Boxing Chen

Yu Wan, Dayiheng Liu, Baosong Yang et al.

Translation quality evaluation plays a crucial role in machine translation. According to the input format, it is mainly separated into three tasks, i.e., reference-only, source-only and source-reference-combined. Recent methods, despite their promising results, are specifically designed and optimized on one of them. This limits the convenience of these methods, and overlooks the commonalities among tasks. In this paper, we propose UniTE, which is the first unified framework engaged with abilities to handle all three evaluation tasks. Concretely, we propose monotonic regional attention to control the interaction among input segments, and unified pretraining to better adapt multi-task learning. We testify our framework on WMT 2019 Metrics and WMT 2020 Quality Estimation benchmarks. Extensive analyses show that our \textit{single model} can universally surpass various state-of-the-art or winner methods across tasks. Both source code and associated models are available at https://github.com/NLP2CT/UniTE.

Yu Wan, Baosong Yang, Dayiheng Liu et al.

Attention mechanism has become the dominant module in natural language processing models. It is computationally intensive and depends on massive power-hungry multiplications. In this paper, we rethink variants of attention mechanism from the energy consumption aspects. After reaching the conclusion that the energy costs of several energy-friendly operations are far less than their multiplication counterparts, we build a novel attention model by replacing multiplications with either selective operations or additions. Empirical results on three machine translation tasks demonstrate that the proposed model, against the vanilla one, achieves competitable accuracy while saving 99\% and 66\% energy during alignment calculation and the whole attention procedure. Code is available at: https://github.com/NLP2CT/E-Att.

Yichao Du, Weizhi Wang, Zhirui Zhang et al. · tencent-ai

End-to-End Speech Translation (E2E-ST) has received increasing attention due to the potential of its less error propagation, lower latency, and fewer parameters. However, the effectiveness of neural-based approaches to this task is severely limited by the available training corpus, especially for domain adaptation where in-domain triplet training data is scarce or nonexistent. In this paper, we propose a novel non-parametric method that leverages domain-specific text translation corpus to achieve domain adaptation for the E2E-ST system. To this end, we first incorporate an additional encoder into the pre-trained E2E-ST model to realize text translation modelling, and then unify the decoder's output representation for text and speech translation tasks by reducing the correspondent representation mismatch in available triplet training data. During domain adaptation, a k-nearest-neighbor (kNN) classifier is introduced to produce the final translation distribution using the external datastore built by the domain-specific text translation corpus, while the universal output representation is adopted to perform a similarity search. Experiments on the Europarl-ST benchmark demonstrate that when in-domain text translation data is involved only, our proposed approach significantly improves baseline by 12.82 BLEU on average in all translation directions, even outperforming the strong in-domain fine-tuning method.

Dexin Wang, Kai Fan, Boxing Chen et al.

k-Nearest-Neighbor Machine Translation (kNN-MT) has been recently proposed as a non-parametric solution for domain adaptation in neural machine translation (NMT). It aims to alleviate the performance degradation of advanced MT systems in translating out-of-domain sentences by coordinating with an additional token-level feature-based retrieval module constructed from in-domain data. Previous studies have already demonstrated that non-parametric NMT is even superior to models fine-tuned on out-of-domain data. In spite of this success, kNN retrieval is at the expense of high latency, in particular for large datastores. To make it practical, in this paper, we explore a more efficient kNN-MT and propose to use clustering to improve the retrieval efficiency. Concretely, we first propose a cluster-based Compact Network for feature reduction in a contrastive learning manner to compress context features into 90+% lower dimensional vectors. We then suggest a cluster-based pruning solution to filter out 10%-40% redundant nodes in large datastores while retaining translation quality. Our proposed methods achieve better or comparable performance while reducing up to 57% inference latency against the advanced non-parametric MT model on several machine translation benchmarks. Experimental results indicate that the proposed methods maintain the most useful information of the original datastore and the Compact Network shows good generalization on unseen domains.

Deze Wang, Boxing Chen, Shanshan Li et al.

As pre-trained models automate many code intelligence tasks, a widely used paradigm is to fine-tune a model on the task dataset for each programming language. A recent study reported that multilingual fine-tuning benefits a range of tasks and models. However, we find that multilingual fine-tuning leads to performance degradation on recent models UniXcoder and CodeT5. To alleviate the potentially catastrophic forgetting issue in multilingual models, we fix all pre-trained model parameters, insert the parameter-efficient structure adapter, and fine-tune it. Updating only 0.6\% of the overall parameters compared to full-model fine-tuning for each programming language, adapter tuning yields consistent improvements on code search and summarization tasks, achieving state-of-the-art results. In addition, we experimentally show its effectiveness in cross-lingual and low-resource scenarios. Multilingual fine-tuning with 200 samples per programming language approaches the results fine-tuned with the entire dataset on code summarization. Our experiments on three probing tasks show that adapter tuning significantly outperforms full-model fine-tuning and effectively overcomes catastrophic forgetting.

Kexin Yang, Dayiheng Liu, Wenqiang Lei et al.

Attribute-based Controlled Text Generation (CTG) refers to generating sentences that satisfy desirable attributes (e.g., emotions and topics). Existing works often utilize fine-tuning or resort to extra attribute classifiers, yet suffer from storage and inference time increases. To address these concerns, we explore attribute-based CTG in a prompt-based manner. In short, the proposed Tailor represents each attribute as a pre-trained continuous vector (i.e., single-attribute prompt) and guides the generation of a fixed PLM switch to a pre-specified attribute. We experimentally find that these prompts can be simply concatenated as a whole to multi-attribute CTG without any re-training, yet raises problems of fluency decrease and position sensitivity. To this end, Tailor provides a multi-attribute prompt mask and a re-indexing position-ids sequence to bridge the gap between the training (one prompt for each task) and testing stage (concatenating more than one prompt). To further enhance such single-attribute prompt combinations, Tailor also introduces a trainable prompt connector, which can be concatenated with any two single-attribute prompts to multi-attribute text generation. Experiments on 11 attribute-specific generation tasks demonstrate strong performances of Tailor on both single-attribute and multi-attribute CTG, with 0.08\% training parameters of a GPT-2.

Yu Wan, Dayiheng Liu, Baosong Yang et al.

In this paper, we present our submission to Shared Metrics Task: RoBLEURT (Robustly Optimizing the training of BLEURT). After investigating the recent advances of trainable metrics, we conclude several aspects of vital importance to obtain a well-performed metric model by: 1) jointly leveraging the advantages of source-included model and reference-only model, 2) continuously pre-training the model with massive synthetic data pairs, and 3) fine-tuning the model with data denoising strategy. Experimental results show that our model reaching state-of-the-art correlations with the WMT2020 human annotations upon 8 out of 10 to-English language pairs.

