerable margin.

Zewei Sun, Qingnan Jiang, Shujian Huang et al. · bytedance

Domain adaptation is an important challenge for neural machine translation. However, the traditional fine-tuning solution requires multiple extra training and yields a high cost. In this paper, we propose a non-tuning paradigm, resolving domain adaptation with a prompt-based method. Specifically, we construct a bilingual phrase-level database and retrieve relevant pairs from it as a prompt for the input sentences. By utilizing Retrieved Phrase-level Prompts (RePP), we effectively boost the translation quality. Experiments show that our method improves domain-specific machine translation for 6.2 BLEU scores and improves translation constraints for 11.5% accuracy without additional training.

Zihan Liu, Zewei Sun, Shanbo Cheng et al. · bytedance

Document-level Neural Machine Translation (DocNMT) has been proven crucial for handling discourse phenomena by introducing document-level context information. One of the most important directions is to input the whole document directly to the standard Transformer model. In this case, efficiency becomes a critical concern due to the quadratic complexity of the attention module. Existing studies either focus on the encoder part, which cannot be deployed on sequence-to-sequence generation tasks, e.g., Machine Translation (MT), or suffer from a significant performance drop. In this work, we keep the translation performance while gaining 20\% speed up by introducing extra selection layer based on lightweight attention that selects a small portion of tokens to be attended. It takes advantage of the original attention to ensure performance and dimension reduction to accelerate inference. Experimental results show that our method could achieve up to 95\% sparsity (only 5\% tokens attended) approximately, and save 93\% computation cost on the attention module compared with the original Transformer, while maintaining the performance.

Wenhao Zhu, Jingjing Xu, Shujian Huang et al.

Neural machine translation has achieved promising results on many translation tasks. However, previous studies have shown that neural models induce a non-smooth representation space, which harms its generalization results. Recently, kNN-MT has provided an effective paradigm to smooth the prediction based on neighbor representations during inference. Despite promising results, kNN-MT usually requires large inference overhead. We propose an effective training framework INK to directly smooth the representation space via adjusting representations of kNN neighbors with a small number of new parameters. The new parameters are then used to refresh the whole representation datastore to get new kNN knowledge asynchronously. This loop keeps running until convergence. Experiments on four benchmark datasets show that \method achieves average gains of 1.99 COMET and 1.0 BLEU, outperforming the state-of-the-art kNN-MT system with 0.02x memory space and 1.9x inference speedup.

Bin Wang, Jiangzhou Ju, Yang Fan et al.

As one of the challenging NLP tasks, designing math word problem (MWP) solvers has attracted increasing research attention for the past few years. In previous work, models designed by taking into account the properties of the binary tree structure of mathematical expressions at the output side have achieved better performance. However, the expressions corresponding to a MWP are often diverse (e.g., $n_1+n_2 \times n_3-n_4$, $n_3\times n_2-n_4+n_1$, etc.), and so are the corresponding binary trees, which creates difficulties in model learning due to the non-deterministic output space. In this paper, we propose the Structure-Unified M-Tree Coding Solver (SUMC-Solver), which applies a tree with any M branches (M-tree) to unify the output structures. To learn the M-tree, we use a mapping to convert the M-tree into the M-tree codes, where codes store the information of the paths from tree root to leaf nodes and the information of leaf nodes themselves, and then devise a Sequence-to-Code (seq2code) model to generate the codes. Experimental results on the widely used MAWPS and Math23K datasets have demonstrated that SUMC-Solver not only outperforms several state-of-the-art models under similar experimental settings but also performs much better under low-resource conditions.

Ziyuan Zhuang, Zhiyang Zhang, Sitao Cheng et al.

Retrieval-augmented generation (RAG) methods encounter difficulties when addressing complex questions like multi-hop queries. While iterative retrieval methods improve performance by gathering additional information, current approaches often rely on multiple calls of large language models (LLMs). In this paper, we introduce EfficientRAG, an efficient retriever for multi-hop question answering. EfficientRAG iteratively generates new queries without the need for LLM calls at each iteration and filters out irrelevant information. Experimental results demonstrate that EfficientRAG surpasses existing RAG methods on three open-domain multi-hop question-answering datasets.

Changjiang Gao, Shujian Huang, Jixing Li et al.

Recent large language models (LLMs) have revealed strong abilities to understand natural language. Since most of them share the same basic structure, i.e. the transformer block, possible contributors to their success in the training process are scaling and instruction tuning. However, how these factors affect the models' language perception is unclear. This work compares the self-attention of several existing LLMs (LLaMA, Alpaca and Vicuna) in different sizes (7B, 13B, 30B, 65B), together with eye saccade, an aspect of human reading attention, to assess the effect of scaling and instruction tuning on language perception. Results show that scaling enhances the human resemblance and improves the effective attention by reducing the trivial pattern reliance, while instruction tuning does not. However, instruction tuning significantly enhances the models' sensitivity to instructions. We also find that current LLMs are consistently closer to non-native than native speakers in attention, suggesting a sub-optimal language perception of all models. Our code and data used in the analysis is available on GitHub.

