While Multimodal Large Language Models (MLLMs) have demonstrated impressive capabilities in general visual understanding, they frequently falter in fine-grained perception tasks that require identifying tiny objects or discerning subtle visual relationships. We attribute this limitation to Visual Attenuation: a phenomenon where sparse fine-grained visual signals are prematurely suppressed or diluted by dominant textual tokens during network propagation, resulting in a “loss of focus” during the deep-level decision-making process. Existing input-centric solutions fail to fundamentally reverse this intrinsic mechanism of information loss. To address this challenge, we propose the Variational Information Flow (VIF) framework. Adopting a probabilistic perspective, VIF leverages a Conditional Variational Autoencoder (CVAE) to model the visual saliency relevant to the question-answer pair as a latent distribution. As a plug-and-play module, VIF can be integrated into existing architectures. Extensive evaluations across diverse benchmarks—covering General VQA, fine-grained perception, and visual grounding—demonstrate that VIF yields competitive improvements over previous methods, validating its effectiveness in enhancing the fine-grained perception of MLLMs. Codes are available at https://github.com/ictnlp/VIF.

Empathy is essential for fostering natural interactions in spoken dialogue systems, as it enables machines to recognize the emotional tone of human speech and deliver empathetic responses. Recent research has made significant progress in developing empathetic spoken chatbots based on large language models (LLMs). However, several challenges still exist when training such models, including reliance on costly empathetic speech instruction data and a lack of emotional expressiveness in the generated speech. Finetuning LLM with cross-modal empathetic instruction data may also lead to catastrophic forgetting and a degradation of its general capability. To address these challenges, we propose FreezeEmpath, an end-to-end empathetic spoken chatbot trained in a simple and efficient manner. The entire training process relies solely on existing speech instruction data and speech emotion recognition (SER) data, while keeping the LLM’s parameters frozen. Experiments demonstrate that FreezeEmpath is able to generate emotionally expressive speech and outperforms other empathetic models in empathetic dialogue, SER, and SpokenQA tasks, demonstrating the effectiveness of our training strategy.

Speculative decoding has emerged as a promising approach to accelerate autoregressive inference in large language models (LLMs). Self-draft methods, which leverage the base LLM itself for speculation, avoid the overhead of auxiliary draft models but face critical limitations: shallow layers often produce overconfident yet incorrect token predictions, and the presence of difficult tokens in a draft sequence forces redundant computation through deeper layers, undermining both draft acceptance and overall speedup. To address these issues, we propose a novel self-draft framework that suppresses spurious confidence via layer-wise temperature annealing in early-exit decision and adaptively bounds speculation length based on token-wise decoding difficulty. By reprocessing the hidden states of draft tokens in a unified parallel pass through deep layers when speculation terminates, our method maintains exact output equivalence with the original model while maximizing computational efficiency. It requires no modifications to the base LLM parameters and achieves up to 2.33× wall-time speedup over standard autoregressive decoding across diverse long-form generation tasks and multiple model architectures.

Currently, large language models (LLMs) predominantly focus on the text modality. To enable more natural human-AI interaction, speech LLMs are emerging, but building effective end-to-end speech LLMs remains challenging due to limited data and the difficulty in expanding to more languages. In this paper, we introduce Cross-lingual Speech Language Model (CSLM), an efficient training method for cross-lingual speech LLMs based on discrete speech tokens. We propose a novel alignment strategy that achieves cross-modal and cross-lingual alignment through continual pre-training. By conducting instruction fine-tuning following a speech-text interleaved chain-of-modality generation process, we enhance modal alignment at a finer granularity, thereby improving generation quality and reducing latency. CSLM aligns different modalities and languages simultaneously without the need for massive speech data, thus exhibiting good language scalability. Evaluations on cross-modal tasks, mono-lingual conversational tasks, and cross-lingual conversational tasks demonstrate CSLM’s strong cross-modal alignment capabilities and general task abilities.