Chengxi Li, Kai Fan, Jiajun Bu et al.

Song translation requires both translation of lyrics and alignment of music notes so that the resulting verse can be sung to the accompanying melody, which is a challenging problem that has attracted some interests in different aspects of the translation process. In this paper, we propose Lyrics-Melody Translation with Adaptive Grouping (LTAG), a holistic solution to automatic song translation by jointly modeling lyrics translation and lyrics-melody alignment. It is a novel encoder-decoder framework that can simultaneously translate the source lyrics and determine the number of aligned notes at each decoding step through an adaptive note grouping module. To address data scarcity, we commissioned a small amount of training data annotated specifically for this task and used large amounts of augmented data through back-translation. Experiments conducted on an English-Chinese song translation data set show the effectiveness of our model in both automatic and human evaluation.

Xiaochen Liu, Yang Gao, Yu Bai et al.

Few-shot abstractive summarization has become a challenging task in natural language generation. To support it, we designed a novel soft prompts architecture coupled with a prompt pre-training plus fine-tuning paradigm that is effective and tunes only extremely light parameters. The soft prompts include continuous input embeddings across an encoder and a decoder to fit the structure of the generation models. Importantly, a novel inner-prompt placed in the text is introduced to capture document-level information. The aim is to devote attention to understanding the document that better prompts the model to generate document-related content. The first step in the summarization procedure is to conduct prompt pre-training with self-supervised pseudo-data. This teaches the model basic summarizing capabilities. The model is then fine-tuned with few-shot examples. Experimental results on the CNN/DailyMail and XSum datasets show that our method, with only 0.1% of the parameters, outperforms full-model tuning where all model parameters are tuned. It also surpasses Prompt Tuning by a large margin and delivers competitive results against Prefix-Tuning with 3% of the parameters.

Chen Wang, Yuchen Liu, Boxing Chen et al.

End-to-end Speech Translation (ST) aims at translating the source language speech into target language text without generating the intermediate transcriptions. However, the training of end-to-end methods relies on parallel ST data, which are difficult and expensive to obtain. Fortunately, the supervised data for automatic speech recognition (ASR) and machine translation (MT) are usually more accessible, making zero-shot speech translation a potential direction. Existing zero-shot methods fail to align the two modalities of speech and text into a shared semantic space, resulting in much worse performance compared to the supervised ST methods. In order to enable zero-shot ST, we propose a novel Discrete Cross-Modal Alignment (DCMA) method that employs a shared discrete vocabulary space to accommodate and match both modalities of speech and text. Specifically, we introduce a vector quantization module to discretize the continuous representations of speech and text into a finite set of virtual tokens, and use ASR data to map corresponding speech and text to the same virtual token in a shared codebook. This way, source language speech can be embedded in the same semantic space as the source language text, which can be then transformed into target language text with an MT module. Experiments on multiple language pairs demonstrate that our zero-shot ST method significantly improves the SOTA, and even performers on par with the strong supervised ST baselines.

Biao Fu, Minpeng Liao, Kai Fan et al.

A popular approach to streaming speech translation is to employ a single offline model with a wait-k policy to support different latency requirements, which is simpler than training multiple online models with different latency constraints. However, there is a mismatch problem in using a model trained with complete utterances for streaming inference with partial input. We demonstrate that speech representations extracted at the end of a streaming input are significantly different from those extracted from a complete utterance. To address this issue, we propose a new approach called Future-Aware Streaming Translation (FAST) that adapts an offline ST model for streaming input. FAST includes a Future-Aware Inference (FAI) strategy that incorporates future context through a trainable masked embedding, and a Future-Aware Distillation (FAD) framework that transfers future context from an approximation of full speech to streaming input. Our experiments on the MuST-C EnDe, EnEs, and EnFr benchmarks show that FAST achieves better trade-offs between translation quality and latency than strong baselines. Extensive analyses suggest that our methods effectively alleviate the aforementioned mismatch problem between offline training and online inference.

Parsa Kavehzadeh, Mojtaba Valipour, Marzieh Tahaei et al.

Large language models (LLMs) have revolutionized natural language processing (NLP) by excelling at understanding and generating human-like text. However, their widespread deployment can be prohibitively expensive. SortedNet is a recent training technique for enabling dynamic inference by leveraging the modularity in networks and sorting sub-models based on computation/accuracy in a nested manner. We extend SortedNet to generative NLP tasks, making large language models dynamic without any Pre-Training and by only replacing Standard Fine-Tuning (SFT) with Sorted Fine-Tuning (SoFT). Our approach boosts model efficiency, eliminating the need for multiple models for various scenarios during inference. We show that this approach can unlock the power of intermediate layers of transformers in generating the target output. Our sub-models remain integral components of the original model, minimizing storage requirements and transition costs between different computational/latency budgets. The efficacy of our proposed method was demonstrated by applying it to tune LLaMA 2 13B on the Stanford Alpaca dataset for instruction following and TriviaQA for closed-book question answering. Our results show the superior performance of sub-models in comparison to Standard Fine-Tuning and SFT+ICT (Early-Exit), all achieved with efficient tuning and without additional memory usage during inference.

Yuan Ge, Yilun Liu, Chi Hu et al.

With contributions from the open-source community, a vast amount of instruction tuning (IT) data has emerged. Given the significant resource allocation required for training and evaluating models, it is advantageous to have an efficient method for selecting high-quality IT data. However, existing methods for instruction data selection have limitations such as relying on fragile external APIs, being affected by biases in GPT models, or reducing the diversity of the selected instruction dataset. In this paper, we propose an industrial-friendly, expert-aligned and diversity-preserved instruction data selection method: Clustering and Ranking (CaR). CaR employs a two-step process: first, it ranks instruction pairs using a high-accuracy (84.25%) scoring model aligned with expert preferences; second, it preserves dataset diversity through clustering. In our experiment, CaR efficiently selected a mere 1.96% of Alpaca's IT data, yet the resulting AlpaCaR model surpassed Alpaca's performance by an average of 32.1% in GPT-4 evaluations. Moreover, we find that data selecting is a consistent paradigm whether the pre-trained model is more capable or the model parameters scaling up. Our approach employs compact models with 550M parameters and incurs just 11.2% of the financial outlay of current methods, enhancing its industrial deployability.

Yilun Liu, Minggui He, Feiyu Yao et al.