Zheng Ma, Mianzhi Pan, Wenhan Wu et al.

Vision-language models (VLMs) have shown impressive performance in substantial downstream multi-modal tasks. However, only comparing the fine-tuned performance on downstream tasks leads to the poor interpretability of VLMs, which is adverse to their future improvement. Several prior works have identified this issue and used various probing methods under a zero-shot setting to detect VLMs' limitations, but they all examine VLMs using general datasets instead of specialized ones. In practical applications, VLMs are usually applied to specific scenarios, such as e-commerce and news fields, so the generalization of VLMs in specific domains should be given more attention. In this paper, we comprehensively investigate the capabilities of popular VLMs in a specific field, the food domain. To this end, we build a food caption dataset, Food-500 Cap, which contains 24,700 food images with 494 categories. Each image is accompanied by a detailed caption, including fine-grained attributes of food, such as the ingredient, shape, and color. We also provide a culinary culture taxonomy that classifies each food category based on its geographic origin in order to better analyze the performance differences of VLM in different regions. Experiments on our proposed datasets demonstrate that popular VLMs underperform in the food domain compared with their performance in the general domain. Furthermore, our research reveals severe bias in VLMs' ability to handle food items from different geographic regions. We adopt diverse probing methods and evaluate nine VLMs belonging to different architectures to verify the aforementioned observations. We hope that our study will bring researchers' attention to VLM's limitations when applying them to the domain of food or culinary cultures, and spur further investigations to address this issue.

Wenhao Zhu, Shujian Huang, Yunzhe Lv et al.

kNN-MT presents a new paradigm for domain adaptation by building an external datastore, which usually saves all target language token occurrences in the parallel corpus. As a result, the constructed datastore is usually large and possibly redundant. In this paper, we investigate the interpretability issue of this approach: what knowledge does the NMT model need? We propose the notion of local correctness (LAC) as a new angle, which describes the potential translation correctness for a single entry and for a given neighborhood. Empirical study shows that our investigation successfully finds the conditions where the NMT model could easily fail and need related knowledge. Experiments on six diverse target domains and two language-pairs show that pruning according to local correctness brings a light and more explainable memory for kNN-MT domain adaptation.

Fei Zhao, Taotian Pang, Zhen Wu et al.

In-context learning (ICL) is a new paradigm for natural language processing (NLP), where a large language model (LLM) observes a small number of demonstrations and a test instance as its input, and directly makes predictions without updating model parameters. Previous studies have revealed that ICL is sensitive to the selection and the ordering of demonstrations. However, there are few studies regarding the impact of the demonstration number on the ICL performance within a limited input length of LLM, because it is commonly believed that the number of demonstrations is positively correlated with model performance. In this paper, we found this conclusion does not always hold true. Through pilot experiments, we discover that increasing the number of demonstrations does not necessarily lead to improved performance. Building upon this insight, we propose a Dynamic Demonstrations Controller (D$^2$Controller), which can improve the ICL performance by adjusting the number of demonstrations dynamically. The experimental results show that D$^2$Controller yields a 4.6% relative improvement on ten different sizes of LLMs across ten datasets. Moreover, we also extend our method to previous ICL models and achieve competitive results.

Jiahuan Li, Shanbo Cheng, Zewei Sun et al. · bytedance

Nearest Neighbor Machine Translation (kNNMT) is a simple and effective method of augmenting neural machine translation (NMT) with a token-level nearest neighbor retrieval mechanism. The effectiveness of kNNMT directly depends on the quality of retrieved neighbors. However, original kNNMT builds datastores based on representations from NMT models, which would result in poor retrieval accuracy when NMT models are not good enough, leading to sub-optimal translation performance. In this paper, we propose PRED, a framework that leverages Pre-trained models for Datastores in kNN-MT. Better representations from pre-trained models allow us to build datastores of better quality. We also design a novel contrastive alignment objective to mitigate the representation gap between the NMT model and pre-trained models, enabling the NMT model to retrieve from better datastores. We conduct extensive experiments on both bilingual and multilingual translation benchmarks, including WMT17 English $\leftrightarrow$ Chinese, WMT14 English $\leftrightarrow$ German, IWSLT14 German $\leftrightarrow$ English, and IWSLT14 multilingual datasets. Empirical results demonstrate the effectiveness of PRED.

Shuaijie She, Xiang Geng, Shujian Huang et al.

