Zewen Chi

Shuming Ma, Hongyu Wang, Shaohan Huang et al. · microsoft-research

Large Transformers have achieved state-of-the-art performance across many tasks. Most open-source libraries on scaling Transformers focus on improving training or inference with better parallelization. In this work, we present TorchScale, an open-source toolkit that allows researchers and developers to scale up Transformers efficiently and effectively. TorchScale has the implementation of several modeling techniques, which can improve modeling generality and capability, as well as training stability and efficiency. Experimental results on language modeling and neural machine translation demonstrate that TorchScale can successfully scale Transformers to different sizes without tears. The library is available at https://aka.ms/torchscale.

Shaohan Huang, Li Dong, Wenhui Wang et al. · cmu, microsoft-research

A big convergence of language, multimodal perception, action, and world modeling is a key step toward artificial general intelligence. In this work, we introduce Kosmos-1, a Multimodal Large Language Model (MLLM) that can perceive general modalities, learn in context (i.e., few-shot), and follow instructions (i.e., zero-shot). Specifically, we train Kosmos-1 from scratch on web-scale multimodal corpora, including arbitrarily interleaved text and images, image-caption pairs, and text data. We evaluate various settings, including zero-shot, few-shot, and multimodal chain-of-thought prompting, on a wide range of tasks without any gradient updates or finetuning. Experimental results show that Kosmos-1 achieves impressive performance on (i) language understanding, generation, and even OCR-free NLP (directly fed with document images), (ii) perception-language tasks, including multimodal dialogue, image captioning, visual question answering, and (iii) vision tasks, such as image recognition with descriptions (specifying classification via text instructions). We also show that MLLMs can benefit from cross-modal transfer, i.e., transfer knowledge from language to multimodal, and from multimodal to language. In addition, we introduce a dataset of Raven IQ test, which diagnoses the nonverbal reasoning capability of MLLMs.

Yaru Hao, Zewen Chi, Li Dong et al. · microsoft-research

Well-designed prompts can guide text-to-image models to generate amazing images. However, the performant prompts are often model-specific and misaligned with user input. Instead of laborious human engineering, we propose prompt adaptation, a general framework that automatically adapts original user input to model-preferred prompts. Specifically, we first perform supervised fine-tuning with a pretrained language model on a small collection of manually engineered prompts. Then we use reinforcement learning to explore better prompts. We define a reward function that encourages the policy to generate more aesthetically pleasing images while preserving the original user intentions. Experimental results on Stable Diffusion show that our method outperforms manual prompt engineering in terms of both automatic metrics and human preference ratings. Moreover, reinforcement learning further boosts performance, especially on out-of-domain prompts. The pretrained checkpoints are available at https://aka.ms/promptist. The demo can be found at https://aka.ms/promptist-demo.

Zewen Chi, Li Dong, Shaohan Huang et al. · cmu, microsoft-research

Sparse mixture of experts provides larger model capacity while requiring a constant computational overhead. It employs the routing mechanism to distribute input tokens to the best-matched experts according to their hidden representations. However, learning such a routing mechanism encourages token clustering around expert centroids, implying a trend toward representation collapse. In this work, we propose to estimate the routing scores between tokens and experts on a low-dimensional hypersphere. We conduct extensive experiments on cross-lingual language model pre-training and fine-tuning on downstream tasks. Experimental results across seven multilingual benchmarks show that our method achieves consistent gains. We also present a comprehensive analysis on the representation and routing behaviors of our models. Our method alleviates the representation collapse issue and achieves more consistent routing than the baseline mixture-of-experts methods.

Yaru Hao, Haoyu Song, Li Dong et al. · microsoft-research

Foundation models have received much attention due to their effectiveness across a broad range of downstream applications. Though there is a big convergence in terms of architecture, most pretrained models are typically still developed for specific tasks or modalities. In this work, we propose to use language models as a general-purpose interface to various foundation models. A collection of pretrained encoders perceive diverse modalities (such as vision, and language), and they dock with a language model that plays the role of a universal task layer. We propose a semi-causal language modeling objective to jointly pretrain the interface and the modular encoders. We subsume the advantages and capabilities from both causal and non-causal modeling, thereby combining the best of two worlds. Specifically, the proposed method not only inherits the capabilities of in-context learning and open-ended generation from causal language modeling, but also is conducive to finetuning because of the bidirectional encoders. More importantly, our approach seamlessly unlocks the combinations of the above capabilities, e.g., enabling in-context learning or instruction following with finetuned encoders. Experimental results across various language-only and vision-language benchmarks show that our model outperforms or is competitive with specialized models on finetuning, zero-shot generalization, and few-shot learning.

