Yun-Shiuan Chuang, Sameer Narendran, Nikunj Harlalka, Alexander Cheung, Sizhe Gao, Siddharth Suresh, Junjie Hu, Timothy T. Rogers. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Recent alignment algorithms such as direct preference optimization (DPO) have been developed to improve the safety of large language models (LLMs) by training these models to match human behaviors exemplified by preference data. However, these methods are both computationally intensive and lacking in controllability and transparency, inhibiting their widespread use. Furthermore, these tuning-based methods require large-scale preference data for training and are susceptible to noisy preference data. In this paper, we introduce a tuning-free alignment alternative, ProFS (Projection Filter for Subspaces), and demonstrate its effectiveness under the use case of toxicity reduction. Grounded on theory from factor analysis, ProFS is a sample-efficient model editing approach that identifies a toxic subspace in the model parameter space and reduces model toxicity by projecting away the detected subspace. The toxic subspace is identified by extracting preference data embeddings from the language model, and removing non-toxic information from these embeddings. We show that ProFS is more sample-efficient than DPO, further showcasing greater robustness to noisy data. Finally, we attempt to connect tuning based alignment with editing, by establishing both theoretical and empirical connections between ProFS and DPO, showing that ProFS can be interpreted as a denoised version of a single DPO step.

Preferences within a group of people are not uniform but follow a distribution. While existing alignment methods like Direct Preference Optimization (DPO) attempt to steer models to reflect human preferences, they struggle to capture the distributional pluralistic preferences within a group. These methods often skew toward dominant preferences, overlooking the diversity of opinions, especially when conflicting preferences arise. To address this issue, we propose Group Distributional Preference Optimization (GDPO), a novel framework that aligns language models with the distribution of preferences within a group by incorporating the concept of beliefs that shape individual preferences. GDPO calibrates a language model using statistical estimation of the group's belief distribution and aligns the model with belief-conditioned preferences, offering a more inclusive alignment framework than traditional methods. In experiments using both synthetic controllable opinion generation and real-world movie review datasets, we show that DPO fails to align with the targeted belief distributions, while GDPO consistently reduces this alignment gap during training. Additionally, our evaluation metrics demonstrate that GDPO outperforms existing approaches in aligning with group distributional preferences, marking a significant advance in pluralistic alignment.

Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 38% of the KV cache. This KV-cache reduction also leads to a 50% memory saving and a 2x speedup over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.

Recent generalist vision-language models (VLMs) have demonstrated impressive reasoning capabilities across diverse multimodal tasks. However, these models still struggle with fine-grained object level understanding and grounding. In terms of modeling, existing VLMs implicitly align text tokens with image patch tokens, which is ineffective for embedding alignment at the same granularity and inevitably introduces noisy spurious background features. Additionally, these models struggle when generalizing to unseen visual concepts and may not be reliable for domain-specific tasks without further fine-tuning. To address these limitations, we propose a novel method to prompt large language models with in-context visual object vectors, thereby enabling controllable object level reasoning. This eliminates the necessity of fusing a lengthy array of image patch features and significantly speeds up training. Furthermore, we propose region-level retrieval using our object representations, facilitating rapid adaptation to new objects without additional training. Our experiments reveal that our method achieves competitive referring object classification and captioning performance, while also offering zero-shot generalization and robustness to visually challenging contexts.

Vision-and-language navigation (VLN) enables the agent to navigate to a remote location following the natural language instruction in 3D environments. At each navigation step, the agent selects from possible candidate locations and then makes the move. For better navigation planning, the lookahead exploration strategy aims to effectively evaluate the agent's next action by accurately anticipating the future environment of candidate locations. To this end, some existing works predict RGB images for future environments, while this strategy suffers from image distortion and high computational cost. To address these issues, we propose the pre-trained hierarchical neural radiance representation model (HNR) to produce multi-level semantic features for future environments, which are more robust and efficient than pixel-wise RGB reconstruction. Furthermore, with the predicted future environmental representations, our lookahead VLN model is able to construct the navigable future path tree and select the optimal path via efficient parallel evaluation. Extensive experiments on the VLN-CE datasets confirm the effectiveness of our method. The code is available at https://github.com/MrZihan/HNR-VLN