The emergence of text-to-image synthesis (TIS) models has significantly influenced digital image creation by producing high-quality visuals from written descriptions. Yet these models are sensitive on textual prompts, posing a challenge for novice users who may not be familiar with TIS prompt writing. Existing solutions relieve this via automatic prompt expansion or generation from a user query. However, this single-turn manner suffers from limited user-centricity in terms of result interpretability and user interactivity. Thus, we propose DialPrompt, a dialogue-based TIS prompt generation model that emphasizes user experience for novice users. DialPrompt is designed to follow a multi-turn workflow, where in each round of dialogue the model guides user to express their preferences on possible optimization dimensions before generating the final TIS prompt. To achieve this, we mined 15 essential dimensions for high-quality prompts from advanced users and curated a multi-turn dataset. Through training on this dataset, DialPrompt improves user-centricity by allowing users to perceive and control the creation process of TIS prompts. Experiments indicate that DialPrompt improves significantly in user-centricity score compared with existing approaches while maintaining a competitive quality of synthesized images. In our user evaluation, DialPrompt is highly rated by 19 human reviewers (especially novices).

Vahid Partovi Nia, Guojun Zhang, Ivan Kobyzev et al.

Deep models are dominating the artificial intelligence (AI) industry since the ImageNet challenge in 2012. The size of deep models is increasing ever since, which brings new challenges to this field with applications in cell phones, personal computers, autonomous cars, and wireless base stations. Here we list a set of problems, ranging from training, inference, generalization bound, and optimization with some formalism to communicate these challenges with mathematicians, statisticians, and theoretical computer scientists. This is a subjective view of the research questions in deep learning that benefits the tech industry in long run.

Hossein Rajabzadeh, Aref Jafari, Aman Sharma et al.

Large Language Models (LLMs), with their increasing depth and number of parameters, have demonstrated outstanding performance across a variety of natural language processing tasks. However, this growth in scale leads to increased computational demands, particularly during inference and fine-tuning. To address these challenges, we introduce EchoAtt, a novel framework aimed at optimizing transformer-based models by analyzing and leveraging the similarity of attention patterns across layers. Our analysis reveals that many inner layers in LLMs, especially larger ones, exhibit highly similar attention matrices. By exploiting this similarity, EchoAtt enables the sharing of attention matrices in less critical layers, significantly reducing computational requirements without compromising performance. We incorporate this approach within a knowledge distillation setup, where a pre-trained teacher model guides the training of a smaller student model. The student model selectively shares attention matrices in layers with high similarity while inheriting key parameters from the teacher. Our best results with TinyLLaMA-1.1B demonstrate that EchoAtt improves inference speed by 15\%, training speed by 25\%, and reduces the number of parameters by approximately 4\%, all while improving zero-shot performance. These findings highlight the potential of attention matrix sharing to enhance the efficiency of LLMs, making them more practical for real-time and resource-limited applications.

Parsa Kavehzadeh, Mohammadreza Pourreza, Mojtaba Valipour et al.

Deployment of autoregressive large language models (LLMs) is costly, and as these models increase in size, the associated costs will become even more considerable. Consequently, different methods have been proposed to accelerate the token generation process and reduce costs. Speculative decoding (SD) is among the most promising approaches to speed up the LLM decoding process by verifying multiple tokens in parallel and using an auxiliary smaller draft model to generate the possible tokens. In SD, usually, one draft model is used to serve a specific target model; however, in practice, LLMs are diverse, and we might need to deal with many target models or more than one target model simultaneously. In this scenario, it is not clear which draft model should be used for which target model, and searching among different draft models or training customized draft models can further increase deployment costs. In this paper, we first introduce a novel multi-target scenario for the deployment of draft models for faster inference. Then, we present a novel, more efficient sorted speculative decoding mechanism that outperforms regular baselines in multi-target settings. We evaluated our method on Spec-Bench in different settings, including base models such as Vicuna 7B, 13B, and LLama Chat 70B. Our results suggest that our draft models perform better than baselines for multiple target models at the same time.

Pufan Zeng, Yilun Liu, Mingchen Dai et al.

Achieving cultural alignment in Large Language Models (LLMs) increasingly depends on synthetic data generation. For such synthesis, the most vital initial step is seed curation; however, current methods lack quantifiable standards for selecting these seeds. Existing approaches rely on unscalable manual curation or bias-prone LLM extraction, treating cultural specificity as an abstract concept rather than a measurable signal. In this paper, we address this "quantification gap" by proposing C-Mining, an unsupervised framework that transforms the discovery of cultural seeds from a subjective selection process into a computable data mining formulation. Our approach exploits a novel geometric insight, leveraging the cross-lingual misalignment of cultural concepts within pre-trained embedding spaces as a quantifiable discovery signal. By systematically identifying these regions characterized by pronounced linguistic exclusivity and geometric isolation, while actively filtering out noise, C-Mining automatically extracts high-fidelity Culture Points (CPs) from raw multilingual corpora without reliance on human or LLM supervision, reducing preparation costs by more than 150-fold. We further leverage the mined knowledge to steer the synthesis of diverse instruction-tuning datasets. Extensive experiments demonstrate that this seed-centric approach significantly enhances cultural understanding and reasoning capabilities, achieving a +6.03 point improvement on CulturalBench-Hard and surpassing state-of-the-art baselines, providing a scalable, quantifiable solution for high-quality cultural data synthesis.

Yuhe Ji, Yilun Liu, Feiyu Yao et al.

Log analysis represents a critical sub-domain within AI applications that facilitates automatic approaches to fault and error management of large-scaled software systems, saving labors of traditional manual methods. While existing solutions using large language models (LLMs) show promise, they are limited by a significant domain gap between natural and log languages (the latter contains rich domain-specific tokens such as status codes, IP addresses, resource pathes), which restricts their effectiveness in real-world applications. However, directly adapting general-purpose LLMs to log analysis using raw logs may degrade their performance due to inconsistent token distribution. In this paper, we present a domain adaptation approach that addresses these limitations by integrating interpretable domain knowledge into open-source LLMs through continual pre-training (CPT), which bridges this domain gap by adapting LLMs on interpretable natural texts with log knowledge (instead of raw logs) to reduce distribution discrepancy. To achieve this, we developed NLPLog, a comprehensive dataset containing over 250,000 question-answer pairs on log-related knowledge. Our resulting model, SuperLog, achieves the best performance across four log analysis tasks, with an average accuracy improvement of 12.01% over the second-best model. Ablation study also suggests advantages of domain adaption using interpretable log knowledge over using raw logs.

Nandan Thakur, Luiz Bonifacio, Xinyu Zhang et al.

