Minghao Wu

Pengfei Li, Liangyou Li, Meng Zhang et al.

Pre-trained sequence-to-sequence models have significantly improved Neural Machine Translation (NMT). Different from prior works where pre-trained models usually adopt an unidirectional decoder, this paper demonstrates that pre-training a sequence-to-sequence model but with a bidirectional decoder can produce notable performance gains for both Autoregressive and Non-autoregressive NMT. Specifically, we propose CeMAT, a conditional masked language model pre-trained on large-scale bilingual and monolingual corpora in many languages. We also introduce two simple but effective methods to enhance the CeMAT, aligned code-switching & masking and dynamic dual-masking. We conduct extensive experiments and show that our CeMAT can achieve significant performance improvement for all scenarios from low- to extremely high-resource languages, i.e., up to +14.4 BLEU on low resource and +7.9 BLEU improvements on average for Autoregressive NMT. For Non-autoregressive NMT, we demonstrate it can also produce consistent performance gains, i.e., up to +5.3 BLEU. To the best of our knowledge, this is the first work to pre-train a unified model for fine-tuning on both NMT tasks. Code, data, and pre-trained models are available at https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/CeMAT.

Chenyang Lyu, Minghao Wu, Longyue Wang et al.

Although instruction-tuned large language models (LLMs) have exhibited remarkable capabilities across various NLP tasks, their effectiveness on other data modalities beyond text has not been fully studied. In this work, we propose Macaw-LLM, a novel multi-modal LLM that seamlessly integrates visual, audio, and textual information. Macaw-LLM consists of three main components: a modality module for encoding multi-modal data, a cognitive module for harnessing pretrained LLMs, and an alignment module for harmonizing diverse representations. Our novel alignment module seamlessly bridges multi-modal features to textual features, simplifying the adaptation process from the modality modules to the cognitive module. In addition, we construct a large-scale multi-modal instruction dataset in terms of multi-turn dialogue, including 69K image instances and 50K video instances. We have made our data, code and model publicly available, which we hope can pave the way for future research in multi-modal LLMs and expand the capabilities of LLMs to handle diverse data modalities and address complex real-world scenarios.

Minghao Wu, Abdul Waheed, Chiyu Zhang et al.

Large language models (LLMs) with instruction fine-tuning demonstrate superior generative capabilities. However, these models are resource-intensive. To alleviate this issue, we explore distilling knowledge from instruction-tuned LLMs into much smaller ones. To this end, we carefully develop a large set of 2.58M instructions based on both existing and newly-generated instructions. In addition to being sizable, we design our instructions to cover a broad set of topics to ensure diversity. Extensive analysis of our instruction dataset confirms its diversity, and we generate responses for these instructions using gpt-3.5-turbo. Leveraging these instructions, we fine-tune a diverse herd of models, collectively referred to as LaMini-LM, which includes models from both the encoder-decoder and decoder-only families, with varying sizes. We evaluate the performance of our models using automatic metrics on 15 different natural language processing (NLP) benchmarks, as well as through human assessment. The results demonstrate that our proposed LaMini-LM models are comparable to competitive baselines, while being much smaller in size.

Huaixia Dou, Jie Zhu, Minghao Wu et al.

As large language models (LLMs) are increasingly deployed in financial services, a single non-compliant interaction can expose institutions to regulatory penalties and direct consumer harm. Existing guard models are built around general harm taxonomies and overlook violations grounded in specific financial regulations. We address this gap with a regulation-driven pipeline that operates directly on regulatory documents, inducing a financial compliance risk taxonomy and synthesizing grounded training data without any predefined violation categories. Instantiating the pipeline on Chinese financial regulations, we release \textbf{FinGuard-Bench}, to our knowledge the first benchmark for financial regulatory compliance detection, with expert-annotated labels at both the query and response levels. We further train \textbf{FinGuard}, a financial compliance detection model built on Qwen3-8B and trained on the regulation-grounded data via supervised fine-tuning and self-play reinforcement learning. On FinGuard-Bench, FinGuard substantially outperforms all baselines, including dedicated guard models and much larger general-purpose LLMs such as Qwen3.5-397B-A17B and GPT-5.1. Furthermore, FinGuard also preserves general safety capabilities and adapts to unseen institution-specific policies using policy documents alone. We will publicly release the code, prompts, and resources used in this work on GitHub.

Minghao Wu, Alham Fikri Aji

As large language models (LLMs) continue to advance, accurately and comprehensively evaluating their performance becomes increasingly challenging. Ranking the relative performance of LLMs based on Elo ratings, according to human judgment, is gaining more popularity. However, the extent to which humans and LLMs are capable evaluators remains uncertain. This study investigates the behavior of crowd-sourced and expert annotators, as well as LLMs, when comparing outputs from different models. To achieve this, we curate a dataset of intentionally flawed machine-generated answers. Our findings reveal a concerning bias in the evaluation process, as answers with factual errors are rated more favorably than answers that are too short or contained grammatical errors. To address this issue, we propose independently evaluating machine-generated text across multiple dimensions, rather than merging all the evaluation aspects into a single score. We instantiate this idea with the Elo rating system, resulting in the Multi-Elo Rating System (MERS). Empirical results from our study reveal that this proposed approach significantly enhances the quality of LLM-based evaluations, particularly in terms of factual accuracy. However, there is no significant improvement in crowd-sourced-based evaluations, indicating the need for further investigation.