Despite the general capabilities of Large Language Models (LLMs) like GPT-4, these models still request fine-tuning or adaptation with customized data when meeting the specific business demands and intricacies of tailored use cases. However, this process inevitably introduces new safety threats, particularly against the Fine-tuning based Jailbreak Attack (FJAttack) under the setting of Language-Model-as-a-Service (LMaaS), where the model's safety has been significantly compromised by fine-tuning on users' uploaded examples that contain just a few harmful examples. Though potential defenses have been proposed that the service providers of LMaaS can integrate safety examples into the fine-tuning dataset to reduce safety issues, such approaches require incorporating a substantial amount of data, making it inefficient. To effectively defend against the FJAttack with limited safety examples under LMaaS, we propose the Backdoor Enhanced Safety Alignment method inspired by an analogy with the concept of backdoor attacks. In particular, service providers will construct prefixed safety examples with a secret prompt, acting as a "backdoor trigger". By integrating prefixed safety examples into the fine-tuning dataset, the subsequent fine-tuning process effectively acts as the "backdoor attack", establishing a strong correlation between the secret prompt and safety generations. Consequently, safe responses are ensured once service providers prepend this secret prompt ahead of any user input during inference. Our comprehensive experiments demonstrate that through the Backdoor Enhanced Safety Alignment with adding as few as 11 prefixed safety examples, the maliciously fine-tuned LLMs will achieve similar safety performance as the original aligned models without harming the benign performance. Furthermore, we also present the effectiveness of our method in a more practical setting where the fine-tuning data consists of both FJAttack examples and the fine-tuning task data.

Recent advances in Large Language Models (LLMs) show the potential to significantly augment or even replace complex human writing activities. However, for complex tasks where people need to make decisions as well as write a justification, the trade offs between making work efficient and hindering decisions remain unclear. In this paper, we explore this question in the context of designing intelligent scaffolding for writing meta-reviews for an academic peer review process. We prototyped a system called "MetaWriter'' trained on five years of open peer review data to support meta-reviewing. The system highlights common topics in the original peer reviews, extracts key points by each reviewer, and on request, provides a preliminary draft of a meta-review that can be further edited. To understand how novice and experienced meta-reviewers use MetaWriter, we conducted a within-subject study with 32 participants. Each participant wrote meta-reviews for two papers: one with and one without MetaWriter. We found that MetaWriter significantly expedited the authoring process and improved the coverage of meta-reviews, as rated by experts, compared to the baseline. While participants recognized the efficiency benefits, they raised concerns around trust, over-reliance, and agency. We also interviewed six paper authors to understand their opinions of using machine intelligence to support the peer review process and reported critical reflections. We discuss implications for future interactive AI writing tools to support complex synthesis work.

Out-of-distribution (OOD) detection is a critical task for reliable predictions over text.Finetuning with pre-trained language models has been a de facto procedure to derive OOD detectors with respect to in-distribution (ID) data.Despite its common use, the understanding of the role of fine-tuning and its necessity for OOD detection is largely unexplored.In this paper, we raise the question: is fine-tuning necessary for OOD detection?We present a study investigating the efficacy of directly leveraging pre-trained language models for OOD detection, without any model fine-tuning on the ID data.We compare the approach with several competitive fine-tuning objectives, and offer new insights under various types of distributional shifts.Extensive evaluations on 8 diverse ID-OOD dataset pairs demonstrate nearperfect OOD detection performance (with 0% FPR95 in many cases), strongly outperforming its fine-tuned counterparts.We show that using distance-based detection methods, pretrained language models are near-perfect OOD detectors when the distribution shift involves a domain change.Furthermore, we study the effect of fine-tuning on OOD detection and identify how to balance ID accuracy with OOD detection performance.Our code is publically available 1 .

The main challenge in the field of unsupervised machine translation (UMT) is to associate source-target sentences in the latent space. As people who speak different languages share biologically similar visual systems, various unsupervised multi-modal machine translation (UMMT) models have been proposed to improve the performances of UMT by employing visual contents in natural images to facilitate alignment. Commonly, relation information is the important semantic in a sentence. Compared with images, videos can better present the interactions between objects and the ways in which an object transforms over time. However, current state-of-the-art methods only explore scene-level or object-level information from images without explicitly modeling objects relation; thus, they are sensitive to spurious correlations, which poses a new challenge for UMMT models. In this paper, we employ a spatial-temporal graph obtained from videos to exploit object interactions in space and time for disambiguation purposes and to promote latent space alignment in UMMT. Our model employs multi-modal back-translation and features pseudo-visual pivoting, in which we learn a shared multilingual visual-semantic embedding space and incorporate visually pivoted captioning as additional weak supervision. Experimental results on the VATEX Translation 2020 and HowToWorld datasets validate the translation capabilities of our model on both sentence-level and word-level and generalizes well when videos are not available during the testing phase.