Barun Patra, Saksham Singhal, Shaohan Huang et al. · cmu, microsoft-research

In this paper, we elaborate upon recipes for building multilingual representation models that are not only competitive with existing state-of-the-art models but are also more parameter efficient, thereby promoting better adoption in resource-constrained scenarios and practical applications. We show that going beyond English-centric bitexts, coupled with a novel sampling strategy aimed at reducing under-utilization of training data, substantially boosts performance across model sizes for both Electra and MLM pre-training objectives. We introduce XY-LENT: X-Y bitext enhanced Language ENcodings using Transformers which not only achieves state-of-the-art performance over 5 cross-lingual tasks within all model size bands, is also competitive across bands. Our XY-LENT XL variant outperforms XLM-RXXL and exhibits competitive performance with mT5 XXL while being 5x and 6x smaller respectively. We then show that our proposed method helps ameliorate the curse of multilinguality, with the XY-LENT XL achieving 99.3% GLUE performance and 98.5% SQuAD 2.0 performance compared to a SoTA English only model in the same size band. We then analyze our models performance on extremely low resource languages and posit that scaling alone may not be sufficient for improving the performance in this scenario

Heqi Zheng, Xiao Zhang, Zewen Chi et al. · microsoft-research

Cross-lingual retrieval aims to retrieve relevant text across languages. Current methods typically achieve cross-lingual retrieval by learning language-agnostic text representations in word or sentence level. However, how to learn phrase representations for cross-lingual phrase retrieval is still an open problem. In this paper, we propose XPR, a cross-lingual phrase retriever that extracts phrase representations from unlabeled example sentences. Moreover, we create a large-scale cross-lingual phrase retrieval dataset, which contains 65K bilingual phrase pairs and 4.2M example sentences in 8 English-centric language pairs. Experimental results show that XPR outperforms state-of-the-art baselines which utilize word-level or sentence-level representations. XPR also shows impressive zero-shot transferability that enables the model to perform retrieval in an unseen language pair during training. Our dataset, code, and trained models are publicly available at www.github.com/cwszz/XPR/.

Xiao Zhang, Heyan Huang, Zewen Chi et al.

Conversational machine reading comprehension (CMRC) aims to assist computers to understand an natural language text and thereafter engage in a multi-turn conversation to answer questions related to the text. Existing methods typically require three steps: (1) decision making based on entailment reasoning; (2) span extraction if required by the above decision; (3) question rephrasing based on the extracted span. However, for nearly all these methods, the span extraction and question rephrasing steps cannot fully exploit the fine-grained entailment reasoning information in decision making step because of their relative independence, which will further enlarge the information gap between decision making and question phrasing. Thus, to tackle this problem, we propose a novel end-to-end framework for conversational machine reading comprehension based on shared parameter mechanism, called entailment reasoning T5 (ET5). Despite the lightweight of our proposed framework, experimental results show that the proposed ET5 achieves new state-of-the-art results on the ShARC leaderboard with the BLEU-4 score of 55.2. Our model and code are publicly available at https://github.com/Yottaxx/ET5.

Yuxiang Nie, Heyan Huang, Zewen Chi et al.

Unsupervised question answering is an attractive task due to its independence on labeled data. Previous works usually make use of heuristic rules as well as pre-trained models to construct data and train QA models. However, most of these works regard named entity (NE) as the only answer type, which ignores the high diversity of answers in the real world. To tackle this problem, we propose a novel unsupervised method by diversifying answers, named DiverseQA. Specifically, the proposed method is composed of three modules: data construction, data augmentation and denoising filter. Firstly, the data construction module extends the extracted named entity into a longer sentence constituent as the new answer span to construct a QA dataset with diverse answers. Secondly, the data augmentation module adopts an answer-type dependent data augmentation process via adversarial training in the embedding level. Thirdly, the denoising filter module is designed to alleviate the noise in the constructed data. Extensive experiments show that the proposed method outperforms previous unsupervised models on five benchmark datasets, including SQuADv1.1, NewsQA, TriviaQA, BioASQ, and DuoRC. Besides, the proposed method shows strong performance in the few-shot learning setting.