Minghao Wu, George Foster, Lizhen Qu et al.

Existing work in document-level neural machine translation commonly concatenates several consecutive sentences as a pseudo-document, and then learns inter-sentential dependencies. This strategy limits the model's ability to leverage information from distant context. We overcome this limitation with a novel Document Flattening (DocFlat) technique that integrates Flat-Batch Attention (FBA) and Neural Context Gate (NCG) into Transformer model to utilize information beyond the pseudo-document boundaries. FBA allows the model to attend to all the positions in the batch and learns the relationships between positions explicitly and NCG identifies the useful information from the distant context. We conduct comprehensive experiments and analyses on three benchmark datasets for English-German translation, and validate the effectiveness of two variants of DocFlat. Empirical results show that our approach outperforms strong baselines with statistical significance on BLEU, COMET and accuracy on the contrastive test set. The analyses highlight that DocFlat is highly effective in capturing the long-range information.

Zhanyu Wang, Longyue Wang, Zhen Zhao et al.

While the recent advances in Multimodal Large Language Models (MLLMs) constitute a significant leap forward in the field, these models are predominantly confined to the realm of input-side multimodal comprehension, lacking the capacity for multimodal content generation. To fill this gap, we present GPT4Video, a unified multi-model framework that empowers Large Language Models (LLMs) with the capability of both video understanding and generation. Specifically, we develop an instruction-following-based approach integrated with the stable diffusion generative model, which has demonstrated to effectively and securely handle video generation scenarios. GPT4Video offers the following benefits: 1) It exhibits impressive capabilities in both video understanding and generation scenarios. For example, GPT4Video outperforms Valley by 11.8\% on the Video Question Answering task, and surpasses NExt-GPT by 2.3\% on the Text to Video generation task. 2) it endows the LLM/MLLM with video generation capabilities without requiring additional training parameters and can flexibly interface with a wide range of models to perform video generation. 3) it maintains a safe and healthy conversation not only in output-side but also the input side in an end-to-end manner. Qualitative and qualitative experiments demonstrate that GPT4Video holds the potential to function as a effective, safe and Humanoid-like video assistant that can handle both video understanding and generation scenarios.

Meng Shen, Minghao Wu, Deepu Rajan

Object hallucination is a significant challenge that hinders the application of large vision-language models (LVLMs) in practice. We hypothesize that one possible origin of hallucination is the model's tendency to prioritize text generation over meaningful interaction with images. To explore this, we examine the generation process and categorize text tokens into three groups: image-positive, invariant, and negative, based on their visual dependence on input image tokens. Our analysis reveals that most generated tokens are minimally influenced by the image information. This suggests that during the model's training stage, more emphasis is placed on learning how to follow textual instructions, rather than extracting information from images. Based on this finding, we propose adjusting the training weights of different tokens depending on their visual dependence to control hallucination. Additionally, we remove a portion of the training data that potentially contains more hallucinations as a data filtering strategy. Both methods achieve a reduction in hallucination without compromising response length or introducing additional computational costs during inference. We validate our methods across three LVLM variants, demonstrating the effectiveness and general applicability.

Weixuan Wang, Minghao Wu, Barry Haddow et al.

Large Language Models (LLMs) have shown remarkable capabilities in natural language processing but exhibit significant performance gaps among different languages. Most existing approaches to address these disparities rely on pretraining or fine-tuning, which are resource-intensive. To overcome these limitations without incurring significant costs, we propose Inference-Time Cross-Lingual Intervention (INCLINE), a novel framework that enhances LLM performance on low-performing (source) languages by aligning their internal representations with those of high-performing (target) languages during inference. INCLINE initially learns alignment matrices using parallel sentences from source and target languages through a Least-Squares optimization, and then applies these matrices during inference to transform the low-performing language representations toward the high-performing language space. Extensive experiments on nine benchmarks with five LLMs demonstrate that INCLINE significantly improves performance across diverse tasks and languages, compared to recent strong baselines. Our analysis demonstrates that INCLINE is highly cost-effective and applicable to a wide range of applications. In addition, we release the code to foster research along this line: https://github.com/weixuan-wang123/INCLINE.

Huifeng Yin, Yu Zhao, Minghao Wu et al.