It has been shown that machine translation models usually generate poor translations for named entities that are infrequent in the training corpus.Earlier named entity translation methods mainly focus on phonetic transliteration, which ignores the sentence context for translation and is limited in domain and language coverage.To address this limitation, we propose DEEP, a DEnoising Entity Pretraining method that leverages large amounts of monolingual data and a knowledge base to improve named entity translation accuracy within sentences.Besides, we investigate a multi-task learning strategy that finetunes a pre-trained neural machine translation model on both entity-augmented monolingual data and parallel data to further improve entity translation.Experimental results on three language pairs demonstrate that DEEP results in significant improvements over strong denoising autoencoding baselines, with a gain of up to 1.3 BLEU and up to 9.2 entity accuracy points for English-Russian translation. 1

Bosheng Ding, Junjie Hu, Lidong Bing, Mahani Aljunied, Shafiq Joty, Luo Si, Chunyan Miao. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Reproducible benchmarks are crucial in driving progress of machine translation research.However, existing machine translation benchmarks have been mostly limited to highresource or well-represented languages.Despite an increasing interest in low-resource machine translation, there are no standardized reproducible benchmarks for many African languages, many of which are used by millions of speakers but have less digitized textual data.To tackle these challenges, we propose AFROMT, a standardized, clean, and reproducible machine translation benchmark for eight widely spoken African languages.We also develop a suite of analysis tools for system diagnosis taking into account unique properties of these languages.Furthermore, we explore the newly considered case of low-resource focused pretraining and develop two novel data augmentation-based strategies, leveraging word-level alignment information and pseudo-monolingual data for pretraining multilingual sequence-to-sequence models.We demonstrate significant improvements when pretraining on 11 languages, with gains of up to 2 BLEU points over strong baselines.We also show gains of up to 12 BLEU points over cross-lingual transfer baselines in data-constrained scenarios.All code and pretrained models will be released as further steps towards larger reproducible benchmarks for African languages.1

Sebastian Ruder, Noah Constant, Jan Botha, Aditya Siddhant, Orhan Firat, Jinlan Fu, Pengfei Liu, Junjie Hu, Dan Garrette, Graham Neubig, Melvin Johnson. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Junjie Hu, Melvin Johnson, Orhan Firat, Aditya Siddhant, Graham Neubig. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Po-Yao Huang, Mandela Patrick, Junjie Hu, Graham Neubig, Florian Metze, Alexander Hauptmann. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Previous storytelling approaches mostly focused on optimizing traditional metrics such as BLEU, ROUGE and CIDEr. In this paper, we re-examine this problem from a different angle, by looking deep into what defines a natural and topically-coherent story. To this end, we propose three assessment criteria: relevance, coherence and expressiveness, which we observe through empirical analysis could constitute a “high-quality” story to the human eye. We further propose a reinforcement learning framework, ReCo-RL, with reward functions designed to capture the essence of these quality criteria. Experiments on the Visual Storytelling Dataset (VIST) with both automatic and human evaluation demonstrate that our ReCo-RL model achieves better performance than state-of-the-art baselines on both traditional metrics and the proposed new criteria.

Unsupervised machine translation (MT) has recently achieved impressive results with monolingual corpora only.However, it is still challenging to associate source-target sentences in the latent space.As people speak different languages biologically share similar visual systems, the potential of achieving better alignment through visual content is promising yet under-explored in unsupervised multimodal MT (MMT).In this paper, we investigate how to utilize visual content for disambiguation and promoting latent space alignment in unsupervised MMT.Our model employs multimodal back-translation and features pseudo visual pivoting in which we learn a shared multilingual visual-semantic embedding space and incorporate visuallypivoted captioning as additional weak supervision.The experimental results on the widely used Multi30K dataset show that the proposed model significantly improves over the state-ofthe-art methods and generalizes well when images are not available at the testing time.