Xiao Zhang, Heyan Huang, Zewen Chi et al.

Open-retrieval conversational machine reading comprehension (OCMRC) simulates real-life conversational interaction scenes. Machines are required to make a decision of "Yes/No/Inquire" or generate a follow-up question when the decision is "Inquire" based on retrieved rule texts, user scenario, user question, and dialogue history. Recent studies explored the methods to reduce the information gap between decision-making and question generation and thus improve the performance of generation. However, the information gap still exists because these pipeline structures are still limited in decision-making, span extraction, and question rephrasing three stages. Decision-making and generation are reasoning separately, and the entailment reasoning utilized in decision-making is hard to share through all stages. To tackle the above problem, we proposed a novel one-stage end-to-end framework, called Entailment Fused-T5 (EFT), to bridge the information gap between decision-making and generation in a global understanding manner. The extensive experimental results demonstrate that our proposed framework achieves new state-of-the-art performance on the OR-ShARC benchmark.

Zewen Chi, Li Dong, Qingxiu Dong et al.

We envision a new era of AI, termed agentic organization, where agents solve complex problems by working collaboratively and concurrently, enabling outcomes beyond individual intelligence. To realize this vision, we introduce asynchronous thinking (AsyncThink) as a new paradigm of reasoning with large language models, which organizes the internal thinking process into concurrently executable structures. Specifically, we propose a thinking protocol where an organizer dynamically assigns sub-queries to workers, merges intermediate knowledge, and produces coherent solutions. More importantly, the thinking structure in this protocol can be further optimized through reinforcement learning. Experiments demonstrate that AsyncThink achieves 28% lower inference latency compared to parallel thinking while improving accuracy on mathematical reasoning. Moreover, AsyncThink generalizes its learned asynchronous thinking capabilities, effectively tackling unseen tasks without additional training.

Zhiliang Peng, Jianwei Yu, Yaoyao Chang et al.

This report presents VibeVoice-ASR, a general-purpose speech understanding framework built upon VibeVoice, designed to address the persistent challenges of context fragmentation and multi-speaker complexity in long-form audio (e.g., meetings, podcasts) that remain despite recent advancements in short-form speech recognition. Unlike traditional pipelined approaches that rely on audio chunking, VibeVoice-ASRsupports single-pass processing for up to 60 minutes of audio. It unifies Automatic Speech Recognition, Speaker Diarization, and Timestamping into a single end-to-end generation task. In addition, VibeVoice-ASR supports over 50 languages, requires no explicit language setting, and natively handles code-switching within and across utterances. Furthermore, we introduce a prompt-based context injection mechanism that allows users to supply customized conetxt, significantly improving accuracy on domain-specific terminology and polyphonic character disambiguation.

Jiaxin Guo, Zewen Chi, Li Dong et al.

Reward models play a critical role in guiding large language models toward outputs that align with human expectations. However, an open challenge remains in effectively utilizing test-time compute to enhance reward model performance. In this work, we introduce Reward Reasoning Models (RRMs), which are specifically designed to execute a deliberate reasoning process before generating final rewards. Through chain-of-thought reasoning, RRMs leverage additional test-time compute for complex queries where appropriate rewards are not immediately apparent. To develop RRMs, we implement a reinforcement learning framework that fosters self-evolved reward reasoning capabilities without requiring explicit reasoning traces as training data. Experimental results demonstrate that RRMs achieve superior performance on reward modeling benchmarks across diverse domains. Notably, we show that RRMs can adaptively exploit test-time compute to further improve reward accuracy. The pretrained reward reasoning models are available at https://huggingface.co/Reward-Reasoning.

Tianzhu Ye, Li Dong, Zewen Chi et al.