Large Reasoning Models(LRMs) such as OpenAI o1 and DeepSeek-R1 have shown remarkable reasoning capabilities by scaling test-time compute and generating long Chain-of-Thought(CoT). Distillation--post-training on LRMs-generated data--is a straightforward yet effective method to enhance the reasoning abilities of smaller models, but faces a critical bottleneck: we found that distilled long CoT data poses learning difficulty for small models and leads to the inheritance of biases (i.e. over-thinking) when using Supervised Fine-tuning (SFT) and Reinforcement Learning (RL) methods. To alleviate this bottleneck, we propose constructing tree-based CoT data from scratch via Monte Carlo Tree Search(MCTS). We then exploit a set of CoT-aware approaches, including Thoughts Length Balance, Fine-grained DPO, and Joint Post-training Objective, to enhance SFT and RL on the constructed data. We conduct evaluation on various benchmarks such as math (GSM8K, MATH, AIME). instruction-following (Multi-IF) and planning (Blocksworld), results demonstrate our approaches substantially improve the reasoning performance of distilled models compared to standard distilled models via reducing the hallucinations in long-time thinking. The project homepage is https://github.com/AIDC-AI/Marco-o1.

Minghao Wu, Yuting Yan, Zhenyang Cai et al.

Sepsis management in the ICU requires sequential treatment decisions under rapidly evolving patient physiology. Although large language models (LLMs) encode broad clinical knowledge and can reason over guidelines, they are not inherently grounded in action-conditioned patient dynamics. We introduce SepsisAgent, a world model-augmented LLM agent for sepsis treatment recommendation. SepsisAgent uses a learned Clinical World Model to simulate patient responses under candidate fluid--vasopressor interventions, and follows a propose--simulate--refine workflow before committing to a prescription. We first show that world-model access alone yields inconsistent LLM decision performance, motivating agent-specific training. We then train SepsisAgent through a three-stage curriculum: patient-dynamics supervised fine-tuning, propose--simulate--refine behavior cloning, and world-model-based agentic reinforcement learning. On MIMIC-IV sepsis trajectories, SepsisAgent outperforms all traditional RL and LLM-based baselines in off-policy value while achieving the best safety profile under guideline adherence and unsafe-action metrics. Further analysis shows that repeated interaction with the Clinical World Model enables the agent to learn regularities in patient evolution, which remain useful even when simulator access is removed.

Bo Zeng, Chenyang Lyu, Sinuo Liu et al.

Instruction-following capability has become a major ability to be evaluated for Large Language Models (LLMs). However, existing datasets, such as IFEval, are either predominantly monolingual and centered on English or simply machine translated to other languages, limiting their applicability in multilingual contexts. In this paper, we present an carefully-curated extension of IFEval to a localized multilingual version named Marco-Bench-MIF, covering 30 languages with varying levels of localization. Our benchmark addresses linguistic constraints (e.g., modifying capitalization requirements for Chinese) and cultural references (e.g., substituting region-specific company names in prompts) via a hybrid pipeline combining translation with verification. Through comprehensive evaluation of 20+ LLMs on our Marco-Bench-MIF, we found that: (1) 25-35% accuracy gap between high/low-resource languages, (2) model scales largely impact performance by 45-60% yet persists script-specific challenges, and (3) machine-translated data underestimates accuracy by7-22% versus localized data. Our analysis identifies challenges in multilingual instruction following, including keyword consistency preservation and compositional constraint adherence across languages. Our Marco-Bench-MIF is available at https://github.com/AIDC-AI/Marco-Bench-MIF.

Haijun Li, Tianqi Shi, Zifu Shang et al.

Machine translation (MT) has become indispensable for cross-border communication in globalized industries like e-commerce, finance, and legal services, with recent advancements in large language models (LLMs) significantly enhancing translation quality. However, applying general-purpose MT models to industrial scenarios reveals critical limitations due to domain-specific terminology, cultural nuances, and stylistic conventions absent in generic benchmarks. Existing evaluation frameworks inadequately assess performance in specialized contexts, creating a gap between academic benchmarks and real-world efficacy. To address this, we propose a three-level translation capability framework: (1) Basic Linguistic Competence, (2) Domain-Specific Proficiency, and (3) Cultural Adaptation, emphasizing the need for holistic evaluation across these dimensions. We introduce TransBench, a benchmark tailored for industrial MT, initially targeting international e-commerce with 17,000 professionally translated sentences spanning 4 main scenarios and 33 language pairs. TransBench integrates traditional metrics (BLEU, TER) with Marco-MOS, a domain-specific evaluation model, and provides guidelines for reproducible benchmark construction. Our contributions include: (1) a structured framework for industrial MT evaluation, (2) the first publicly available benchmark for e-commerce translation, (3) novel metrics probing multi-level translation quality, and (4) open-sourced evaluation tools. This work bridges the evaluation gap, enabling researchers and practitioners to systematically assess and enhance MT systems for industry-specific needs.

Xintong Wang, Jingheng Pan, Yixiao Liu et al.