Retrieval-Augmented Generation (RAG) grounds Large Language Model (LLM) output by leveraging external knowledge sources to reduce factual hallucinations. However, prior work lacks a comprehensive evaluation of different language families, making it challenging to evaluate LLM robustness against errors in external retrieved knowledge. To overcome this, we establish NoMIRACL, a human-annotated dataset for evaluating LLM robustness in RAG across 18 typologically diverse languages. NoMIRACL includes both a non-relevant and a relevant subset. Queries in the non-relevant subset contain passages judged as non-relevant, whereas queries in the relevant subset include at least a single judged relevant passage. We measure relevance assessment using: (i) hallucination rate, measuring model tendency to hallucinate, when the answer is not present in passages in the non-relevant subset, and (ii) error rate, measuring model inaccuracy to recognize relevant passages in the relevant subset.In our work, we observe that most models struggle to balance the two capacities. Models such as LLAMA-2 and Orca-2 achieve over 88% hallucination rate on the non-relevant subset. Mistral and LLAMA-3 hallucinate less but can achieve up to a 74.9% error rate on the relevant subset. Overall, GPT-4 is observed to provide the best tradeoff on both subsets, highlighting future work necessary to improve LLM robustness. NoMIRACL dataset and evaluation code are available at: https://github.com/project-miracl/nomiracl.

Peng Cheng, Jiucheng Zang, Qingnan Li et al.

Muon-style optimizers leverage Newton-Schulz (NS) iterations to orthogonalize updates, yielding update geometries that often outperform Adam-series methods. However, this orthogonalization discards magnitude information, rendering training sensitive to step-size hyperparameters and vulnerable to high-energy bursts. To mitigate this, we introduce TrasMuon (\textbf{T}rust \textbf{R}egion \textbf{A}daptive \textbf{S}caling \textbf{Muon}). TrasMuon preserves the near-isometric geometry of Muon while stabilizing magnitudes through (i) global RMS calibration and (ii) energy-based trust-region clipping. We demonstrate that while reintroducing adaptive scaling improves optimization efficiency, it typically exacerbates instability due to high-energy outliers. TrasMuon addresses this by defining a trust region based on relative energy ratios, confining updates to a stable zone. Empirical experiments on vision and language models demonstrate that TrasMuon converges faster than baselines. Furthermore, experiments without warmup stages confirm TrasMuon's superior stability and robustness.

Mohammad Dehghan, Mohammad Ali Alomrani, Sunyam Bagga et al.

The emerging citation-based QA systems are gaining more attention especially in generative AI search applications. The importance of extracted knowledge provided to these systems is vital from both accuracy (completeness of information) and efficiency (extracting the information in a timely manner). In this regard, citation-based QA systems are suffering from two shortcomings. First, they usually rely only on web as a source of extracted knowledge and adding other external knowledge sources can hamper the efficiency of the system. Second, web-retrieved contents are usually obtained by some simple heuristics such as fixed length or breakpoints which might lead to splitting information into pieces. To mitigate these issues, we propose our enhanced web and efficient knowledge graph (KG) retrieval solution (EWEK-QA) to enrich the content of the extracted knowledge fed to the system. This has been done through designing an adaptive web retriever and incorporating KGs triples in an efficient manner. We demonstrate the effectiveness of EWEK-QA over the open-source state-of-the-art (SoTA) web-based and KG baseline models using a comprehensive set of quantitative and human evaluation experiments. Our model is able to: first, improve the web-retriever baseline in terms of extracting more relevant passages (>20\%), the coverage of answer span (>25\%) and self containment (>35\%); second, obtain and integrate KG triples into its pipeline very efficiently (by avoiding any LLM calls) to outperform the web-only and KG-only SoTA baselines significantly in 7 quantitative QA tasks and our human evaluation.

Fengran Mo, Abbas Ghaddar, Kelong Mao et al.

In this paper, we study how open-source large language models (LLMs) can be effectively deployed for improving query rewriting in conversational search, especially for ambiguous queries. We introduce CHIQ, a two-step method that leverages the capabilities of LLMs to resolve ambiguities in the conversation history before query rewriting. This approach contrasts with prior studies that predominantly use closed-source LLMs to directly generate search queries from conversation history. We demonstrate on five well-established benchmarks that CHIQ leads to state-of-the-art results across most settings, showing highly competitive performances with systems leveraging closed-source LLMs. Our study provides a first step towards leveraging open-source LLMs in conversational search, as a competitive alternative to the prevailing reliance on commercial LLMs. Data, models, and source code will be publicly available upon acceptance at https://github.com/fengranMark/CHIQ.

Abbas Ghaddar, Philippe Langlais, Mehdi Rezagholizadeh et al.

Pretraining monolingual language models have been proven to be vital for performance in Arabic Natural Language Processing (NLP) tasks. In this paper, we conduct a comprehensive study on the role of data in Arabic Pretrained Language Models (PLMs). More precisely, we reassess the performance of a suite of state-of-the-art Arabic PLMs by retraining them on massive-scale, high-quality Arabic corpora. We have significantly improved the performance of the leading Arabic encoder-only BERT-base and encoder-decoder T5-base models on the ALUE and ORCA leaderboards, thereby reporting state-of-the-art results in their respective model categories. In addition, our analysis strongly suggests that pretraining data by far is the primary contributor to performance, surpassing other factors. Our models and source code are publicly available at https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/JABER-PyTorch.

Yichao Du, Zhirui Zhang, Weizhi Wang et al.

End-to-end speech-to-text translation (E2E-ST) is becoming increasingly popular due to the potential of its less error propagation, lower latency, and fewer parameters. Given the triplet training corpus $\langle speech, transcription, translation\rangle$, the conventional high-quality E2E-ST system leverages the $\langle speech, transcription\rangle$ pair to pre-train the model and then utilizes the $\langle speech, translation\rangle$ pair to optimize it further. However, this process only involves two-tuple data at each stage, and this loose coupling fails to fully exploit the association between triplet data. In this paper, we attempt to model the joint probability of transcription and translation based on the speech input to directly leverage such triplet data. Based on that, we propose a novel regularization method for model training to improve the agreement of dual-path decomposition within triplet data, which should be equal in theory. To achieve this goal, we introduce two Kullback-Leibler divergence regularization terms into the model training objective to reduce the mismatch between output probabilities of dual-path. Then the well-trained model can be naturally transformed as the E2E-ST models by the pre-defined early stop tag. Experiments on the MuST-C benchmark demonstrate that our proposed approach significantly outperforms state-of-the-art E2E-ST baselines on all 8 language pairs, while achieving better performance in the automatic speech recognition task. Our code is open-sourced at https://github.com/duyichao/E2E-ST-TDA.