Abstractive summarization is the process of generating a summary given a document as input. Although significant progress has been made, the factual inconsistency between the document and the generated summary still limits its practical applications. Previous work found that the probabilities assigned by the generation model reflect its preferences for the generated summary, including the preference for factual consistency, and the preference for the language or knowledge prior as well. To separate the preference for factual consistency, we propose an unsupervised framework named CoP by controlling the preference of the generation model with the help of prompt. More specifically, the framework performs an extra inference step in which a text prompt is introduced as an additional input. In this way, another preference is described by the generation probability of this extra inference process. The difference between the above two preferences, i.e. the difference between the probabilities, could be used as measurements for detecting factual inconsistencies. Interestingly, we found that with the properly designed prompt, our framework could evaluate specific preferences and serve as measurements for fine-grained categories of inconsistency, such as entity-related inconsistency, coreference-related inconsistency, etc. Moreover, our framework could also be extended to the supervised setting to learn better prompt from the labeled data as well. Experiments show that our framework achieves new SOTA results on three factual inconsistency detection tasks.

Zheng Ma, Shi Zong, Mianzhi Pan et al.

In recent years, vision and language pre-training (VLP) models have advanced the state-of-the-art results in a variety of cross-modal downstream tasks. Aligning cross-modal semantics is claimed to be one of the essential capabilities of VLP models. However, it still remains unclear about the inner working mechanism of alignment in VLP models. In this paper, we propose a new probing method that is based on image captioning to first empirically study the cross-modal semantics alignment of VLP models. Our probing method is built upon the fact that given an image-caption pair, the VLP models will give a score, indicating how well two modalities are aligned; maximizing such scores will generate sentences that VLP models believe are of good alignment. Analyzing these sentences thus will reveal in what way different modalities are aligned and how well these alignments are in VLP models. We apply our probing method to five popular VLP models, including UNITER, ROSITA, ViLBERT, CLIP, and LXMERT, and provide a comprehensive analysis of the generated captions guided by these models. Our results show that VLP models (1) focus more on just aligning objects with visual words, while neglecting global semantics; (2) prefer fixed sentence patterns, thus ignoring more important textual information including fluency and grammar; and (3) deem the captions with more visual words are better aligned with images. These findings indicate that VLP models still have weaknesses in cross-modal semantics alignment and we hope this work will draw researchers' attention to such problems when designing a new VLP model.

Bin Wang, Jiangzhou Ju, Yunlin Mao et al.

The numerical reasoning in the financial domain -- performing quantitative analysis and summarizing the information from financial reports -- can greatly increase business efficiency and reduce costs of billions of dollars. Here, we propose a numerical reasoning question answering system to answer numerical reasoning questions among financial text and table data sources, consisting of a retriever module, a generator module, and an ensemble module. Specifically, in the retriever module, in addition to retrieving the whole row data, we innovatively design a cell retriever that retrieves the gold cells to avoid bringing unrelated and similar cells in the same row to the inputs of the generator module. In the generator module, we utilize multiple generators to produce programs, which are operation steps to answer the question. Finally, in the ensemble module, we integrate multiple programs to choose the best program as the output of our system. In the final private test set in FinQA Competition, our system obtains 69.79 execution accuracy.

Ming Fang, Shi Zong, Jing Li et al.

Complaining is a speech act that expresses a negative inconsistency between reality and human expectations. While prior studies mostly focus on identifying the existence or the type of complaints, in this work, we present the first study in computational linguistics of measuring the intensity of complaints from text. Analyzing complaints from such perspective is particularly useful, as complaints of certain degrees may cause severe consequences for companies or organizations. We create the first Chinese dataset containing 3,103 posts about complaints from Weibo, a popular Chinese social media platform. These posts are then annotated with complaints intensity scores using Best-Worst Scaling (BWS) method. We show that complaints intensity can be accurately estimated by computational models with the best mean square error achieving 0.11. Furthermore, we conduct a comprehensive linguistic analysis around complaints, including the connections between complaints and sentiment, and a cross-lingual comparison for complaints expressions used by Chinese and English speakers. We finally show that our complaints intensity scores can be incorporated for better estimating the popularity of posts on social media.

Xinyou Wang, Zaixiang Zheng, Fei Ye et al.

This paper introduces diffusion protein language model (DPLM), a versatile protein language model that demonstrates strong generative and predictive capabilities for protein sequences. We first pre-train scalable DPLMs from evolutionary-scale protein sequences within a generative self-supervised discrete diffusion probabilistic framework, which generalizes language modeling for proteins in a principled way. After pre-training, DPLM exhibits the ability to generate structurally plausible, novel, and diverse protein sequences for unconditional generation. We further demonstrate the proposed diffusion generative pre-training makes DPLM possess a better understanding of proteins, making it a superior representation learner, which can be fine-tuned for various predictive tasks, comparing favorably to ESM2 (Lin et al., 2022). Moreover, DPLM can be tailored for various needs, which showcases its prowess of conditional generation in several ways: (1) conditioning on partial peptide sequences, e.g., generating scaffolds for functional motifs with high success rate; (2) incorporating other modalities as conditioner, e.g., structure-conditioned generation for inverse folding; and (3) steering sequence generation towards desired properties, e.g., satisfying specified secondary structures, through a plug-and-play classifier guidance. Code is released at \url{https://github.com/bytedance/dplm}.