Vision-Language Translation (VLT) is a challenging task that requires accurately recognizing multilingual text embedded in images and translating it into the target language with the support of visual context. While recent Large Vision-Language Models (LVLMs) have demonstrated strong multilingual and visual understanding capabilities, there is a lack of systematic evaluation and understanding of their performance on VLT. In this work, we present a comprehensive study of VLT from three key perspectives: data quality, model architecture, and evaluation metrics. (1) We identify critical limitations in existing datasets, particularly in semantic and cultural fidelity, and introduce AibTrans -- a multilingual, parallel, human-verified dataset with OCR-corrected annotations. (2) We benchmark 11 commercial LVLMs/LLMs and 6 state-of-the-art open-source models across end-to-end and cascaded architectures, revealing their OCR dependency and contrasting generation versus reasoning behaviors. (3) We propose Density-Aware Evaluation to address metric reliability issues under varying contextual complexity, introducing the DA Score as a more robust measure of translation quality. Building upon these findings, we establish a new evaluation benchmark for VLT. Notably, we observe that fine-tuning LVLMs on high-resource language pairs degrades cross-lingual performance, and we propose a balanced multilingual fine-tuning strategy that effectively adapts LVLMs to VLT without sacrificing their generalization ability.

Haonan Li, Fajri Koto, Minghao Wu et al.

Instruction tuning has shown great promise in improving the performance of large language models. However, research on multilingual instruction tuning has been limited due to the scarcity of high-quality instruction-response datasets across different languages. To bridge this gap, we present Bactrian-X, a comprehensive multilingual parallel dataset of 3.4 million instruction-response pairs across 52 languages. Leveraging this dataset, we train a set of adapters using low-rank adaptation (LoRA), which are lightweight components that seamlessly integrate with large language models. These adapters have a substantially lower parameter count than the base model, making them easily replaceable and usable as plug-ins for different languages or language groups. Extensive experiments in various multilingual evaluation settings demonstrate that models derived from LoRA-based training over Bactrian-X outperform both the vanilla models and existing instruction-tuned models. The code and models are publicly available at https://github.com/mbzuai-nlp/bactrian-x

Minghao Wu, Thuy-Trang Vu, Lizhen Qu et al.

Large language models (LLMs) have significantly advanced various natural language processing (NLP) tasks. Recent research indicates that moderately-sized LLMs often outperform larger ones after task-specific fine-tuning. This study focuses on adapting LLMs for document-level machine translation (DocMT) for specific language pairs. We first investigate the impact of prompt strategies on translation performance and then conduct extensive experiments using two fine-tuning methods, three LLM backbones, and 18 translation tasks across nine language pairs. Our results show that specialized models can sometimes surpass GPT-4 in translation performance but still face issues like off-target translation due to error propagation in decoding. We provide an in-depth analysis of these LLMs tailored for DocMT, examining translation errors, discourse phenomena, strategies for training and inference, the data efficiency of parallel documents, recent test set evaluations, and zero-shot crosslingual transfer. Our findings highlight the strengths and limitations of LLM-based DocMT models and provide a foundation for future research.

Minghao Wu, Jiahao Xu, Yulin Yuan et al.

Literary translation remains one of the most challenging frontiers in machine translation due to the complexity of capturing figurative language, cultural nuances, and unique stylistic elements. In this work, we introduce TransAgents, a novel multi-agent framework that simulates the roles and collaborative practices of a human translation company, including a CEO, Senior Editor, Junior Editor, Translator, Localization Specialist, and Proofreader. The translation process is divided into two stages: a preparation stage where the team is assembled and comprehensive translation guidelines are drafted, and an execution stage that involves sequential translation, localization, proofreading, and a final quality check. Furthermore, we propose two innovative evaluation strategies: Monolingual Human Preference (MHP), which evaluates translations based solely on target language quality and cultural appropriateness, and Bilingual LLM Preference (BLP), which leverages large language models like GPT-4} for direct text comparison. Although TransAgents achieves lower d-BLEU scores, due to the limited diversity of references, its translations are significantly better than those of other baselines and are preferred by both human evaluators and LLMs over traditional human references and GPT-4} translations. Our findings highlight the potential of multi-agent collaboration in enhancing translation quality, particularly for longer texts.

Chenyang Lyu, Minghao Wu, Alham Fikri Aji

Large Language Models (LLMs) have demonstrated remarkable capabilities across various applications, fundamentally reshaping the landscape of natural language processing (NLP) research. However, recent evaluation frameworks often rely on the output probabilities of LLMs for predictions, primarily due to computational constraints, diverging from real-world LLM usage scenarios. While widely employed, the efficacy of these probability-based evaluation strategies remains an open research question. This study aims to scrutinize the validity of such probability-based evaluation methods within the context of using LLMs for Multiple Choice Questions (MCQs), highlighting their inherent limitations. Our empirical investigation reveals that the prevalent probability-based evaluation method inadequately aligns with generation-based prediction. Furthermore, current evaluation frameworks typically assess LLMs through predictive tasks based on output probabilities rather than directly generating responses, owing to computational limitations. We illustrate that these probability-based approaches do not effectively correspond with generative predictions. The outcomes of our study can enhance the understanding of LLM evaluation methodologies and provide insights for future research in this domain.