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.

Weizhi Wang, Zhirui Zhang, Yichao Du et al.

Zero-shot translation, directly translating between language pairs unseen in training, is a promising capability of multilingual neural machine translation (NMT). However, it usually suffers from capturing spurious correlations between the output language and language invariant semantics due to the maximum likelihood training objective, leading to poor transfer performance on zero-shot translation. In this paper, we introduce a denoising autoencoder objective based on pivot language into traditional training objective to improve the translation accuracy on zero-shot directions. The theoretical analysis from the perspective of latent variables shows that our approach actually implicitly maximizes the probability distributions for zero-shot directions. On two benchmark machine translation datasets, we demonstrate that the proposed method is able to effectively eliminate the spurious correlations and significantly outperforms state-of-the-art methods with a remarkable performance. Our code is available at https://github.com/Victorwz/zs-nmt-dae.

Xin Zheng, Zhirui Zhang, Junliang Guo et al.

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

Hossein Rajabzadeh, Mojtaba Valipour, Tianshu Zhu et al.

Finetuning large language models requires huge GPU memory, restricting the choice to acquire Larger models. While the quantized version of the Low-Rank Adaptation technique, named QLoRA, significantly alleviates this issue, finding the efficient LoRA rank is still challenging. Moreover, QLoRA is trained on a pre-defined rank and, therefore, cannot be reconfigured for its lower ranks without requiring further fine-tuning steps. This paper proposes QDyLoRA -Quantized Dynamic Low-Rank Adaptation-, as an efficient quantization approach for dynamic low-rank adaptation. Motivated by Dynamic LoRA, QDyLoRA is able to efficiently finetune LLMs on a set of pre-defined LoRA ranks. QDyLoRA enables fine-tuning Falcon-40b for ranks 1 to 64 on a single 32 GB V100-GPU through one round of fine-tuning. Experimental results show that QDyLoRA is competitive to QLoRA and outperforms when employing its optimal rank.

Yilun Liu, Yuhe Ji, Shimin Tao et al.

Automatic log analysis is essential for the efficient Operation and Maintenance (O&M) of software systems, providing critical insights into system behaviors. However, existing approaches mostly treat log analysis as training a model to perform an isolated task ( e.g., anomaly detection, log parsing, etc.) using task-specific log-label pairs. These task-based approaches are inflexible in generalizing to complex scenarios, depend on task-specific training data, and cost significantly when deploying multiple models. In this paper, we propose an instruction-based training approach that transforms log-label pairs from multiple tasks and domains into a unified format of instruction-response pairs. Our trained model, LogLM, can follow complex user instructions and generalize better across different tasks, thereby increasing flexibility and reducing the dependence on task-specific training data. By integrating major log analysis tasks into a single model, our approach also relieves model deployment burden. Experimentally, LogLM outperforms existing approaches across five log analysis capabilities, and exhibits strong generalization abilities on complex instructions and unseen tasks.

Minggui He, Yilun Liu, Shimin Tao et al.

Despite recent breakthroughs in reasoning-enhanced large language models (LLMs) like DeepSeek-R1, incorporating inference-time reasoning into machine translation (MT), where human translators naturally employ structured, multi-layered reasoning chain-of-thoughts (CoTs), is yet underexplored. Existing methods either design a fixed CoT tailored for a specific MT sub-task (e.g., literature translation), or rely on synthesizing CoTs unaligned with humans and supervised fine-tuning (SFT) prone to overfitting, limiting their adaptability to diverse translation scenarios. This paper introduces R1-Translator (R1-T1), a novel framework to achieve inference-time reasoning for general MT via reinforcement learning (RL) with human-aligned CoTs comprising six common patterns. Our approach pioneers three innovations: (1) extending reasoning-based translation to broader MT scenarios (e.g., multilingual MT, domain MT) unseen in the training phase; (2) formalizing six expert-curated CoT templates that mirror hybrid human strategies like context-aware paraphrasing and back translation; and (3) enabling self-evolving CoT discovery through RL. Both human and automatic evaluation results indicate a steady translation performance improvement in a total of 10+ languages and 40+ translation directions on Flores-101 test set and four domain-specific MT tasks, especially on the languages unseen from training.

Qiuhao Zeng, Jerry Huang, Peng Lu et al.

Over recent years, the Transformer has become a fundamental building block for sequence modeling architectures. Yet at its core is the use of self-attention, whose memory and computational cost grow quadratically with the sequence length $N$, rendering it prohibitively expensive for long sequences. A promising approach is top-$k$ attention, which selects only the $k$ most relevant tokens and achieves performance comparable to vanilla self-attention while significantly reducing space and computational demands. However, causal masks require the current query token to only attend to past tokens, preventing the existing top-$k$ attention method from efficiently searching for the most relevant tokens in parallel, thereby limiting training efficiency. In this work, we propose ZETA, leveraging \textbf{Z}-Order Curves for \textbf{E}fficient \textbf{T}op-$k$ \textbf{A}ttention, to enable parallel querying of past tokens for entire sequences. % in both space and time complexity of $\mathcal{O}(N \log N)$. We first theoretically show that the choice of key and query dimensions involves a trade-off between the curse of dimensionality and the preservation of relative distances after projection. In light of this insight, we propose reducing the dimensionality of keys and queries in contrast to values and further leverage $Z$-order curves to map low-dimensional keys and queries into \emph{one}-dimensional space, which permits parallel sorting, thereby largely improving the efficiency for top-$k$ token selection. Experimental results demonstrate that ZETA matches the performance of standard attention on the synthetic \textsc{Multi-Query Associative Recall} task and outperforms attention and its variants on \textsc{Long Range Arena} and \textsc{WikiText-103} language modeling.

Peng Lu, Ivan Kobyzev, Mehdi Rezagholizadeh et al.

Recent advancements in Large Language Models (LLMs) have set themselves apart with their exceptional performance in complex language modelling tasks. However, these models are also known for their significant computational and storage requirements, primarily due to the quadratic computation complexity of softmax attention. To mitigate this issue, linear attention has been designed to reduce the quadratic space-time complexity that is inherent in standard transformers. In this work, we embarked on a comprehensive exploration of three key components that substantially impact the performance of the Gated Linear Attention module: feature maps, normalization, and the gating mechanism. We developed a feature mapping function to address some crucial issues that previous suggestions overlooked. Then we offered further rationale for the integration of normalization layers to stabilize the training process. Moreover, we explored the saturation phenomenon of the gating mechanism and augmented it with a refining module. We conducted extensive experiments and showed our architecture outperforms previous Gated Linear Attention mechanisms in extensive tasks including training from scratch and post-linearization with continual pre-training.