Streaming speech translation (StreamST) requires determining appropriate timing, known as policy, to generate translations while continuously receiving source speech inputs, balancing low latency with high translation quality. Existing StreamST methods typically operate on sentence-level speech segments, referred to as simultaneous speech translation (SimulST). In practice, they require collaboration with upstream segmentation models to accomplish StreamST, where the truncated speech segments constrain SimulST models to make policy decisions and generate translations based on pre-defined contextual information preset by the upstream models. Moreover, SimulST models struggle to learn effective policies due to the complexity of speech inputs and cross-lingual generation. To address these challenges, we propose StreamUni, which achieves StreamST through a unified Large Speech-Language Model (LSLM). Specifically, StreamUni incorporates speech Chain-of-Thought (CoT) in guiding the LSLM to generate multi-stage outputs. Leveraging these multi-stage outputs, StreamUni simultaneously accomplishes speech segmentation, policy decision, and translation generation. Additionally, we propose a streaming CoT training method that enhances low-latency policy decisions and generation capabilities using limited CoT data. Experiments demonstrate that our approach achieves state-of-the-art performance on both SimulST and StreamST tasks.

Long-context LLMs and Retrieval-Augmented Generation (RAG) systems expand accessible context, but still defer state tracking and evidence consolidation to query time. This passive paradigm is brittle in streaming settings where facts are updated and answers depend on latent states rather than explicit spans. We introduce Unified Memory Agent (UMA), an end-to-end reinforcement-learning framework that trains a single policy to maintain external memory and answer questions. UMA updates a structured memory bank via CRUD operations and combines it with raw-context retrieval for QA. To jointly optimize memory maintenance and question answering, we propose Task-Stratified GRPO, which normalizes heterogeneous trajectories separately for more effective credit assignment. We also introduce Ledger-QA, a controlled benchmark for long-horizon state tracking over accumulated updates. Experiments on Ledger-QA and 13 TTL/AR benchmarks show that UMA improves long-horizon tracking, achieves the best average TTL/AR score among compared methods, and remains competitive on retrieval-centric benchmarks. Our results highlight learned proactive memory management as a complement to longer context windows and query-time retrieval.

Large language models (LLMs) exhibit strong general intelligence, yet their multilingual performance remains highly imbalanced. Although LLMs encode substantial cross-lingual knowledge in a unified semantic space, they often struggle to reliably interface this knowledge with low-resource or unseen languages. Fortunately, pretrained encoder-decoder translation models already possess balanced multilingual capacity, suggesting a natural complement to LLMs. In this work, we propose XBridge, a compositional encoder-LLM-decoder architecture that offloads multilingual understanding and generation to external pretrained translation models, while preserving the LLM as an English-centric core for general knowledge processing and reasoning. To address the resulting representation misalignment across models, we introduce lightweight cross-model mapping layers together with an optimal transport-based alignment objective, enabling fine-grained semantic consistency for multilingual generation. Experiments on four LLMs across multilingual understanding, reasoning, and generation indicate that XBridge outperforms strong baselines, especially on low-resource and previously unseen languages, without retraining the LLM.

Byte-based machine translation systems have shown significant potential in massively multilingual settings. Unicode encoding, which maps each character to specific byte(s), eliminates the emergence of unknown words, even in new languages, enabling broad language scalability. However, byte-level tokenization results in sequences that are hard to interpret due to limited semantic information per byte. Local contextualization has proven effective in assigning initial semantics to tokens, improving sentence comprehension. Nevertheless, variations in encoding rules across languages necessitate an adaptive approach for effective contextualization. To this end, we propose Adaptive MultiScale-Headed Attention (Ada-MSHA), adaptively selecting and mixing attention heads, which are treated as contextualization experts. This enhances the flexibility of contextualization scales and improves the potential to discover a better strategy than previous methods. Experiment results show that our method outperforms existing methods without extensive manual adjustment of hyper-parameters and surpasses subword-based models with fewer parameters in Ted-59 dataset.

The remarkable understanding and generation capabilities of large language models (LLMs) have greatly improved translation performance. However, incorrect understanding of the sentence to be translated can degrade translation quality. To address this issue, we proposed a novel Iterative Bilingual Understanding Translation (IBUT) method based on the cross-lingual capabilities of LLMs and the dual characteristics of translation tasks. The cross-lingual capability of LLMs enables the generation of contextual understanding for both the source and target languages separately. Furthermore, the dual characteristics allow IBUT to generate effective cross-lingual feedback, iteratively refining contextual understanding, thereby reducing errors and improving translation performance. Experimental results showed that the proposed IBUT outperforms several strong comparison methods, especially being generalized to multiple domains (e.g., news, commonsense, and cultural translation benchmarks).