Sinuo Liu, Chenyang Lyu, Minghao Wu et al.

Recent advances in Large Reasoning Models (LRMs), particularly those leveraging Chain-of-Thought reasoning (CoT), have opened brand new possibility for Machine Translation (MT). This position paper argues that LRMs substantially transformed traditional neural MT as well as LLMs-based MT paradigms by reframing translation as a dynamic reasoning task that requires contextual, cultural, and linguistic understanding and reasoning. We identify three foundational shifts: 1) contextual coherence, where LRMs resolve ambiguities and preserve discourse structure through explicit reasoning over cross-sentence and complex context or even lack of context; 2) cultural intentionality, enabling models to adapt outputs by inferring speaker intent, audience expectations, and socio-linguistic norms; 3) self-reflection, LRMs can perform self-reflection during the inference time to correct the potential errors in translation especially extremely noisy cases, showing better robustness compared to simply mapping X->Y translation. We explore various scenarios in translation including stylized translation, document-level translation and multimodal translation by showcasing empirical examples that demonstrate the superiority of LRMs in translation. We also identify several interesting phenomenons for LRMs for MT including auto-pivot translation as well as the critical challenges such as over-localisation in translation and inference efficiency. In conclusion, we think that LRMs redefine translation systems not merely as text converters but as multilingual cognitive agents capable of reasoning about meaning beyond the text. This paradigm shift reminds us to think of problems in translation beyond traditional translation scenarios in a much broader context with LRMs - what we can achieve on top of it.

Minghao Wu, Weixuan Wang, Sinuo Liu et al.

As large language models (LLMs) continue to advance in linguistic capabilities, robust multilingual evaluation has become essential for promoting equitable technological progress. This position paper examines over 2,000 multilingual (non-English) benchmarks from 148 countries, published between 2021 and 2024, to evaluate past, present, and future practices in multilingual benchmarking. Our findings reveal that, despite significant investments amounting to tens of millions of dollars, English remains significantly overrepresented in these benchmarks. Additionally, most benchmarks rely on original language content rather than translations, with the majority sourced from high-resource countries such as China, India, Germany, the UK, and the USA. Furthermore, a comparison of benchmark performance with human judgments highlights notable disparities. STEM-related tasks exhibit strong correlations with human evaluations (0.70 to 0.85), while traditional NLP tasks like question answering (e.g., XQuAD) show much weaker correlations (0.11 to 0.30). Moreover, translating English benchmarks into other languages proves insufficient, as localized benchmarks demonstrate significantly higher alignment with local human judgments (0.68) than their translated counterparts (0.47). This underscores the importance of creating culturally and linguistically tailored benchmarks rather than relying solely on translations. Through this comprehensive analysis, we highlight six key limitations in current multilingual evaluation practices, propose the guiding principles accordingly for effective multilingual benchmarking, and outline five critical research directions to drive progress in the field. Finally, we call for a global collaborative effort to develop human-aligned benchmarks that prioritize real-world applications.

Longyue Wang, Siyou Liu, Chenyang Lyu et al.

Following last year, we have continued to host the WMT translation shared task this year, the second edition of the Discourse-Level Literary Translation. We focus on three language directions: Chinese-English, Chinese-German, and Chinese-Russian, with the latter two ones newly added. This year, we totally received 10 submissions from 5 academia and industry teams. We employ both automatic and human evaluations to measure the performance of the submitted systems. The official ranking of the systems is based on the overall human judgments. We release data, system outputs, and leaderboard at https://www2.statmt.org/wmt24/literary-translation-task.html.

Chiyu Zhang, Yifei Sun, Minghao Wu et al.

Content-based recommendation systems play a crucial role in delivering personalized content to users in the digital world. In this work, we introduce EmbSum, a novel framework that enables offline pre-computations of users and candidate items while capturing the interactions within the user engagement history. By utilizing the pretrained encoder-decoder model and poly-attention layers, EmbSum derives User Poly-Embedding (UPE) and Content Poly-Embedding (CPE) to calculate relevance scores between users and candidate items. EmbSum actively learns the long user engagement histories by generating user-interest summary with supervision from large language model (LLM). The effectiveness of EmbSum is validated on two datasets from different domains, surpassing state-of-the-art (SoTA) methods with higher accuracy and fewer parameters. Additionally, the model's ability to generate summaries of user interests serves as a valuable by-product, enhancing its usefulness for personalized content recommendations.

Minghao Wu, Yufei Wang, George Foster et al.

Document-level neural machine translation (DocNMT) aims to generate translations that are both coherent and cohesive, in contrast to its sentence-level counterpart. However, due to its longer input length and limited availability of training data, DocNMT often faces the challenge of data sparsity. To overcome this issue, we propose a novel Importance-Aware Data Augmentation (IADA) algorithm for DocNMT that augments the training data based on token importance information estimated by the norm of hidden states and training gradients. We conduct comprehensive experiments on three widely-used DocNMT benchmarks. Our empirical results show that our proposed IADA outperforms strong DocNMT baselines as well as several data augmentation approaches, with statistical significance on both sentence-level and document-level BLEU.