The goal of universal machine translation is to learn to translate between any pair of languages. Despite impressive empirical results and an increasing interest in massively multilingual models, theoretical analysis on translation errors made by such universal machine translation models is only nascent. In this paper, we formally prove certain impossibilities of this endeavour in general, as well as prove positive results in the presence of additional (but natural) structure of data. For the former, we derive a lower bound on the translation error in the many-to-many translation setting, which shows that any algorithm aiming to learn shared sentence representations among multiple language pairs has to make a large translation error on at least one of the translation tasks, if no assumption on the structure of the languages is made. For the latter, we show that if the paired documents in the corpus follow a natural \emph{encoder-decoder} generative process, we can expect a natural notion of “generalization”: a linear number of language pairs, rather than quadratic, suffices to learn a good representation. Our theory also explains what kinds of connection graphs between pairs of languages are better suited: ones with longer paths result in worse sample complexity. We believe our theoretical insights and implications contribute to the future algorithmic design of universal machine translation.

Much recent progress in applications of machine learning models to NLP has been driven by benchmarks that evaluate models across a wide variety of tasks. However, these broad-coverage benchmarks have been mostly limited to English, and despite an increasing interest in multilingual models, a benchmark that enables the comprehensive evaluation of such methods on a diverse range of languages and tasks is still missing. To this end, we introduce the Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark, a multi-task benchmark for evaluating the cross-lingual generalization capabilities of multilingual representations across 40 languages and 9 tasks. We demonstrate that while models tested on English reach human performance on many tasks, there is still a sizable gap in the performance of cross-lingually transferred models, particularly on syntactic and sentence retrieval tasks. There is also a wide spread of results across languages. We will release the benchmark to encourage research on cross-lingual learning methods that transfer linguistic knowledge across a diverse and representative set of languages and tasks.

It has been previously noted that neural machine translation (NMT) is very sensitive to domain shift.In this paper, we argue that this is a dual effect of the highly lexicalized nature of NMT, resulting in failure for sentences with large numbers of unknown words, and lack of supervision for domain-specific words.To remedy this problem, we propose an unsupervised adaptation method which finetunes a pre-trained out-of-domain NMT model using a pseudo-in-domain corpus.Specifically, we perform lexicon induction to extract an in-domain lexicon, and construct a pseudo-parallel in-domain corpus by performing word-for-word back-translation of monolingual in-domain target sentences.In five domains over twenty pairwise adaptation settings and two model architectures, our method achieves consistent improvements without using any in-domain parallel sentences, improving up to 14 BLEU over unadapted models, and up to 2 BLEU over strong back-translation baselines.

Intelligent personal assistant systems that are able to have multi-turn conversations with human users are becoming increasingly popular. Most previous research has been focused on using either retrieval-based or generation-based methods to develop such systems. Retrieval-based methods have the advantage of returning fluent and informative responses with great diversity. However, the performance of the methods is limited by the size of the response repository. On the other hand, generation-based methods can produce highly coherent responses on any topics. But the generated responses are often generic and not informative due to the lack of grounding knowledge. In this paper, we propose a hybrid neural conversation model that combines the merits of both response retrieval and generation methods. Experimental results on Twitter and Foursquare data show that the proposed model outperforms both retrieval-based methods and generation-based methods (including a recently proposed knowledge-grounded neural conversation model) under both automatic evaluation metrics and human evaluation. We hope that the findings in this study provide new insights on how to integrate text retrieval and text generation models for building conversation systems.

Zi-Yi Dou, Junjie Hu, Antonios Anastasopoulos, Graham Neubig. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Ming Jiang, Junjie Hu, Qiuyuan Huang, Lei Zhang, Jana Diesner, Jianfeng Gao. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Chunting Zhou, Xuezhe Ma, Junjie Hu, Graham Neubig. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

This paper examines the problem of adapting neural machine translation systems to new, low-resourced languages (LRLs) as effectively and rapidly as possible.We propose methods based on starting with massively multilingual "seed models", which can be trained ahead-of-time, and then continuing training on data related to the LRL.We contrast a number of strategies, leading to a novel, simple, yet effective method of "similar-language regularization", where we jointly train on both a LRL of interest and a similar high-resourced language to prevent over-fitting to small LRL data.Experiments demonstrate that massively multilingual models, even without any explicit adaptation, are surprisingly effective, achieving BLEU scores of up to 15.5 with no data from the LRL, and that the proposed similarlanguage regularization method improves over other adaptation methods by 1.7 BLEU points average over 4 LRL settings.1