With the growing computational demands of large language models (LLMs), efficient inference has become increasingly critical for practical deployment. Depth pruning has emerged as a promising approach for reducing the computational costs of large language models by removing transformer layers. However, existing methods typically rely on fixed block masks, which can lead to suboptimal performance across different tasks and inputs. In this paper, we propose IG-Pruning, a novel input-aware block-wise pruning method that dynamically selects layer masks at inference time. Our approach consists of two stages: (1) Discovering diverse mask candidates through semantic clustering and L0 optimization, and (2) Implementing efficient dynamic pruning without the need for extensive training. Experimental results demonstrate that our method consistently outperforms state-of-the-art static depth pruning methods, making it particularly suitable for resource-constrained deployment scenarios.

Multilingual large language models (LLMs) possess impressive multilingual understanding and generation capabilities. However, their performance and cross-lingual alignment often lag for non-dominant languages. A common solution is to fine-tune LLMs on large-scale and more balanced multilingual corpus, but such approaches often lead to imprecise alignment and suboptimal knowledge transfer, struggling with limited improvements across languages. In this paper, we propose AlignX to bridge the multilingual performance gap, which is a two-stage representation-level framework for enhancing multilingual performance of pre-trained LLMs. In the first stage, we align multilingual representations with multilingual semantic alignment and language feature integration. In the second stage, we stimulate the multilingual capability of LLMs via multilingual instruction fine-tuning. Experimental results on several pre-trained LLMs demonstrate that our approach enhances LLMs’ multilingual general and cross-lingual generation capability. Further analysis indicates that AlignX brings the multilingual representations closer and improves the cross-lingual alignment.

Real-time, intelligent, and natural speech interaction is an essential part of the next-generation human-computer interaction. Recent advancements have showcased the potential of building intelligent spoken chatbots based on large language models (LLMs). In this paper, we introduce LLaMA-Omni 2, a series of speech language models (SpeechLMs) ranging from 0.5B to 14B parameters, capable of achieving high-quality real-time speech interaction. LLaMA-Omni 2 is built upon the Qwen2.5 series models, integrating a speech encoder and an autoregressive streaming speech decoder. Despite being trained on only 200K multi-turn speech dialogue samples, LLaMA-Omni 2 demonstrates strong performance on several spoken question answering and speech instruction following benchmarks, surpassing previous state-of-the-art SpeechLMs like GLM-4-Voice, which was trained on millions of hours of speech data.

Retrieval-Augmented Generation (RAG) plays a pivotal role in modern large language model applications, with numerous existing frameworks offering a wide range of functionalities to facilitate the development of RAG systems.However, we have identified several persistent challenges in these frameworks, including lack of new techniques, difficulties in algorithm reproduction and sharing, and high system overhead.To address these limitations, we introduce FlexRAG, an open-source framework specifically designed for research and prototyping.FlexRAG supports text-based, multimodal, and network-based RAG, providing comprehensive lifecycle support alongside efficient asynchronous processing and persistent caching capabilities.By offering a robust and flexible solution, FlexRAG enables researchers to rapidly develop, deploy, and share advanced RAG systems.Our toolkit and resources are available at https://github.com/ictnlp/FlexRAG.

Although Large Language Models (LLMs) have demonstrated impressive text generation capabilities, they are easily misled by untruthful contexts provided by users or knowledge augmentation tools, leading to hallucinations. To alleviate LLMs from being misled by untruthful context and take advantage of knowledge augmentation, we propose Truth-Aware Context Selection (TACS), a lightweight method to adaptively recognize and mask untruthful context from the inputs. TACS begins by performing truth detection on the input context, leveraging the parameterized knowledge within the LLM. Subsequently, it constructs a corresponding attention mask based on the truthfulness of each position, selecting the truthful context and discarding the untruthful context. Additionally, we introduce a new evaluation metric, Disturbance Adaption Rate, to further study the LLMs’ ability to accept truthful information and resist untruthful information.Experimental results indicate that TACS can effectively filter untruthful context and significantly improve the overall quality of LLMs’ responses when presented with misleading information.

The many-to-many multilingual neural machine translation can be regarded as the process of integrating semantic features from the source sentences and linguistic features from the target sentences. To enhance zero-shot translation, models need to share knowledge across languages, which can be achieved through auxiliary tasks for learning a universal representation or cross-lingual mapping. To this end, we propose to exploit both semantic and linguistic features between multiple languages to enhance multilingual translation. On the encoder side, we introduce a disentangling learning task that aligns encoder representations by disentangling semantic and linguistic features, thus facilitating knowledge transfer while preserving complete information. On the decoder side, we leverage a linguistic encoder to integrate low-level linguistic features to assist in the target language generation. Experimental results on multilingual datasets demonstrate significant improvement in zero-shot translation compared to the baseline system, while maintaining performance in supervised translation. Further analysis validates the effectiveness of our method in leveraging both semantic and linguistic features.