Black-box distillation creates student large language models (LLMs) by learning from a proprietary teacher model's text outputs alone, without access to its internal logits or parameters. In this work, we introduce Generative Adversarial Distillation (GAD), which enables on-policy and black-box distillation. GAD frames the student LLM as a generator and trains a discriminator to distinguish its responses from the teacher LLM's, creating a minimax game. The discriminator acts as an on-policy reward model that co-evolves with the student, providing stable, adaptive feedback. Experimental results show that GAD consistently surpasses the commonly used sequence-level knowledge distillation. In particular, Qwen2.5-14B-Instruct (student) trained with GAD becomes comparable to its teacher, GPT-5-Chat, on the LMSYS-Chat automatic evaluation. The results establish GAD as a promising and effective paradigm for black-box LLM distillation.

Hanyong Shao, Yingbo Hao, Ting Song et al.

NVIDIA's 2:4 Sparse Tensor Cores deliver 2x throughput but demand strict 50% pruning -- a ratio that collapses LLM reasoning accuracy (Qwen3: 54% to 15%). Milder $(2N-2):2N$ patterns (e.g., 6:8, 25% pruning) preserve accuracy yet receive no hardware support, falling back to dense execution without any benefit from sparsity. We present SlideSparse, the first system to unlock Sparse Tensor Core acceleration for the $(2N-2):2N$ model family on commodity GPUs. Our Sliding Window Decomposition reconstructs any $(2N-2):2N$ weight block into $N-1$ overlapping 2:4-compliant windows without any accuracy loss; Activation Lifting fuses the corresponding activation rearrangement into per-token quantization at near-zero cost. Integrated into vLLM, SlideSparse is evaluated across various GPUs (A100, H100, B200, RTX 4090, RTX 5080, DGX-spark), precisions (FP4, INT8, FP8, BF16, FP16), and model families (Llama, Qwen, BitNet). On compute-bound workloads, the measured speedup ratio (1.33x) approaches the theoretical upper-bound $N/(N-1)=4/3$ at 6:8 weight sparsity in Qwen2.5-7B, establishing $(2N-2):2N$ as a practical path to accuracy-preserving LLM acceleration. Code available at https://github.com/bcacdwk/vllmbench.

Di Zhang, Xun Wu, Shaohan Huang et al.

Semi-structured N:M sparsity and low-bit quantization (e.g., 1.58-bit BitNet) are two promising approaches for improving the efficiency of large language models (LLMs), yet they have largely been studied in isolation. In this work, we investigate their interaction and show that 1.58-bit BitNet is naturally more compatible with N:M sparsity than full-precision models. To study this effect, we propose Sparse-BitNet, a unified framework that jointly applies 1.58-bit quantization and dynamic N:M sparsification while ensuring stable training for the first time. Across multiple model scales and training regimes (sparse pretraining and dense-to-sparse schedules), 1.58-bit BitNet consistently exhibits smaller performance degradation than full-precision baselines at the same sparsity levels and can tolerate higher structured sparsity before accuracy collapse. Moreover, using our custom sparse tensor core, Sparse-BitNet achieves substantial speedups in both training and inference, reaching up to 1.30X. These results highlight that combining extremely low-bit quantization with semi-structured N:M sparsity is a promising direction for efficient LLMs. Code available at https://github.com/AAzdi/Sparse-BitNet

Yu Bai, Heyan Huang, Kai Fan et al.

Cross-Lingual Summarization (CLS) is a task that extracts important information from a source document and summarizes it into a summary in another language. It is a challenging task that requires a system to understand, summarize, and translate at the same time, making it highly related to Monolingual Summarization (MS) and Machine Translation (MT). In practice, the training resources for Machine Translation are far more than that for cross-lingual and monolingual summarization. Thus incorporating the Machine Translation corpus into CLS would be beneficial for its performance. However, the present work only leverages a simple multi-task framework to bring Machine Translation in, lacking deeper exploration. In this paper, we propose a novel task, Cross-lingual Summarization with Compression rate (CSC), to benefit Cross-Lingual Summarization by large-scale Machine Translation corpus. Through introducing compression rate, the information ratio between the source and the target text, we regard the MT task as a special CLS task with a compression rate of 100%. Hence they can be trained as a unified task, sharing knowledge more effectively. However, a huge gap exists between the MT task and the CLS task, where samples with compression rates between 30% and 90% are extremely rare. Hence, to bridge these two tasks smoothly, we propose an effective data augmentation method to produce document-summary pairs with different compression rates. The proposed method not only improves the performance of the CLS task, but also provides controllability to generate summaries in desired lengths. Experiments demonstrate that our method outperforms various strong baselines in three cross-lingual summarization datasets. We released our code and data at https://github.com/ybai-nlp/CLS_CR.