Shuaijie She, Shujian Huang, Xingyun Wang et al.

LLMs (Large Language Models) usually interact with users in the form of dialogue and generate responses following their instructions, which naturally require dialogue comprehension abilities. However, dialogue comprehension is a general language ability which is hard to be evaluated directly. In this work, we propose to perform the evaluation focusing on the factual consistency issue with the help of the dialogue summarization task. Besides evaluating and analyzing the dialogue summarization performance (DIAC-Sum) of different LLMs, we also derive factual questions from the generated summaries and use them as a more flexible measurement of dialogue comprehension (DIAC-QA). Our evaluation shows that, on average, 26.8% of the summaries generated by LLMs contain factual inconsistency. Even ChatGPT, the strongest model evaluated, has such errors in 16% of its summaries. For answering the factual questions, which is more challenging, the average error rate of all evaluated LLMs is 36.1%. Both results indicate serious deficiencies. Detailed analysis shows that the understanding of subject/object of the conversation is still challenging for LLMs. Furthermore, to stimulate and enhance the dialogue comprehension ability of LLMs, we propose a fine-tuning paradigm with auto-constructed multi-task data, which achieved a relative error rate reduction of 11% on DIAC-QA.

Xinyou Wang, Zaixiang Zheng, Shujian Huang

Recently, non-autoregressive (NAR) neural machine translation models have received increasing attention due to their efficient parallel decoding. However, the probabilistic framework of NAR models necessitates conditional independence assumption on target sequences, falling short of characterizing human language data. This drawback results in less informative learning signals for NAR models under conventional MLE training, thereby yielding unsatisfactory accuracy compared to their autoregressive (AR) counterparts. In this paper, we propose a simple and model-agnostic multi-task learning framework to provide more informative learning signals. During training stage, we introduce a set of sufficiently weak AR decoders that solely rely on the information provided by NAR decoder to make prediction, forcing the NAR decoder to become stronger or else it will be unable to support its weak AR partners. Experiments on WMT and IWSLT datasets show that our approach can consistently improve accuracy of multiple NAR baselines without adding any additional decoding overhead.

Wenhao Zhu, Shujian Huang, Fei Yuan et al.

Large language models show compelling performance on reasoning tasks but they tend to perform much worse in languages other than English. This is unsurprising given that their training data largely consists of English text and instructions. A typical solution is to translate instruction data into all languages of interest, and then train on the resulting multilingual data, which is called translate-training. This approach not only incurs high cost, but also results in poorly translated data due to the non-standard formatting of mathematical chain-of-thought. In this paper, we explore the benefits of question alignment, where we train the model to translate reasoning questions into English by finetuning on X-English parallel question data. In this way we perform targeted, in-domain language alignment which makes best use of English instruction data to unlock the LLMs' multilingual reasoning abilities. Experimental results on LLaMA2-13B show that question alignment leads to consistent improvements over the translate-training approach: an average improvement of 11.3% and 16.1% accuracy across ten languages on the MGSM and MSVAMP multilingual reasoning benchmarks. The project will be available at: https://github.com/NJUNLP/QAlign.

Min Liu, Yu Bao, Chengqi Zhao et al.

Benefiting from the sequence-level knowledge distillation, the Non-Autoregressive Transformer (NAT) achieves great success in neural machine translation tasks. However, existing knowledge distillation has side effects, such as propagating errors from the teacher to NAT students, which may limit further improvements of NAT models and are rarely discussed in existing research. In this paper, we introduce selective knowledge distillation by introducing an NAT evaluator to select NAT-friendly targets that are of high quality and easy to learn. In addition, we introduce a simple yet effective progressive distillation method to boost NAT performance. Experiment results on multiple WMT language directions and several representative NAT models show that our approach can realize a flexible trade-off between the quality and complexity of training data for NAT models, achieving strong performances. Further analysis shows that distilling only 5% of the raw translations can help an NAT outperform its counterpart trained on raw data by about 2.4 BLEU.

Peng Ding, Jun Kuang, Zongyu Wang et al.

Large Language Models (LLMs) have shown impressive capabilities across various tasks but remain vulnerable to meticulously crafted jailbreak attacks. In this paper, we identify a critical safety gap: while LLMs are adept at detecting jailbreak prompts, they often produce unsafe responses when directly processing these inputs. Inspired by this insight, we propose SAGE (Self-Aware Guard Enhancement), a training-free defense strategy designed to align LLMs' strong safety discrimination performance with their relatively weaker safety generation ability. SAGE consists of two core components: a Discriminative Analysis Module and a Discriminative Response Module, enhancing resilience against sophisticated jailbreak attempts through flexible safety discrimination instructions. Extensive experiments demonstrate SAGE's effectiveness and robustness across various open-source and closed-source LLMs of different sizes and architectures, achieving an average 99% defense success rate against numerous complex and covert jailbreak methods while maintaining helpfulness on general benchmarks. We further conduct mechanistic interpretability analysis through hidden states and attention distributions, revealing the underlying mechanisms of this detection-generation discrepancy. Our work thus contributes to developing future LLMs with coherent safety awareness and generation behavior. Our code and datasets are publicly available at https://github.com/NJUNLP/SAGE.