Jerry Huang, Prasanna Parthasarathi, Mehdi Rezagholizadeh et al. · mila

The growth in prominence of large language models (LLMs) in everyday life can be largely attributed to their generative abilities, yet some of this is also owed to the risks and costs associated with their use. On one front is their tendency to hallucinate false or misleading information, limiting their reliability. On another is the increasing focus on the computational limitations associated with traditional self-attention based LLMs, which has brought about new alternatives, in particular recurrent models, meant to overcome them. Yet it remains uncommon to consider these two concerns simultaneously. Do changes in architecture exacerbate/alleviate existing concerns about hallucinations? Do they affect how and where they occur? Through an extensive evaluation, we study how these architecture-based inductive biases affect the propensity to hallucinate. While hallucination remains a general phenomenon not limited to specific architectures, the situations in which they occur and the ease with which specific types of hallucinations can be induced can significantly differ based on the model architecture. These findings highlight the need for better understanding both these problems in conjunction with each other, as well as consider how to design more universal techniques for handling hallucinations.

Xinyu Wang, Linrui Ma, Jerry Huang et al. · mila

Recent shifts in the space of large language model (LLM) research have shown an increasing focus on novel architectures to compete with prototypical Transformer-based models that have long dominated this space. Linear recurrent models have proven to be a viable competitor due to their computational efficiency. However, such models still demonstrate a sizable gap compared to Transformers in terms of in-context learning among other tasks that require recalling information from a context. In this work, we introduce Resona, a simple and scalable framework for augmenting linear recurrent models with retrieval. Resona augments models with the ability to integrate retrieved information from the provided input context, enabling tailored behavior to diverse task requirements. Experiments on a variety of linear recurrent models demonstrate that Resona-augmented models observe significant performance gains on a variety of synthetic as well as real-world natural language tasks, highlighting its ability to act as a general purpose method to improve the in-context learning and language modeling abilities of linear recurrent LLMs.

Xinyu Wang, Vahid Partovi Nia, Peng Lu et al.

Large Language Models (LLMs) have demonstrated remarkable performance across various natural language processing (NLP) tasks. However, their deployment is challenging due to the substantial computational resources required. Power-of-two (PoT) quantization is a general tool to counteract this difficulty. Albeit previous works on PoT quantization can be efficiently dequantized on CPUs using fixed-point addition, it showed less effectiveness on GPUs. The reason is entanglement of the sign bit and sequential bit manipulations needed for dequantization. We propose a novel POT quantization framework for LLM weights that (i) outperforms state-of-the-art accuracy in extremely low-precision number formats, and (ii) enables faster inference through more efficient dequantization. To maintain the accuracy of the quantized model, we introduce a two-step post-training algorithm: (i) initialize the quantization scales with a robust starting point, and (ii) refine these scales using a minimal calibration set. The performance of our PoT post-training algorithm surpasses the current state-of-the-art in integer quantization, particularly at low precisions such as 2- and 3-bit formats. Our PoT quantization accelerates the dequantization step required for the floating point inference and leads to $3.67\times$ speed up on a NVIDIA V100, and $1.63\times$ on a NVIDIA RTX 4090, compared to uniform integer dequantization.

Alireza Ghaffari, Justin Yu, Mahsa Ghazvini Nejad et al.

Low-precision fine-tuning of language models has gained prominence as a cost-effective and energy-efficient approach to deploying large-scale models in various applications. However, this approach is susceptible to the existence of outlier values in activation. The outlier values in the activation can negatively affect the performance of fine-tuning language models in the low-precision regime since they affect the scaling factor and thus make representing smaller values harder. This paper investigates techniques for mitigating outlier activation in low-precision integer fine-tuning of the language models. Our proposed novel approach enables us to represent the outlier activation values in 8-bit integers instead of floating-point (FP16) values. The benefit of using integers for outlier values is that it enables us to use operator tiling to avoid performing 16-bit integer matrix multiplication to address this problem effectively. We provide theoretical analysis and supporting experiments to demonstrate the effectiveness of our approach in improving the robustness and performance of low-precision fine-tuned language models.

Ali Edalati, Alireza Ghaffari, Mahsa Ghazvini Nejad et al.

Deployment of Large Language Models (LLMs) has major computational costs, due to their rapidly expanding size. Compression of LLMs reduces the memory footprint, latency, and energy required for their inference. Post-training Quantization (PTQ) techniques have been developed to compress LLMs while avoiding expensive re-training. Most PTQ approaches formulate the quantization error based on a layer-wise Euclidean loss, ignoring the model output. Then, each layer is calibrated using its layer-wise Hessian to update the weights towards minimizing the quantization error. The Hessian is also used for detecting the most salient weights to quantization. Such PTQ approaches are prone to accuracy drop in low-precision quantization. We propose Output-adaptive Calibration (OAC) to incorporate the model output in the calibration process. We formulate the quantization error based on the distortion of the output cross-entropy loss. OAC approximates the output-adaptive Hessian for each layer under reasonable assumptions to reduce the computational complexity. The output-adaptive Hessians are used to update the weight matrices and detect the salient weights towards maintaining the model output. Our proposed method outperforms the state-of-the-art baselines such as SpQR and BiLLM, especially, at extreme low-precision (2-bit and binary) quantization.

Abbas Ghaddar, Ivan Kobyzev, Boxing Chen et al.

Post-training hybridization of large language models (LLMs) often replaces quadratic self-attention with sliding-window attention (SWA) to reduce KV cache usage and improve latency. Existing hybridization schemes are typically defined either at the layer level (e.g., interleaving) or at the head level via static rankings from local to global. Layer-level schemes ignore that local and global dependencies are routed through heads within the same layer, while static head-level rankings suffer from entanglement: a head's local/global behavior can change after hybridization. We propose BOSCH, Black-box Binary Optimization for Short-context Head Selection, a training-free method that formulates the problem as a Large Neighborhood Search and decomposes it into three subproblems: (i) layer-importance detection via small-budget black-box probes, (ii) adaptive per-layer SWA-ratio assignment based on these sensitivities, and (iii) grouped head-level optimization within ratio buckets. Extensive experiments on 4 LLMs ranging from 1.7B to 30B parameters, across 4 SWA ratios, show that BOSCH consistently outperforms layer-level heuristics and 6 strong static head-level methods, with larger gains at higher SWA ratios. Under continual pretraining, BOSCH recover original long-context performance faster and to a higher level. Analysis of the selected heads reveals substantial turnover for BOSCH across different SWA ratios, underscoring the importance of performing head-level selection for each target ratio rather than relying on fixed locality rankings.