Zewen Chi, Li Dong, Bo Zheng et al.

The cross-lingual language models are typically pretrained with masked language modeling on multilingual text or parallel sentences. In this paper, we introduce denoising word alignment as a new cross-lingual pre-training task. Specifically, the model first self-labels word alignments for parallel sentences. Then we randomly mask tokens in a bitext pair. Given a masked token, the model uses a pointer network to predict the aligned token in the other language. We alternately perform the above two steps in an expectation-maximization manner. Experimental results show that our method improves cross-lingual transferability on various datasets, especially on the token-level tasks, such as question answering, and structured prediction. Moreover, the model can serve as a pretrained word aligner, which achieves reasonably low error rates on the alignment benchmarks. The code and pretrained parameters are available at https://github.com/CZWin32768/XLM-Align.

Houjin Yu, Xian-Ling Mao, Zewen Chi et al.

Recently, it has attracted much attention to build reliable named entity recognition (NER) systems using limited annotated data. Nearly all existing works heavily rely on domain-specific resources, such as external lexicons and knowledge bases. However, such domain-specific resources are often not available, meanwhile it's difficult and expensive to construct the resources, which has become a key obstacle to wider adoption. To tackle the problem, in this work, we propose a novel robust and domain-adaptive approach RDANER for low-resource NER, which only uses cheap and easily obtainable resources. Extensive experiments on three benchmark datasets demonstrate that our approach achieves the best performance when only using cheap and easily obtainable resources, and delivers competitive results against state-of-the-art methods which use difficultly obtainable domainspecific resources. All our code and corpora can be found on https://github.com/houking-can/RDANER.

Zewen Chi, Li Dong, Furu Wei et al.

In this work we focus on transferring supervision signals of natural language generation (NLG) tasks between multiple languages. We propose to pretrain the encoder and the decoder of a sequence-to-sequence model under both monolingual and cross-lingual settings. The pre-training objective encourages the model to represent different languages in the shared space, so that we can conduct zero-shot cross-lingual transfer. After the pre-training procedure, we use monolingual data to fine-tune the pre-trained model on downstream NLG tasks. Then the sequence-to-sequence model trained in a single language can be directly evaluated beyond that language (i.e., accepting multi-lingual input and producing multi-lingual output). Experimental results on question generation and abstractive summarization show that our model outperforms the machine-translation-based pipeline methods for zero-shot cross-lingual generation. Moreover, cross-lingual transfer improves NLG performance of low-resource languages by leveraging rich-resource language data. Our implementation and data are available at https://github.com/CZWin32768/xnlg.

Lingjie Jiang, Xun Wu, Shaohan Huang et al.

Recent Large Reasoning Models (LRMs) have shown substantially improved reasoning capabilities over traditional Large Language Models (LLMs) by incorporating extended thinking processes prior to producing final responses. However, excessively lengthy thinking introduces substantial overhead in terms of token consumption and latency, which is particularly unnecessary for simple queries. In this work, we introduce Large Hybrid-Reasoning Models (LHRMs), the first kind of model capable of adaptively determining whether to perform thinking based on the contextual information of user queries. To achieve this, we propose a two-stage training pipeline comprising Hybrid Fine-Tuning (HFT) as a cold start, followed by online reinforcement learning with the proposed Hybrid Group Policy Optimization (HGPO) to implicitly learn to select the appropriate thinking mode. Furthermore, we introduce a metric called Hybrid Accuracy to quantitatively assess the model's capability for hybrid thinking. Extensive experimental results show that LHRMs can adaptively perform hybrid thinking on queries of varying difficulty and type. It outperforms existing LRMs and LLMs in reasoning and general capabilities while significantly improving efficiency. Together, our work advocates for a reconsideration of the appropriate use of extended thinking processes and provides a solid starting point for building hybrid thinking systems.