Xiang Geng, Zhejian Lai, Jiajun Chen et al.

Quality Estimation (QE) models evaluate the quality of machine translations without reference translations, serving as the reward models for the translation task. Due to the data scarcity, synthetic data generation has emerged as a promising solution. However, synthetic QE data often suffers from distribution shift, which can manifest as discrepancies between pseudo and real translations, or in pseudo labels that do not align with human preferences. To tackle this issue, we introduce DCSQE, a novel framework for alleviating distribution shift in synthetic QE data. To reduce the difference between pseudo and real translations, we employ the constrained beam search algorithm and enhance translation diversity through the use of distinct generation models. DCSQE uses references, i.e., translation supervision signals, to guide both the generation and annotation processes, enhancing the quality of token-level labels. DCSQE further identifies the shortest phrase covering consecutive error tokens, mimicking human annotation behavior, to assign the final phrase-level labels. Specially, we underscore that the translation model can not annotate translations of itself accurately. Extensive experiments demonstrate that DCSQE outperforms SOTA baselines like CometKiwi in both supervised and unsupervised settings. Further analysis offers insights into synthetic data generation that could benefit reward models for other tasks. The code is available at https://github.com/NJUNLP/njuqe.

Peng Ding, Jun Kuang, Wen Sun et al.

Large language models (LLMs) remain vulnerable to jailbreaking attacks despite their impressive capabilities. Investigating these weaknesses is crucial for robust safety mechanisms. Existing attacks primarily distract LLMs by introducing additional context or adversarial tokens, leaving the core harmful intent unchanged. In this paper, we introduce ISA (Intent Shift Attack), which obfuscates LLMs about the intent of the attacks. More specifically, we establish a taxonomy of intent transformations and leverage them to generate attacks that may be misperceived by LLMs as benign requests for information. Unlike prior methods relying on complex tokens or lengthy context, our approach only needs minimal edits to the original request, and yields natural, human-readable, and seemingly harmless prompts. Extensive experiments on both open-source and commercial LLMs show that ISA achieves over 70% improvement in attack success rate compared to direct harmful prompts. More critically, fine-tuning models on only benign data reformulated with ISA templates elevates success rates to nearly 100%. For defense, we evaluate existing methods and demonstrate their inadequacy against ISA, while exploring both training-free and training-based mitigation strategies. Our findings reveal fundamental challenges in intent inference for LLMs safety and underscore the need for more effective defenses. Our code and datasets are available at https://github.com/NJUNLP/ISA.

Zhejian Lai, Xiang Geng, Zhijun Wang et al.

Mathematical reasoning is a primary indicator of large language models (LLMs) intelligence. However, existing LLMs exhibit failures of robustness and generalization. This paper attributes these deficiencies to spurious reasoning, i.e., producing answers from superficial features. To address this challenge, we propose the AdaR framework to enable adaptive reasoning, wherein models rely on problem-solving logic to produce answers. AdaR synthesizes logically equivalent queries by varying variable values, and trains models with RLVR on these data to penalize spurious logic while encouraging adaptive logic. To improve data quality, we extract the problem-solving logic from the original query and generate the corresponding answer by code execution, then apply a sanity check. Experimental results demonstrate that AdaR improves robustness and generalization, achieving substantial improvement in mathematical reasoning while maintaining high data efficiency. Analysis indicates that data synthesis and RLVR function in a coordinated manner to enable adaptive reasoning in LLMs. Subsequent analyses derive key design insights into the effect of critical factors and the applicability to instruct LLMs. Our project is available at https://github.com/NJUNLP/AdaR.

Peng Hu, Sizhe Liu, Changjiang Gao et al.

Large Language Models have demonstrated impressive reasoning capabilities across multiple languages. However, the relationship between capabilities in different languages is less explored. In this work, we decompose the process of reasoning tasks into two separated components: knowledge retrieval and knowledge-free reasoning, and analyze the relationship between cross-lingual transferability and these two components. With adapted commonsense reasoning datasets and constructed knowledge-free reasoning datasets, we show that the knowledge-free reasoning capability can be nearly perfectly transferred across various source-target language directions despite the secondary impact of resource in some specific target languages, while cross-lingual knowledge retrieval significantly hinders the transfer. Moreover, by analyzing the hidden states and feed-forward network neuron activation during the reasoning, we show that higher similarity of hidden representations and larger overlap of activated neurons could explain the better cross-lingual transferability of knowledge-free reasoning than knowledge retrieval. Thus, we hypothesize that knowledge-free reasoning shares similar neurons in different languages for reasoning, while knowledge is stored separately in different languages. Our code and data is available at: https://github.com/NJUNLP/Knowledge-Free-Reasoning.