Qingyang Zhang, Xinke Kong, Haitao Wu et al.

Test-time training (TTT) adapts model parameters on unlabeled test instances during inference time, which continuously extends capabilities beyond the reach of offline training. Despite initial gains, existing TTT methods for LRMs plateau quickly and do not benefit from additional test-time compute. Without external calibration, the self-generated reward signal increasingly drifts as the policy model evolves, leading to both performance plateaus and diversity collapse. We propose TEMPO, a TTT framework that interleaves policy refinement on unlabeled questions with periodic critic recalibration on a labeled dataset. By formalizing this alternating procedure through the Expectation-Maximization (EM) algorithm, we reveal that prior methods can be interpreted as incomplete variants that omit the crucial recalibration step. Reintroducing this step tightens the evidence lower bound (ELBO) and enables sustained improvement. Across diverse model families (Qwen3 and OLMO3) and reasoning tasks, TEMPO improves OLMO3-7B on AIME 2024 from 33.0% to 51.1% and Qwen3-14B from 42.3% to 65.8%, while maintaining high diversity.

Weixuan Wang, Minghao Wu, Barry Haddow et al.

Large language models (LLMs) are designed to perform a wide range of tasks. To improve their ability to solve complex problems requiring multi-step reasoning, recent research leverages process reward modeling to provide fine-grained feedback at each step of the reasoning process for reinforcement learning (RL), but it predominantly focuses on English. In this paper, we tackle the critical challenge of extending process reward models (PRMs) to multilingual settings. To achieve this, we train multilingual PRMs on a dataset spanning seven languages, which is translated from English. Through comprehensive evaluations on two widely used reasoning benchmarks across 11 languages, we demonstrate that multilingual PRMs not only improve average accuracy but also reduce early-stage reasoning errors. Furthermore, our results highlight the sensitivity of multilingual PRMs to both the number of training languages and the volume of English data, while also uncovering the benefits arising from more candidate responses and trainable parameters. This work opens promising avenues for robust multilingual applications in complex, multi-step reasoning tasks. In addition, we release the code to foster research along this line.

Renxi Wang, Haonan Li, Minghao Wu et al.

Instruction tuning significantly enhances the performance of large language models (LLMs) across various tasks. However, the procedure to optimizing the mixing of instruction datasets for LLM fine-tuning is still poorly understood. This study categorizes instructions into three primary types: NLP downstream tasks, coding, and general chat. We explore the effects of instruction tuning on different combinations of datasets on LLM performance, and find that certain instruction types are more advantageous for specific applications but can negatively impact other areas. This work provides insights into instruction mixtures, laying the foundations for future research.

Weixuan Wang, Minghao Wu, Barry Haddow et al.

Large Language Models (LLMs) exhibit remarkable capabilities across various tasks, yet guiding them to follow desired behaviours during inference remains a significant challenge. Activation steering offers a promising method to control the generation process of LLMs by modifying their internal activations. However, existing methods commonly intervene in the model's behaviour using steering vectors generated by the model itself, which constrains their effectiveness to that specific model and excludes the possibility of leveraging powerful external expert models for steering. To address these limitations, we propose ExpertSteer, a novel approach that leverages arbitrary specialized expert models to generate steering vectors, enabling intervention in any LLMs. ExpertSteer transfers the knowledge from an expert model to a target LLM through a cohesive four-step process: first aligning representation dimensions with auto-encoders to enable cross-model transfer, then identifying intervention layer pairs based on mutual information analysis, next generating steering vectors from the expert model using Recursive Feature Machines, and finally applying these vectors on the identified layers during inference to selectively guide the target LLM without updating model parameters. We conduct comprehensive experiments using three LLMs on 15 popular benchmarks across four distinct domains. Experiments demonstrate that ExpertSteer significantly outperforms established baselines across diverse tasks at minimal cost.

Weixuan Wang, Minghao Wu, Barry Haddow et al.

Long document summarization remains a significant challenge for current large language models (LLMs), as existing approaches commonly struggle with information loss, factual inconsistencies, and coherence issues when processing excessively long documents. We propose SummQ, a novel adversarial multi-agent framework that addresses these limitations through collaborative intelligence between specialized agents operating in two complementary domains: summarization and quizzing. Our approach employs summary generators and reviewers that work collaboratively to create and evaluate comprehensive summaries, while quiz generators and reviewers create comprehension questions that serve as continuous quality checks for the summarization process. This adversarial dynamic, enhanced by an examinee agent that validates whether the generated summary contains the information needed to answer the quiz questions, enables iterative refinement through multifaceted feedback mechanisms. We evaluate SummQ on three widely used long document summarization benchmarks. Experimental results demonstrate that our framework significantly outperforms existing state-of-the-art methods across ROUGE and BERTScore metrics, as well as in LLM-as-a-Judge and human evaluations. Our comprehensive analyses reveal the effectiveness of the multi-agent collaboration dynamics, the influence of different agent configurations, and the impact of the quizzing mechanism. This work establishes a new approach for long document summarization that uses adversarial agentic collaboration to improve summarization quality.