Song Xu, Yilun Liu, Minggui He et al.

Logs constitute a form of evidence signaling the operational status of software systems. Automated log anomaly detection is crucial for ensuring the reliability of modern software systems. However, existing approaches face significant limitations: traditional deep learning models lack interpretability and generalization, while methods leveraging Large Language Models are often hindered by unreliability and factual inaccuracies. To address these issues, we propose RationAnomaly, a novel framework that enhances log anomaly detection by synergizing Chain-of-Thought (CoT) fine-tuning with reinforcement learning. Our approach first instills expert-like reasoning patterns using CoT-guided supervised fine-tuning, grounded in a high-quality dataset corrected through a rigorous expert-driven process. Subsequently, a reinforcement learning phase with a multi-faceted reward function optimizes for accuracy and logical consistency, effectively mitigating hallucinations. Experimentally, RationAnomaly outperforms state-of-the-art baselines, achieving superior F1-scores on key benchmarks while providing transparent, step-by-step analytical outputs. We have released the corresponding resources, including code and datasets.

Aman Sharma, Saeed Najafi, Parsa Farinneya et al.

Transformers achieve state-of-the-art results across many tasks, but their uniform application of quadratic self-attention to every token at every layer makes them computationally expensive. We introduce DTRNet (Dynamic Token Routing Network), an improved Transformer architecture that allows tokens to dynamically skip the quadratic cost of cross-token mixing while still receiving lightweight linear updates. By preserving the MLP module and reducing the attention cost for most tokens to linear, DTRNet ensures that every token is explicitly updated while significantly lowering overall computation. This design offers an efficient and effective alternative to standard dense attention. Once trained, DTRNet blocks routes only ~10% of tokens through attention at each layer while maintaining performance comparable to a full Transformer. It consistently outperforms routing-based layer skipping methods such as MoD and D-LLM in both accuracy and memory at matched FLOPs, while routing fewer tokens to full attention. Its efficiency gains, scales with sequence length, offering significant reduction in FLOPs for long-context inputs. By decoupling token updates from attention mixing, DTRNet substantially reduces the quadratic share of computation, providing a simple, efficient, and scalable alternative to Transformers.

Yan Yu, Yilun Liu, Minggui He et al.

Pairwise evaluation of large language models (LLMs) has become the dominant paradigm for benchmarking open-ended tasks, yet non-transitive preferences, where evaluators prefer A over B, B over C, but C over A, fundamentally undermine ranking reliability. We show that this critical issue stems largely from low-quality data that contains inherently ambiguous preference pairs. To address this challenge, we propose ELSPR, a principled graph-theoretic framework that models pairwise preferences as tournament graphs and systematically identifies problematic training data. ELSPR quantifies non-transitivity through strongly connected components (SCCs) analysis and measures overall preference clarity using a novel normalized directed graph structural entropy metric. Our filtering methodology selectively removes preference data that induce non-transitivity while preserving transitive preferences. Extensive experiments on the AlpacaEval benchmark demonstrate that models fine-tuned on ELSPR-filtered data achieve substantial improvements: a 13.8% reduction in non-transitivity, a 0.088 decrease in structural entropy, and significantly enhanced discriminative power in real-world evaluation systems. Human validation confirms that discarded data exhibit dramatically lower inter-annotator agreement (34.4% vs. 52.6%) and model-human consistency (51.2% vs. 80.6%) compared to cleaned data. These findings establish ELSPR as an effective data self-purification approach for developing more robust, consistent, and human-aligned LLM evaluation systems.

Alireza Ghaffari, Sharareh Younesian, Vahid Partovi Nia et al.

The ever-growing computational complexity of Large Language Models (LLMs) necessitates efficient deployment strategies. The current state-of-the-art approaches for Post-training Quantization (PTQ) often require calibration to achieve the desired accuracy. This paper presents AdpQ, a novel zero-shot adaptive PTQ method for LLMs that achieves the state-of-the-art performance in low-precision quantization (e.g. 3-bit) without requiring any calibration data. Inspired by Adaptive LASSO regression model, our proposed approach tackles the challenge of outlier activations by separating salient weights using an adaptive soft-thresholding method. Guided by Adaptive LASSO, this method ensures that the quantized weights distribution closely follows the originally trained weights and eliminates the need for calibration data entirely, setting our method apart from popular approaches such as SpQR and AWQ. Furthermore, our method offers an additional benefit in terms of privacy preservation by eliminating any calibration or training data. We also delve deeper into the information-theoretic underpinnings of the proposed method. We demonstrate that it leverages the Adaptive LASSO to minimize the Kullback-Leibler divergence between the quantized weights and the originally trained weights. This minimization ensures the quantized model retains the Shannon information content of the original model to a great extent, guaranteeing efficient deployment without sacrificing accuracy or information. Our results achieve the same accuracy as the existing methods on various LLM benchmarks while the quantization time is reduced by at least 10x, solidifying our contribution to efficient and privacy-preserving LLM deployment.

Michael R. Metel, Yufei Cui, Boxing Chen et al.

Sequential test-time scaling is a promising training-free method to improve large reasoning model accuracy, but as currently implemented, significant limitations have been observed. Inducing models to think for longer can increase their accuracy, but as the length of reasoning is further extended, it has also been shown to result in accuracy degradation and model instability. This work presents a novel sequential test-time scaling method, Min-Seek, which improves model accuracy significantly over a wide range of induced thoughts, stabilizing the accuracy of sequential scaling, and removing the need for reasoning length fine-tuning. Beyond improving model accuracy over a variety of reasoning tasks, our method is inherently efficient, as only the KV pairs of one additional induced thought are kept in the KV cache during reasoning. With a custom KV cache which stores keys without position embeddings, by dynamically encoding them contiguously before each new generated thought, our method can continue to reason well beyond a model's maximum context length, and under mild conditions has linear computational complexity.

Rezaul Karim, Maryam Dialameh, Yang Liu et al.

We present a novel method for Efficient training with Progressive Activation Sharing (EPAS). This method bridges progressive training paradigm with the phenomenon of redundant QK (or KV ) activations across deeper layers of transformers. EPAS gradually grows a sharing region during training by switching decoder layers to activation sharing mode. This results in throughput increase due to reduced compute. To utilize deeper layer redundancy, the sharing region starts from the deep end of the model and grows towards the shallow end. The EPAS trained models allow for variable region lengths of activation sharing for different compute budgets during inference. Empirical evaluations with QK activation sharing in LLaMA models ranging from 125M to 7B parameters show up to an 11.1% improvement in training throughput and up to a 29% improvement in inference throughput while maintaining similar loss curve to the baseline models. Furthermore, applying EPAS in continual pretraining to transform TinyLLaMA into an attention-sharing model yields up to a 10% improvement in average accuracy over state-of-the-art methods, emphasizing the significance of progressive training in cross layer activation sharing models.