Classifying binary imbalanced streaming data is a significant task in both machine learning and data mining. Previously, online area under the receiver operating characteristic (ROC) curve (AUC) maximization has been proposed to seek a linear classifier. However, it is not well suited for handling nonlinearity and heterogeneity of the data. In this paper, we propose the kernelized online imbalanced learning (KOIL) algorithm, which produces a nonlinear classifier for the data by maximizing the AUC score while minimizing a functional regularizer. We address four major challenges that arise from our approach. First, to control the number of support vectors without sacrificing the model performance, we introduce two buffers with fixed budgets to capture the global information on the decision boundary by storing the corresponding learned support vectors. Second, to restrict the fluctuation of the learned decision function and achieve smooth updating, we confine the influence on a new support vector to its -nearest opposite support vectors. Third, to avoid information loss, we propose an effective compensation scheme after the replacement is conducted when either buffer is full. With such a compensation scheme, the performance of the learned model is comparable to the one learned with infinite budgets. Fourth, to determine good kernels for data similarity representation, we exploit the multiple kernel learning framework to automatically learn a set of kernels. Extensive experiments on both synthetic and real-world benchmark data sets demonstrate the efficacy of our proposed approach.

We study the problem of semi-supervised question answering--utilizing unlabeled text to boost the performance of question answering models.We propose a novel training framework, the Generative Domain-Adaptive Nets.In this framework, we train a generative model to generate questions based on the unlabeled text, and combine model-generated questions with human-generated questions for training question answering models.We develop novel domain adaptation algorithms, based on reinforcement learning, to alleviate the discrepancy between the modelgenerated data distribution and the humangenerated data distribution.Experiments show that our proposed framework obtains substantial improvement from unlabeled text.

Deep neural networks for machine comprehension typically utilizes only word or character embeddings without explicitly taking advantage of structured linguistic information such as constituency trees and dependency trees.In this paper, we propose structural embedding of syntactic trees (SEST), an algorithm framework to utilize structured information and encode them into vector representations that can boost the performance of algorithms for the machine comprehension.We evaluate our approach using a state-of-the-art neural attention model on the SQuAD dataset.Experimental results demonstrate that our model can accurately identify the syntactic boundaries of the sentences and extract answers that are syntactically coherent over the baseline methods.

Online learning from imbalanced streaming data to capture the nonlinearity and heterogeneity of the data is significant in machine learning and data mining. To tackle this problem, we propose a kernelized online imbalanced learning (KOIL) algorithm to directly maximize the area under the ROC curve (AUC). We address two more challenges: 1) How to control the number of support vectors without sacrificing model performance; and 2) how to restrict the fluctuation of the learned decision function to attain smooth updating. To this end, we introduce two buffers with fixed budgets (buffer sizes) for positive class and negative class, respectively, to store the learned support vectors, which can allow us to capture the global information of the decision boundary. When determining the weight of a new support vector, we confine its influence only to its $k$-nearest opposite support vectors. This can restrict the effect of new instances and prevent the harm of outliers. More importantly, we design a sophisticated scheme to compensate the model after replacement is conducted when either buffer is full. With this compensation, the learned model approaches the one learned with infinite budgets. We present both theoretical analysis and extensive experimental comparison to demonstrate the effectiveness of our proposed KOIL.

Social information between users has been widely used to improve the traditional Recommender System in many previous works. However, in many websites such as Amazon and eBay, there is no explicit social graph that can be used to improve the recommendation performance. Hence in this work, in order to make it possible to employ social recommendation methods in those non-social information websites, we propose a general framework to construct a homophily-based implicit social network by utilizing both the rating and comments of items given by the users. Our scalable framework can be easily extended to enhance the performance of any recommender systems without social network by replacing the homophily-based implicit social relation definition. We propose four methods to extract and analyze the implicit social links between users, and then conduct the experiments on Amazon dataset. Experimental results show that our proposed methods work better than traditional recommendation methods without social information.