Chang Hong, Minghao Wu, Qingying Xiao et al.

The integration of large language models into healthcare necessitates a rigorous evaluation of their ethical reasoning, an area current benchmarks often overlook. We introduce PrinciplismQA, a comprehensive benchmark with 3,648 questions designed to systematically assess LLMs' alignment with core medical ethics. Grounded in Principlism, our benchmark features a high-quality dataset. This includes multiple-choice questions curated from authoritative textbooks and open-ended questions sourced from authoritative medical ethics case study literature, all validated by medical experts. Our experiments reveal a significant gap between models' ethical knowledge and their practical application, especially in dynamically applying ethical principles to real-world scenarios. Most LLMs struggle with dilemmas concerning Beneficence, often over-emphasizing other principles. Frontier closed-source models, driven by strong general capabilities, currently lead the benchmark. Notably, medical domain fine-tuning can enhance models' overall ethical competence, but further progress requires better alignment with medical ethical knowledge. PrinciplismQA offers a scalable framework to diagnose these specific ethical weaknesses, paving the way for more balanced and responsible medical AI.

Weixuan Wang, Minghao Wu, Barry Haddow et al.

Fine-tuning large language models (LLMs) on a mixture of diverse datasets poses challenges due to data imbalance and heterogeneity. Existing methods often address these issues across datasets (globally) but overlook the imbalance and heterogeneity within individual datasets (locally), which limits their effectiveness. We introduce Hierarchical Balancing Optimization (HBO), a novel method that enables LLMs to autonomously adjust data allocation during fine-tuning both across datasets (globally) and within each individual dataset (locally). HBO employs a bilevel optimization strategy with two types of actors: a Global Actor, which balances data sampling across different subsets of the training mixture, and several Local Actors, which optimizes data usage within each subset based on difficulty levels. These actors are guided by reward functions derived from the LLM's training state, which measure learning progress and relative performance improvement. We evaluate HBO on three LLM backbones across nine diverse tasks in multilingual and multitask setups. Results show that HBO consistently outperforms existing baselines, achieving significant accuracy gains. Our in-depth analysis further demonstrates that both the global actor and local actors of HBO effectively adjust data usage during fine-tuning. HBO provides a comprehensive solution to the challenges of data imbalance and heterogeneity in LLM fine-tuning, enabling more effective training across diverse datasets.

Minghao Wu, Thuy-Trang Vu, Lizhen Qu et al.

The performance of large language models (LLMs) is strongly influenced by the quality and diversity of data used during supervised fine-tuning (SFT). However, current data selection methods often prioritize one aspect over the other, resulting in suboptimal training outcomes. To address this, we formulate data selection as a set cover problem and present GraphFilter, a novel approach that balances both quality and diversity in data selection. GraphFilter models the dataset as a bipartite graph connecting sentences to their constituent n-grams, then employs a priority function that combines quality and diversity metrics multiplicatively. GraphFilter iteratively selects sentences with the highest priority, removes covered n-grams from the bipartite graph, and recomputes priorities to reflect the changing data landscape. We validate GraphFilter using three model backbones across six widely-used benchmarks, demonstrating that it outperforms nine existing baselines in both model performance and computational efficiency. Further analysis shows that our design choices lead to more effective subset selection, underscores the value of instruction diversity, and provides insights into how quality and diversity interact with different subset sizes.

Huifang Du, Shuqin Li, Minghao Wu et al.

Reinforcement learning (RL) is a powerful approach to enhance task-oriented dialogue (TOD) systems. However, existing RL methods tend to mainly focus on generation tasks, such as dialogue policy learning (DPL) or response generation (RG), while neglecting dialogue state tracking (DST) for understanding. This narrow focus limits the systems to achieve globally optimal performance by overlooking the interdependence between understanding and generation. Additionally, RL methods face challenges with sparse and delayed rewards, which complicates training and optimization. To address these issues, we extend RL into both understanding and generation tasks by introducing step-by-step rewards throughout the token generation. The understanding reward increases as more slots are correctly filled in DST, while the generation reward grows with the accurate inclusion of user requests. Our approach provides a balanced optimization aligned with task completion. Experimental results demonstrate that our approach effectively enhances the performance of TOD systems and achieves new state-of-the-art results on three widely used datasets, including MultiWOZ2.0, MultiWOZ2.1, and In-Car. Our approach also shows superior few-shot ability in low-resource settings compared to current models.

Weixuan Wang, Barry Haddow, Minghao Wu et al.