Qingnan Jiang, Mingxuan Wang, Jun Cao et al.

How to effectively adapt neural machine translation (NMT) models according to emerging cases without retraining? Despite the great success of neural machine translation, updating the deployed models online remains a challenge. Existing non-parametric approaches that retrieve similar examples from a database to guide the translation process are promising but are prone to overfit the retrieved examples. In this work, we propose to learn Kernel-Smoothed Translation with Example Retrieval (KSTER), an effective approach to adapt neural machine translation models online. Experiments on domain adaptation and multi-domain machine translation datasets show that even without expensive retraining, KSTER is able to achieve improvement of 1.1 to 1.5 BLEU scores over the best existing online adaptation methods. The code and trained models are released at https://github.com/jiangqn/KSTER.

Xin Zheng, Zhirui Zhang, Junliang Guo et al.

kNN-MT, recently proposed by Khandelwal et al. (2020a), successfully combines pre-trained neural machine translation (NMT) model with token-level k-nearest-neighbor (kNN) retrieval to improve the translation accuracy. However, the traditional kNN algorithm used in kNN-MT simply retrieves a same number of nearest neighbors for each target token, which may cause prediction errors when the retrieved neighbors include noises. In this paper, we propose Adaptive kNN-MT to dynamically determine the number of k for each target token. We achieve this by introducing a light-weight Meta-k Network, which can be efficiently trained with only a few training samples. On four benchmark machine translation datasets, we demonstrate that the proposed method is able to effectively filter out the noises in retrieval results and significantly outperforms the vanilla kNN-MT model. Even more noteworthy is that the Meta-k Network learned on one domain could be directly applied to other domains and obtain consistent improvements, illustrating the generality of our method. Our implementation is open-sourced at https://github.com/zhengxxn/adaptive-knn-mt.

Peng Wu, Shujian Huang, Rongxiang Weng et al.

Relation detection is a core step in many natural language process applications including knowledge base question answering. Previous efforts show that single-fact questions could be answered with high accuracy. However, one critical problem is that current approaches only get high accuracy for questions whose relations have been seen in the training data. But for unseen relations, the performance will drop rapidly. The main reason for this problem is that the representations for unseen relations are missing. In this paper, we propose a simple mapping method, named representation adapter, to learn the representation mapping for both seen and unseen relations based on previously learned relation embedding. We employ the adversarial objective and the reconstruction objective to improve the mapping performance. We re-organize the popular SimpleQuestion dataset to reveal and evaluate the problem of detecting unseen relations. Experiments show that our method can greatly improve the performance of unseen relations while the performance for those seen part is kept comparable to the state-of-the-art. Our code and data are available at https://github.com/wudapeng268/KBQA-Adapter.

Shuaijie She, Wei Zou, Shujian Huang et al.

Though reasoning abilities are considered language-agnostic, existing LLMs exhibit inconsistent reasoning abilities across different languages, e.g., reasoning in the dominant language like English is superior to other languages due to the imbalance of multilingual training data. To enhance reasoning abilities in non-dominant languages, we propose a Multilingual-Alignment-as-Preference Optimization framework (MAPO), aiming to align the reasoning processes in other languages with the dominant language. Specifically, we harness an off-the-shelf translation model for the consistency between answers in non-dominant and dominant languages, which we adopt as the preference for optimization, e.g., Direct Preference Optimization (DPO) or Proximal Policy Optimization (PPO). Experiments show that MAPO stably achieves significant improvements in the multilingual reasoning of various models on all three benchmarks (MSVAMP +16.2%, MGSM +6.1%, and MNumGLUESub +13.3%), with improved reasoning consistency across languages.

Xinyou Wang, Zaixiang Zheng, Fei Ye et al.

Proteins are essential macromolecules defined by their amino acid sequences, which determine their three-dimensional structures and, consequently, their functions in all living organisms. Therefore, generative protein modeling necessitates a multimodal approach to simultaneously model, understand, and generate both sequences and structures. However, existing methods typically use separate models for each modality, limiting their ability to capture the intricate relationships between sequence and structure. This results in suboptimal performance in tasks that requires joint understanding and generation of both modalities. In this paper, we introduce DPLM-2, a multimodal protein foundation model that extends discrete diffusion protein language model (DPLM) to accommodate both sequences and structures. To enable structural learning with the language model, 3D coordinates are converted to discrete tokens using a lookup-free quantization-based tokenizer. By training on both experimental and high-quality synthetic structures, DPLM-2 learns the joint distribution of sequence and structure, as well as their marginals and conditionals. We also implement an efficient warm-up strategy to exploit the connection between large-scale evolutionary data and structural inductive biases from pre-trained sequence-based protein language models. Empirical evaluation shows that DPLM-2 can simultaneously generate highly compatible amino acid sequences and their corresponding 3D structures eliminating the need for a two-stage generation approach. Moreover, DPLM-2 demonstrates competitive performance in various conditional generation tasks, including folding, inverse folding, and scaffolding with multimodal motif inputs, as well as providing structure-aware representations for predictive tasks.