Hossam Amer, Rezaul Karim, Ali Pourranjbar et al.

With the rapid growth of large language models (LLMs), a wide range of methods have been developed to distribute computation and memory across hardware devices for efficient training and inference. While existing surveys provide descriptive overviews of these techniques, systematic analysis of their benefits and trade offs and how such insights can inform principled methodology for designing optimal distributed systems remain limited. This paper offers a comprehensive review of collective operations and distributed parallel strategies, complemented by mathematical formulations to deepen theoretical understanding. We further examine hybrid parallelization designs, emphasizing communication computation overlap across different stages of model deployment, including both training and inference. Recent advances in automated search for optimal hybrid parallelization strategies using cost models are also discussed. Moreover, we present case studies with mainstream architecture categories to reveal empirical insights to guide researchers and practitioners in parallelism strategy selection. Finally, we highlight open challenges and limitations of current LLM training paradigms and outline promising directions for the next generation of large scale model development.

Prasanna Parthasarathi, Mathieu Reymond, Boxing Chen et al. · mila

Large language models (LLMs) are increasingly deployed for tasks requiring complex reasoning, prompting significant interest in improving their reasoning abilities through post-training. Especially RL based methods using verifiable reward, like the state-of-the-art GRPO, have shown to tremendously improve reasoning behaviors when applied as post-training methods. However, the lack of an explicit reward or critic model limits GRPO's ability to assign fine-grained credit across token sequences. In this work, we present GRPO-$λ$, a novel extension to GRPO that enhances credit assignment in RL finetuning of LLMs for complex reasoning tasks. We approximate learning from $λ$-return with a reformulation of eligibility traces using token-level log-probabilities applied after each sequence generation, and a novel critic-free approximation of the temporal-difference error. We introduce a few variations for the weighting of the $λ$-return, and their applications to the eligibility-trace, where all the variations provide significant gains over GRPO. We compare GRPO-$λ$ against GRPO by training models from 1.5B to 7B parameters on $4$ different math reasoning datasets. The training plots demonstrate 30-40% improved performance during RL training on both LLaMA-3.1 and Qwen-2.5 architectures. Finally, we show that with GRPO-$λ$, the resulting average performance on AIME24, Math500, OlympiadMath, MinervaMath, and AMC improves over GRPO by over $3$ points and a $4.5$ points improvement on the 7B model.

Peng Lu, Jerry Huang, Qiuhao Zeng et al.

The quadratic complexity of the attention mechanism in Transformer models has motivated the development of alternative architectures with sub-quadratic scaling, such as state-space models. Among these, Mamba has emerged as a leading architecture, achieving state-of-the-art results across a range of language modeling tasks. However, Mamba's performance significantly deteriorates when applied to contexts longer than those seen during pre-training, revealing a sharp sensitivity to context length extension. Through detailed analysis, we attribute this limitation to the out-of-distribution behaviour of its state-space dynamics, particularly within the parameterization of the state transition matrix $\mathbf{A}$. Unlike recent works which attribute this sensitivity to the vanished accumulation of discretization time steps, $\exp(-\sum_{t=1}^NΔ_t)$, we establish a connection between state convergence behavior as the input length approaches infinity and the spectrum of the transition matrix $\mathbf{A}$, offering a well-founded explanation of its role in length extension. Next, to overcome this challenge, we propose an approach that applies spectrum scaling to pre-trained Mamba models to enable robust long-context generalization by selectively modulating the spectrum of $\mathbf{A}$ matrices in each layer. We show that this can significantly improve performance in settings where simply modulating $Δ_t$ fails, validating our insights and providing avenues for better length generalization of state-space models with structured transition matrices.

Chunguang Zhao, Yilun Liu, Pufan Zeng et al.

Multilingual Instruction Fine-Tuning (IFT) is essential for enabling large language models (LLMs) to generalize effectively across diverse linguistic and cultural contexts. However, the scarcity of high-quality multilingual training data and corresponding building method remains a critical bottleneck. While data selection has shown promise in English settings, existing methods often fail to generalize across languages due to reliance on simplistic heuristics or language-specific assumptions. In this work, we introduce Multilingual Data Quality and Diversity (M-DaQ), a novel method for improving LLMs multilinguality, by selecting high-quality and semantically diverse multilingual IFT samples. We further conduct the first systematic investigation of the Superficial Alignment Hypothesis (SAH) in multilingual setting. Empirical results across 18 languages demonstrate that models fine-tuned with M-DaQ method achieve significant performance gains over vanilla baselines over 60% win rate. Human evaluations further validate these gains, highlighting the increment of cultural points in the response. We release the M-DaQ code to support future research.

Aref Jafari, Yuhe Fan, Benyamin Jamialahmadi et al.

Transformers have demonstrated strong performance across a wide range of sequence modeling tasks, but their quadratic attention complexity limits scalability to long sequences. Linear models such as Mamba and sliding-window attention (SWA) address this by mixing tokens through recurrent or localized operations with fixed-size memory, achieving efficient inference. However, these methods risk degrading performance on long sequences due to their inability to retain detailed information from distant tokens. We propose SCOUT (Segment Compression for Optimized Utility in Transformers), a hybrid architecture that compresses tokens locally within fixed-size segments and applies attention only over these compressed representations. Each token embedding is first enriched via a linear local mixer, Mamba or SWA, that integrates recent context. Then, instead of attending to all previous tokens, each token sparsely attends to a small number of compressed checkpoint tokens that summarize the input history. This design retains much of the expressivity of full attention while substantially reducing the computational and memory cost. By attending to compressed history rather than all previous tokens, SCOUT incurs slightly higher memory than purely linear models, but its growth rate remains sub-quadratic and far more scalable than that of full Transformers. We analyze SCOUT's computational and memory efficiency and evaluate it empirically on long-context language modeling and reasoning tasks. SCOUT with both Mamba and SWA mixers outperforms strong long-sequence baselines under the same computational budget, matches full-attention Transformers on language modeling and common-sense reasoning tasks at 400M and 1.3B scales. Moreover, our SCOUT achieves higher end-to-end throughput than SOTA models, while delivering comparable results on long sequence benchmarks.