Subword tokenization is a common method for vocabulary building in Neural Machine Translation (NMT) models. However, increasingly complex tasks have revealed its disadvantages. First, a vocabulary cannot be modified once it is learned, making it hard to adapt to new words. Second, in multilingual translation, the imbalance in data volumes across different languages spreads to the vocabulary, exacerbating translations involving low-resource languages. While byte-based tokenization addresses these issues, byte-based models struggle with the low information density inherent in UTF-8 byte sequences. Previous works enhance token semantics through local contextualization but fail to select an appropriate contextualizing scope based on the input. Consequently, we propose the Multi-Scale Contextualization (MSC) method, which learns contextualized information of varying scales across different hidden state dimensions. It then leverages the attention module to dynamically integrate the multi-scale contextualized information. Experiments show that MSC significantly outperforms subword-based and other byte-based methods in both multilingual and out-of-domain scenarios. Code can be found in https://github.com/ictnlp/Multiscale-Contextualization.

Direct speech-to-speech translation (S2ST) has achieved impressive translation quality, but it often faces the challenge of slow decoding due to the considerable length of speech sequences. Recently, some research has turned to non-autoregressive (NAR) models to expedite decoding, yet the translation quality typically lags behind autoregressive (AR) models significantly. In this paper, we investigate the performance of CTC-based NAR models in S2ST, as these models have shown impressive results in machine translation. Experimental results demonstrate that by combining pretraining, knowledge distillation, and advanced NAR training techniques such as glancing training and non-monotonic latent alignments, CTC-based NAR models achieve translation quality comparable to the AR model, while preserving up to 26.81× decoding speedup.

Non-autoregressive neural machine translation (NAT) suffers from the multi-modality problem: the source sentence may have multiple correct translations, but the loss function is calculated only according to the reference sentence. Sequence-level knowledge distillation makes the target more deterministic by replacing the target with the output from an autoregressive model. However, the multi-modality problem in the distilled dataset is still nonnegligible. Furthermore, learning from a specific teacher limits the upper bound of the model capability, restricting the potential of NAT models. In this paper, we argue that one reference is not enough and propose diverse distillation with reference selection (DDRS) for NAT. Specifically, we first propose a method called SeedDiv for diverse machine translation, which enables us to generate a dataset containing multiple high-quality reference translations for each source sentence. During the training, we compare the NAT output with all references and select the one that best fits the NAT output to train the model. Experiments on widely-used machine translation benchmarks demonstrate the effectiveness of DDRS, which achieves 29.82 BLEU with only one decoding pass on WMT14 En-De, improving the state-of-the-art performance for NAT by over 1 BLEU.

Simultaneous machine translation (SiMT) outputs translation while receiving the streaming source inputs, and hence needs a policy to determine where to start translating. The alignment between target and source words often implies the most informative source word for each target word, and hence provides the unified control over translation quality and latency, but unfortunately the existing SiMT methods do not explicitly model the alignment to perform the control. In this paper, we propose Gaussian Multi-head Attention (GMA) to develop a new SiMT policy by modeling alignment and translation in a unified manner. For SiMT policy, GMA models the aligned source position of each target word, and accordingly waits until its aligned position to start translating. To integrate the learning of alignment into the translation model, a Gaussian distribution centered on predicted aligned position is introduced as an alignment-related prior, which cooperates with translation-related soft attention to determine the final attention. Experiments on En-Vi and De-En tasks show that our method outperforms strong baselines on the trade-off between translation and latency.

Regularization methods applying input perturbation have drawn considerable attention and have been frequently explored for NMT tasks in recent years. Despite their simplicity and effectiveness, we argue that these methods are limited by the under-fitting of training data. In this paper, we utilize prediction difference for ground-truth tokens to analyze the fitting of token-level samples and find that under-fitting is almost as common as over-fitting. We introduce prediction difference regularization (PD-R), a simple and effective method that can reduce over-fitting and under-fitting at the same time. For all token-level samples, PD-R minimizes the prediction difference between the original pass and the input-perturbed pass, making the model less sensitive to small input changes, thus more robust to both perturbations and under-fitted training data. Experiments on three widely used WMT translation tasks show that our approach can significantly improve over existing perturbation regularization methods. On WMT16 En-De task, our model achieves 1.80 SacreBLEU improvement over vanilla transformer.