Zhu Danhao, Shen Si, Huang Shujian et al.

Recently, neural network based methods have shown their power in learning more expressive features on the task of knowledge graph embedding (KGE). However, the performance of deep methods often falls behind the shallow ones on simple graphs. One possible reason is that deep models are difficult to train, while shallow models might suffice for accurately representing the structure of the simple KGs. In this paper, we propose a neural network based model, named DeepE, to address the problem, which stacks multiple building blocks to predict the tail entity based on the head entity and the relation. Each building block is an addition of a linear and a non-linear function. The stacked building blocks are equivalent to a group of learning functions with different non-linear depth. Hence, DeepE allows deep functions to learn deep features, and shallow functions to learn shallow features. Through extensive experiments, we find DeepE outperforms other state-of-the-art baseline methods. A major advantage of DeepE is the robustness. DeepE achieves a Mean Rank (MR) score that is 6%, 30%, 65% lower than the best baseline methods on FB15k-237, WN18RR and YAGO3-10. Our design makes it possible to train much deeper networks on KGE, e.g. 40 layers on FB15k-237, and without scarifying precision on simple relations.

Changjiang Gao, Hongda Hu, Peng Hu et al.

Despite their strong ability to retrieve knowledge in English, current large language models show imbalance abilities in different languages. Two approaches are proposed to address this, i.e., multilingual pretraining and multilingual instruction tuning. However, whether and how do such methods contribute to the cross-lingual knowledge alignment inside the models is unknown. In this paper, we propose CLiKA, a systematic framework to assess the cross-lingual knowledge alignment of LLMs in the Performance, Consistency and Conductivity levels, and explored the effect of multilingual pretraining and instruction tuning on the degree of alignment. Results show that: while both multilingual pretraining and instruction tuning are beneficial for cross-lingual knowledge alignment, the training strategy needs to be carefully designed. Namely, continued pretraining improves the alignment of the target language at the cost of other languages, while mixed pretraining affect other languages less. Also, the overall cross-lingual knowledge alignment, especially in the conductivity level, is unsatisfactory for all tested LLMs, and neither multilingual pretraining nor instruction tuning can substantially improve the cross-lingual knowledge conductivity.

Shimao Zhang, Yu Bao, Shujian Huang

Recently, Large Language Models (LLMs) have demonstrated outstanding performance across a wide range of downstream language tasks. Temperature sampling is a commonly used decoding strategy for LLMs' generation process. However, a fixed temperature parameter is used in most cases, which may not always be an optimal choice for balancing generation quality and diversity. In this paper, we propose an effective Entropy-based Dynamic Temperature (EDT) Sampling method, to achieve a more balanced performance in terms of both generation quality and diversity by dynamically selecting the temperature parameter. Additionally, we also show model performance and comprehensive analyses for 4 different generation benchmarks. Our experiments show that EDT significantly outperforms the existing strategies across different tasks.

Shimao Zhang, Xiao Liu, Xin Zhang et al.

Large Language Models have demonstrated outstanding performance across various downstream tasks and have been widely applied in multiple scenarios. Human-annotated preference data is used for training to further improve LLMs' performance, which is constrained by the upper limit of human performance. Therefore, Self-Rewarding method has been proposed, where LLMs generate training data by rewarding their own outputs. However, the existing self-rewarding paradigm is not effective in mathematical reasoning scenarios and may even lead to a decline in performance. In this work, we propose the Process-based Self-Rewarding pipeline for language models, which introduces long-thought reasoning, step-wise LLM-as-a-Judge, and step-wise preference optimization within the self-rewarding paradigm. Our new paradigm successfully enhances the performance of LLMs on multiple mathematical reasoning benchmarks through iterative Process-based Self-Rewarding, demonstrating the immense potential of self-rewarding to achieve LLM reasoning that may surpass human capabilities.

Shuaijie She, Junxiao Liu, Yifeng Liu et al.

Large language models (LLMs) inevitably make mistakes when performing step-by-step mathematical reasoning. Process Reward Models (PRMs) have emerged as a promising solution by evaluating each reasoning step. However, existing PRMs typically output evaluation scores directly, limiting both learning efficiency and evaluation accuracy, which is further exacerbated by the scarcity of annotated data. To address these issues, we propose Reasoning-Driven Process Reward Modeling (R-PRM). First, we leverage stronger LLMs to generate seed data from limited annotations, effectively bootstrapping our model's reasoning capabilities and enabling comprehensive step-by-step evaluation. Second, we further enhance performance through preference optimization, without requiring additional annotated data. Third, we introduce inference-time scaling to fully harness the model's reasoning potential. Extensive experiments demonstrate R-PRM's effectiveness: on ProcessBench and PRMBench, it surpasses strong baselines by 11.9 and 8.5 points in F1 scores, respectively. When applied to guide mathematical reasoning, R-PRM achieves consistent accuracy improvements of over 8.5 points across six challenging datasets. Further analysis reveals that R-PRM exhibits more comprehensive evaluation and stronger generalization capabilities, thereby highlighting its significant potential.