Le Zhuo, Zewen Chi, Minghao Xu et al.

We propose ProtLLM, a versatile cross-modal large language model (LLM) for both protein-centric and protein-language tasks. ProtLLM features a unique dynamic protein mounting mechanism, enabling it to handle complex inputs where the natural language text is interspersed with an arbitrary number of proteins. Besides, we propose the protein-as-word language modeling approach to train ProtLLM. By developing a specialized protein vocabulary, we equip the model with the capability to predict not just natural language but also proteins from a vast pool of candidates. Additionally, we construct a large-scale interleaved protein-text dataset, named InterPT, for pre-training. This dataset comprehensively encompasses both (1) structured data sources like protein annotations and (2) unstructured data sources like biological research papers, thereby endowing ProtLLM with crucial knowledge for understanding proteins. We evaluate ProtLLM on classic supervised protein-centric tasks and explore its novel protein-language applications. Experimental results demonstrate that ProtLLM not only achieves superior performance against protein-specialized baselines on protein-centric tasks but also induces zero-shot and in-context learning capabilities on protein-language tasks.

Yaru Hao, Li Dong, Xun Wu et al.

Reinforcement learning algorithms are fundamental to align large language models with human preferences and to enhance their reasoning capabilities. However, current reinforcement learning algorithms often suffer from training instability due to loose on-policy constraints and computational inefficiency due to auxiliary models. In this work, we propose On-Policy RL with Optimal reward baseline (OPO), a novel and simplified reinforcement learning algorithm designed to address these challenges. OPO emphasizes the importance of exact on-policy training, which empirically stabilizes the training process and enhances exploration. Moreover, OPO integrates a practically feasible formulation of the optimal reward baseline that minimizes gradient variance. We evaluate OPO on mathematical reasoning benchmarks. The results demonstrate its superior performance and training stability without additional models or regularization terms. Furthermore, OPO achieves lower policy shifts and higher output entropy, encouraging more diverse and less repetitive responses. These results highlight OPO as a promising direction for stable and effective reinforcement learning in large language model alignment and reasoning tasks. The implementation is merged into the verl library at https://verl.readthedocs.io/en/latest/algo/opo.html.

Runheng Liu, Xingchen Xiao, Heyan Huang et al.

Retrieval-Augmented Language Modeling (RALM) by integrating large language models (LLM) with relevant documents from an external corpus is a proven method for enabling the LLM to generate information beyond the scope of its pre-training corpus. Previous work utilizing retrieved content by simply prepending it to the input poses a high runtime issue, which degrades the inference efficiency of the LLMs because they fail to use the Key-Value (KV) cache efficiently. In this paper, we propose FlashBack, a modular RALM designed to improve the inference efficiency of RALM with appending context pattern while maintaining decent performance after fine-tuning by Low-Rank Adaption. FlashBack appends retrieved documents at the end of the context for efficiently utilizing the KV cache instead of prepending them. And we introduce Marking Token as two special prompt tokens for marking the boundary of the appending context during fine-tuning. Our experiments on testing generation quality show that FlashBack can remain decent generation quality in perplexity. And the inference speed of FlashBack is up to $4\times$ faster than the prepending counterpart on a 7B LLM (Llama 2) in the runtime test. Via bypassing unnecessary re-computation, it demonstrates an advancement by achieving significantly faster inference speed, and this heightened efficiency will substantially reduce inferential cost.

Zongqian Li, Shaohan Huang, Zewen Chi et al.

Modern code generation models exhibit longer outputs, accelerated capability growth, and changed training dynamics, rendering traditional training methodologies, algorithms, and datasets ineffective for improving their performance. To address these training bottlenecks, we propose MicroCoder-GRPO, an improved Group Relative Policy Optimization approach with three innovations: conditional truncation masking to improve long output potential while maintaining training stability, diversity-determined temperature selection to maintain and encourage output diversity, and removal of KL loss with high clipping ratios to facilitate solution diversity. MicroCoder-GRPO achieves up to 17.6% relative improvement over strong baselines on LiveCodeBench v6, with more pronounced gains under extended context evaluation. Additionally, we release MicroCoder-Dataset, a more challenging training corpus that achieves 3x larger performance gains than mainstream datasets on LiveCodeBench v6 within 300 training steps, and MicroCoder-Evaluator, a robust framework with approximately 25% improved evaluation accuracy and around 40% faster execution. Through comprehensive analysis across more than thirty controlled experiments, we reveal 34 training insights across seven main aspects, demonstrating that properly trained models can achieve competitive performance with larger counterparts.