How to learn a better speech representation for end-to-end speech-to-text translation (ST) with limited labeled data? Existing techniques often attempt to transfer powerful machine translation (MT) capabilities to ST, but neglect the representation discrepancy across modalities. In this paper, we propose the Speech-TExt Manifold Mixup (STEMM) method to calibrate such discrepancy. Specifically, we mix up the representation sequences of different modalities, and take both unimodal speech sequences and multimodal mixed sequences as input to the translation model in parallel, and regularize their output predictions with a self-learning framework. Experiments on MuST-C speech translation benchmark and further analysis show that our method effectively alleviates the cross-modal representation discrepancy, and achieves significant improvements over a strong baseline on eight translation directions.

Simultaneous machine translation (SiMT) starts translating while receiving the streaming source inputs, and hence the source sentence is always incomplete during translating. Different from the full-sentence MT using the conventional seq-to-seq architecture, SiMT often applies prefix-to-prefix architecture, which forces each target word to only align with a partial source prefix to adapt to the incomplete source in streaming inputs. However, the source words in the front positions are always illusoryly considered more important since they appear in more prefixes, resulting in position bias, which makes the model pay more attention on the front source positions in testing. In this paper, we first analyze the phenomenon of position bias in SiMT, and develop a Length-Aware Framework to reduce the position bias by bridging the structural gap between SiMT and full-sentence MT. Specifically, given the streaming inputs, we first predict the full-sentence length and then fill the future source position with positional encoding, thereby turning the streaming inputs into a pseudo full-sentence. The proposed framework can be integrated into most existing SiMT methods to further improve performance. Experiments on two representative SiMT methods, including the state-of-the-art adaptive policy, show that our method successfully reduces the position bias and thereby achieves better SiMT performance.

Multimodal machine translation (MMT) aims to improve neural machine translation (NMT) with additional visual information, but most existing MMT methods require paired input of source sentence and image, which makes them suffer from shortage of sentence-image pairs. In this paper, we propose a phrase-level retrieval-based method for MMT to get visual information for the source input from existing sentence-image data sets so that MMT can break the limitation of paired sentence-image input. Our method performs retrieval at the phrase level and hence learns visual information from pairs of source phrase and grounded region, which can mitigate data sparsity. Furthermore, our method employs the conditional variational auto-encoder to learn visual representations which can filter redundant visual information and only retain visual information related to the phrase. Experiments show that the proposed method significantly outperforms strong baselines on multiple MMT datasets, especially when the textual context is limited.

Simultaneous machine translation (SiMT) outputs translation while reading source sentence and hence requires a policy to decide whether to wait for the next source word (READ) or generate a target word (WRITE), the actions of which form a read/write path. Although the read/write path is essential to SiMT performance, no direct supervision is given to the path in the existing methods. In this paper, we propose a method of dual-path SiMT which introduces duality constraints to direct the read/write path. According to duality constraints, the read/write path in source-to-target and target-to-source SiMT models can be mapped to each other. As a result, the two SiMT models can be optimized jointly by forcing their read/write paths to satisfy the mapping. Experiments on En-Vi and De-En tasks show that our method can outperform strong baselines under all latency.

Neural networks tend to gradually forget the previously learned knowledge when learning multiple tasks sequentially from dynamic data distributions. This problem is called catastrophic forgetting, which is a fundamental challenge in the continual learning of neural networks. In this work, we observe that catastrophic forgetting not only occurs in continual learning but also affects the traditional static training. Neural networks, especially neural machine translation models, suffer from catastrophic forgetting even if they learn from a static training set. To be specific, the final model pays imbalanced attention to training samples, where recently exposed samples attract more attention than earlier samples. The underlying cause is that training samples do not get balanced training in each model update, so we name this problem imbalanced training. To alleviate this problem, we propose Complementary Online Knowledge Distillation (COKD), which uses dynamically updated teacher models trained on specific data orders to iteratively provide complementary knowledge to the student model. Experimental results on multiple machine translation tasks show that our method successfully alleviates the problem of imbalanced training and achieves substantial improvements over strong baseline systems.