Xu Huang, Zhirui Zhang, Xiang Geng et al.

This study investigates how Large Language Models (LLMs) leverage source and reference data in machine translation evaluation task, aiming to better understand the mechanisms behind their remarkable performance in this task. We design the controlled experiments across various input modes and model types, and employ both coarse-grained and fine-grained prompts to discern the utility of source versus reference information. We find that reference information significantly enhances the evaluation accuracy, while surprisingly, source information sometimes is counterproductive, indicating LLMs' inability to fully leverage the cross-lingual capability when evaluating translations. Further analysis of the fine-grained evaluation and fine-tuning experiments show similar results. These findings also suggest a potential research direction for LLMs that fully exploits the cross-lingual capability of LLMs to achieve better performance in machine translation evaluation tasks.

Xu Huang, Wenhao Zhu, Hanxu Hu et al.

Previous multilingual benchmarks focus primarily on simple understanding tasks, but for large language models(LLMs), we emphasize proficiency in instruction following, reasoning, long context understanding, code generation, and so on. However, measuring these advanced capabilities across languages is underexplored. To address the disparity, we introduce BenchMAX, a multi-way multilingual evaluation benchmark that allows for fair comparisons of these important abilities across languages. To maintain high quality, three distinct native-speaking annotators independently annotate each sample within all tasks after the data was machine-translated from English into 16 other languages. Additionally, we present a novel translation challenge stemming from dataset construction. Extensive experiments on BenchMAX reveal varying effectiveness of core capabilities across languages, highlighting performance gaps that cannot be bridged by simply scaling up model size. BenchMAX serves as a comprehensive multilingual evaluation platform, providing a promising test bed to promote the development of multilingual language models. The dataset and code are publicly accessible.

Wenhao Zhu, Shujian Huang, Fei Yuan et al.

Bridging the significant gap between large language model's English and non-English performance presents a great challenge. While some previous studies attempt to mitigate this gap with translated training data, the recently proposed question alignment framework leverages the model's English expertise to improve multilingual performance with minimum usage of expensive, error-prone translation. In this paper, we explore how broadly this method can be applied by examining its effects in reasoning with and without chain-of-thought, as well as with program-of-thought. We also explore applying this framework to extremely large language models in an efficient manner, such as through proxy-tuning. Experiment results on multilingual reasoning benchmarks mGSM, mSVAMP, xCSQA and xNLI demonstrate that we can extend question alignment framework to boost multilingual performance across diverse reasoning scenarios, model families, and sizes. For instance, when applied to the LLaMA2 models, it brings an average accuracy improvements of 12.2% on mGSM even with the 70B model. To understand the mechanism of its success, we analyze representation space, generated response and data scales, and reveal how question translation training strengthens language alignment within LLMs and shapes their working patterns.

Changjiang Gao, Jixing Li, Jiajun Chen et al.

The process of meaning composition, wherein smaller units like morphemes or words combine to form the meaning of phrases and sentences, is essential for human sentence comprehension. Despite extensive neurolinguistic research into the brain regions involved in meaning composition, a computational metric to quantify the extent of composition is still lacking. Drawing on the key-value memory interpretation of transformer feed-forward network blocks, we introduce the Composition Score, a novel model-based metric designed to quantify the degree of meaning composition during sentence comprehension. Experimental findings show that this metric correlates with brain clusters associated with word frequency, structural processing, and general sensitivity to words, suggesting the multifaceted nature of meaning composition during human sentence comprehension.

Cheng-Yen Hsieh, Xinyou Wang, Daiheng Zhang et al.

Multimodal protein language models (PLMs) integrate sequence and token-based structural information, serving as a powerful foundation for protein modeling, generation, and design. However, the reliance on tokenizing 3D structures into discrete tokens causes substantial loss of fidelity about fine-grained structural details and correlations. In this paper, we systematically elucidate the design space of multimodal PLMs to overcome their limitations. We identify tokenization loss and inaccurate structure token predictions by the PLMs as major bottlenecks. To address these, our proposed design space covers improved generative modeling, structure-aware architectures and representation learning, and data exploration. Our advancements approach finer-grained supervision, demonstrating that token-based multimodal PLMs can achieve robust structural modeling. The effective design methods dramatically improve the structure generation diversity, and notably, folding abilities of our 650M model by reducing the RMSD from 5.52 to 2.36 on PDB testset, even outperforming 3B baselines and on par with the specialized folding models. Project page and code: https://bytedance.github.io/dplm/dplm-2.1/.