Di Zhang, Xun Wu, Shaohan Huang et al.

Recent advances in reinforcement learning (RL) have substantially improved the training of large-scale language models, leading to significant gains in generation quality and reasoning ability. However, most existing research focuses on dense models, while RL training for Mixture-of-Experts (MoE) architectures remains underexplored. To address the instability commonly observed in MoE training, we propose a novel router-aware approach to optimize importance sampling (IS) weights in off-policy RL. Specifically, we design a rescaling strategy guided by router logits, which effectively reduces gradient variance and mitigates training divergence. Experimental results demonstrate that our method significantly improves both the convergence stability and the final performance of MoE models, highlighting the potential of RL algorithmic innovations tailored to MoE architectures and providing a promising direction for efficient training of large-scale expert models.

Zewen Chi, Heyan Huang, Xian-Ling Mao

Recent studies have exhibited remarkable capabilities of pre-trained multilingual Transformers, especially cross-lingual transferability. However, current methods do not measure cross-lingual transferability well, hindering the understanding of multilingual Transformers. In this paper, we propose IGap, a cross-lingual transferability metric for multilingual Transformers on sentence classification tasks. IGap takes training error into consideration, and can also estimate transferability without end-task data. Experimental results show that IGap outperforms baseline metrics for transferability measuring and transfer direction ranking. Besides, we conduct extensive systematic experiments where we compare transferability among various multilingual Transformers, fine-tuning algorithms, and transfer directions. More importantly, our results reveal three findings about cross-lingual transfer, which helps us to better understand multilingual Transformers.

Zewen Chi, Heyan Huang, Luyang Liu et al.

The success of pretrained cross-lingual language models relies on two essential abilities, i.e., generalization ability for learning downstream tasks in a source language, and cross-lingual transferability for transferring the task knowledge to other languages. However, current methods jointly learn the two abilities in a single-phase cross-lingual pretraining process, resulting in a trade-off between generalization and cross-lingual transfer. In this paper, we propose cross-lingual language model meta-pretraining, which learns the two abilities in different training phases. Our method introduces an additional meta-pretraining phase before cross-lingual pretraining, where the model learns generalization ability on a large-scale monolingual corpus. Then, the model focuses on learning cross-lingual transfer on a multilingual corpus. Experimental results show that our method improves both generalization and cross-lingual transfer, and produces better-aligned representations across different languages.

Zewen Chi, Shaohan Huang, Li Dong et al.

In this paper, we introduce ELECTRA-style tasks to cross-lingual language model pre-training. Specifically, we present two pre-training tasks, namely multilingual replaced token detection, and translation replaced token detection. Besides, we pretrain the model, named as XLM-E, on both multilingual and parallel corpora. Our model outperforms the baseline models on various cross-lingual understanding tasks with much less computation cost. Moreover, analysis shows that XLM-E tends to obtain better cross-lingual transferability.

Bo Zheng, Li Dong, Shaohan Huang et al.

Fine-tuning pre-trained cross-lingual language models can transfer task-specific supervision from one language to the others. In this work, we propose to improve cross-lingual fine-tuning with consistency regularization. Specifically, we use example consistency regularization to penalize the prediction sensitivity to four types of data augmentations, i.e., subword sampling, Gaussian noise, code-switch substitution, and machine translation. In addition, we employ model consistency to regularize the models trained with two augmented versions of the same training set. Experimental results on the XTREME benchmark show that our method significantly improves cross-lingual fine-tuning across various tasks, including text classification, question answering, and sequence labeling.

Zewen Chi, Li Dong, Shuming Ma et al.