Large language models (LLMs) have revolutionized the field of natural language processing (NLP), and recent studies have aimed to understand their underlying mechanisms. However, most of this research is conducted within a monolingual setting, primarily focusing on English. Few studies have attempted to explore the internal workings of LLMs in multilingual settings. In this study, we aim to fill this research gap by examining how neuron activation is shared across tasks and languages. We classify neurons into four distinct categories based on their responses to a specific input across different languages: all-shared, partial-shared, specific, and non-activated. Building upon this categorisation, we conduct extensive experiments on three tasks across nine languages using several LLMs and present an in-depth analysis in this work. Our findings reveal that: (i) deactivating the all-shared neurons significantly decreases performance; (ii) the shared neurons play a vital role in generating responses, especially for the all-shared neurons; (iii) neuron activation patterns are highly sensitive and vary across tasks, LLMs, and languages. These findings shed light on the internal workings of multilingual LLMs and pave the way for future research. We release the code to foster research in this area.

Minghao Wu, Thuy-Trang Vu, Lizhen Qu et al.

Large language models (LLMs) are typically fine-tuned on diverse and extensive datasets sourced from various origins to develop a comprehensive range of skills, such as writing, reasoning, chatting, coding, and more. Each skill has unique characteristics, and these datasets are often heterogeneous and imbalanced, making the fine-tuning process highly challenging. Balancing the development of each skill while ensuring the model maintains its overall performance requires sophisticated techniques and careful dataset curation. In this work, we propose a general, model-agnostic, reinforcement learning framework, Mixture-of-Skills (MoS), that learns to optimize data usage automatically during the fine-tuning process. This framework ensures the optimal comprehensive skill development of LLMs by dynamically adjusting the focus on different datasets based on their current learning state. To validate the effectiveness of MoS, we conduct extensive experiments using three diverse LLM backbones on two widely used benchmarks and demonstrate that MoS substantially enhances model performance. Building on the success of MoS, we propose MoSpec, an adaptation for task-specific fine-tuning, which harnesses the utilities of various datasets for a specific purpose. Our work underlines the significance of dataset rebalancing and present MoS as a powerful, general solution for optimizing data usage in the fine-tuning of LLMs for various purposes.

Chenyang Lyu, Zefeng Du, Jitao Xu et al.

Machine Translation (MT) has greatly advanced over the years due to the developments in deep neural networks. However, the emergence of Large Language Models (LLMs) like GPT-4 and ChatGPT is introducing a new phase in the MT domain. In this context, we believe that the future of MT is intricately tied to the capabilities of LLMs. These models not only offer vast linguistic understandings but also bring innovative methodologies, such as prompt-based techniques, that have the potential to further elevate MT. In this paper, we provide an overview of the significant enhancements in MT that are influenced by LLMs and advocate for their pivotal role in upcoming MT research and implementations. We highlight several new MT directions, emphasizing the benefits of LLMs in scenarios such as Long-Document Translation, Stylized Translation, and Interactive Translation. Additionally, we address the important concern of privacy in LLM-driven MT and suggest essential privacy-preserving strategies. By showcasing practical instances, we aim to demonstrate the advantages that LLMs offer, particularly in tasks like translating extended documents. We conclude by emphasizing the critical role of LLMs in guiding the future evolution of MT and offer a roadmap for future exploration in the sector.

Minghao Wu, Yitong Li, Meng Zhang et al.

Learning multilingual and multi-domain translation model is challenging as the heterogeneous and imbalanced data make the model converge inconsistently over different corpora in real world. One common practice is to adjust the share of each corpus in the training, so that the learning process is balanced and low-resource cases can benefit from the high resource ones. However, automatic balancing methods usually depend on the intra- and inter-dataset characteristics, which is usually agnostic or requires human priors. In this work, we propose an approach, MultiUAT, that dynamically adjusts the training data usage based on the model's uncertainty on a small set of trusted clean data for multi-corpus machine translation. We experiments with two classes of uncertainty measures on multilingual (16 languages with 4 settings) and multi-domain settings (4 for in-domain and 2 for out-of-domain on English-German translation) and demonstrate our approach MultiUAT substantially outperforms its baselines, including both static and dynamic strategies. We analyze the cross-domain transfer and show the deficiency of static and similarity based methods.

Minghao Wu, Fei Liu, Trevor Cohn

Conventional wisdom is that hand-crafted features are redundant for deep learning models, as they already learn adequate representations of text automatically from corpora. In this work, we test this claim by proposing a new method for exploiting handcrafted features as part of a novel hybrid learning approach, incorporating a feature auto-encoder loss component. We evaluate on the task of named entity recognition (NER), where we show that including manual features for part-of-speech, word shapes and gazetteers can improve the performance of a neural CRF model. We obtain a $F_1$ of 91.89 for the CoNLL-2003 English shared task, which significantly outperforms a collection of highly competitive baseline models. We also present an ablation study showing the importance of auto-encoding, over using features as either inputs or outputs alone, and moreover, show including the autoencoder components reduces training requirements to 60\%, while retaining the same predictive accuracy.