Multilingual T5 (mT5) pretrains a sequence-to-sequence model on massive monolingual texts, which has shown promising results on many cross-lingual tasks. In this paper, we improve multilingual text-to-text transfer Transformer with translation pairs (mT6). Specifically, we explore three cross-lingual text-to-text pre-training tasks, namely, machine translation, translation pair span corruption, and translation span corruption. In addition, we propose a partially non-autoregressive objective for text-to-text pre-training. We evaluate the methods on eight multilingual benchmark datasets, including sentence classification, named entity recognition, question answering, and abstractive summarization. Experimental results show that the proposed mT6 improves cross-lingual transferability over mT5.

Shuming Ma, Jian Yang, Haoyang Huang et al.

Multilingual machine translation enables a single model to translate between different languages. Most existing multilingual machine translation systems adopt a randomly initialized Transformer backbone. In this work, inspired by the recent success of language model pre-training, we present XLM-T, which initializes the model with an off-the-shelf pretrained cross-lingual Transformer encoder and fine-tunes it with multilingual parallel data. This simple method achieves significant improvements on a WMT dataset with 10 language pairs and the OPUS-100 corpus with 94 pairs. Surprisingly, the method is also effective even upon the strong baseline with back-translation. Moreover, extensive analysis of XLM-T on unsupervised syntactic parsing, word alignment, and multilingual classification explains its effectiveness for machine translation. The code will be at https://aka.ms/xlm-t.

Zewen Chi, Li Dong, Furu Wei et al.

In this work, we present an information-theoretic framework that formulates cross-lingual language model pre-training as maximizing mutual information between multilingual-multi-granularity texts. The unified view helps us to better understand the existing methods for learning cross-lingual representations. More importantly, inspired by the framework, we propose a new pre-training task based on contrastive learning. Specifically, we regard a bilingual sentence pair as two views of the same meaning and encourage their encoded representations to be more similar than the negative examples. By leveraging both monolingual and parallel corpora, we jointly train the pretext tasks to improve the cross-lingual transferability of pre-trained models. Experimental results on several benchmarks show that our approach achieves considerably better performance. The code and pre-trained models are available at https://aka.ms/infoxlm.

Heng-Da Xu, Xian-Ling Mao, Zewen Chi et al.

Recently, open-domain dialogue systems have attracted growing attention. Most of them use the sequence-to-sequence (Seq2Seq) architecture to generate responses. However, traditional Seq2Seq-based open-domain dialogue models tend to generate generic and safe responses, which are less informative, unlike human responses. In this paper, we propose a simple but effective keywords-guided Sequence-to-Sequence model (KW-Seq2Seq) which uses keywords information as guidance to generate open-domain dialogue responses. Specifically, KW-Seq2Seq first uses a keywords decoder to predict some topic keywords, and then generates the final response under the guidance of them. Extensive experiments demonstrate that the KW-Seq2Seq model produces more informative, coherent and fluent responses, yielding substantive gain in both automatic and human evaluation metrics.

Zewen Chi, Li Dong, Furu Wei et al.

Multilingual pretrained language models (such as multilingual BERT) have achieved impressive results for cross-lingual transfer. However, due to the constant model capacity, multilingual pre-training usually lags behind the monolingual competitors. In this work, we present two approaches to improve zero-shot cross-lingual classification, by transferring the knowledge from monolingual pretrained models to multilingual ones. Experimental results on two cross-lingual classification benchmarks show that our methods outperform vanilla multilingual fine-tuning.

Zewen Chi, Heyan Huang, Heng-Da Xu et al.

The task of table structure recognition aims to recognize the internal structure of a table, which is a key step to make machines understand tables. Currently, there are lots of studies on this task for different file formats such as ASCII text and HTML. It also attracts lots of attention to recognize the table structures in PDF files. However, it is hard for the existing methods to accurately recognize the structure of complicated tables in PDF files. The complicated tables contain spanning cells which occupy at least two columns or rows. To address the issue, we propose a novel graph neural network for recognizing the table structure in PDF files, named GraphTSR. Specifically, it takes table cells as input, and then recognizes the table structures by predicting relations among cells. Moreover, to evaluate the task better, we construct a large-scale table structure recognition dataset from scientific papers, named SciTSR, which contains 15,000 tables from PDF files and their corresponding structure labels. Extensive experiments demonstrate that our proposed model is highly effective for complicated tables and outperforms state-of-the-art baselines over a benchmark dataset and our new constructed dataset.