Hangyu Lin

Yong Lin, Hangyu Lin, Wei Xiong et al.

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax. To investigate alignment tax, we conducted experiments with existing RLHF algorithms using OpenLLaMA-3B, which revealed a pronounced alignment tax in NLP tasks. Whereas, despite various techniques to mitigate forgetting, they are often at odds with the RLHF performance, leading to a trade-off between alignment performance and forgetting mitigation, leading to an alignment-forgetting trade-off. In this paper we show that model averaging, which simply interpolates between pre and post RLHF model weights, surprisingly achieves the most strongest alignment-forgetting Pareto front among a wide range of competing methods. To understand its effectiveness, we offer theoretical insights into model averaging, revealing that it enhances performance Pareto front by increasing feature diversity on the layers where tasks share overlapped feature spaces. Empirical evidence corroborates our analysis by showing the benefits of averaging low-level transformer layers. Building on the analysis and the observation that averaging different layers of the transformer leads to significantly different alignment-forgetting trade-offs, we propose Heterogeneous Model Averaging (HMA) to Heterogeneously find various combination ratios of model layers. HMA seeks to maximize the alignment performance while incurring minimal alignment tax. Moreover, we validate HMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B which is evaluated by open-sourced preference model and GPT4. Code available here: https://github.com/avalonstrel/Mitigating-the-Alignment-Tax-of-RLHF.git.

Alejandro R. Salamanca, Diana Abagyan, Daniel D'souza et al. · microsoft-research

Tiny Aya redefines what a small multilingual language model can achieve. Trained on 70 languages and refined through region-aware posttraining, it delivers state-of-the-art in translation quality, strong multilingual understanding, and high-quality target-language generation, all with just 3.35B parameters. The release includes a pretrained foundation model, a globally balanced instruction-tuned variant, and three region-specialized models targeting languages from Africa, South Asia, Europe, Asia-Pacific, and West Asia. This report details the training strategy, data composition, and comprehensive evaluation framework behind Tiny Aya, and presents an alternative scaling path for multilingual AI: one centered on efficiency, balanced performance across languages, and practical deployment.

Hangyu Lin, Chao Wen, Chengming Xu et al.

Flow matching based video generative models have been increasingly relying on prepended Vision-Language Models (VLMs) to handle complex, instruction-based video editing. The prevailing assumption underlying this paradigm is that a connector module can seamlessly align the VLM's rich multi-modal reasoning with the original text embedding space of DiTs. However, we hypothesize that this alignment acts as a severe semantic bottleneck, degrading fine-grained structural variables. Verifying this is challenging, as end-to-end evaluations conflate alignment failures with generation errors, and natural datasets lack disentangled annotations. To rigorously investigate this, we propose a controlled data processing pipeline based on video composition that results in TRACE-Edit, a diagnostic dataset focusing on relation-based editing. Leveraging this dataset, we propose a comprehensive diagnostic protocol to analyze two important designs of meta-query and connector in the existing video editing models. Systematic evaluation of four representative model cases reveals that fine-grained structural semantics can be severely degraded during alignment. Our findings overturn the assumption of lossless semantic transfer, identifying the VLM-to-DiT alignment as a major bottleneck and providing a new diagnostic foundation for future multi-modal alignment architectures.

Viraat Aryabumi, John Dang, Dwarak Talupuru et al.

This technical report introduces Aya 23, a family of multilingual language models. Aya 23 builds on the recent release of the Aya model (Üstün et al., 2024), focusing on pairing a highly performant pre-trained model with the recently released Aya collection (Singh et al., 2024). The result is a powerful multilingual large language model serving 23 languages, expanding state-of-art language modeling capabilities to approximately half of the world's population. The Aya model covered 101 languages whereas Aya 23 is an experiment in depth vs breadth, exploring the impact of allocating more capacity to fewer languages that are included during pre-training. Aya 23 outperforms both previous massively multilingual models like Aya 101 for the languages it covers, as well as widely used models like Gemma, Mistral and Mixtral on an extensive range of discriminative and generative tasks. We release the open weights for both the 8B and 35B models as part of our continued commitment for expanding access to multilingual progress.

John Dang, Shivalika Singh, Daniel D'souza et al.

We introduce the Aya Expanse model family, a new generation of 8B and 32B parameter multilingual language models, aiming to address the critical challenge of developing highly performant multilingual models that match or surpass the capabilities of monolingual models. By leveraging several years of research at Cohere For AI and Cohere, including advancements in data arbitrage, multilingual preference training, and model merging, Aya Expanse sets a new state-of-the-art in multilingual performance. Our evaluations on the Arena-Hard-Auto dataset, translated into 23 languages, demonstrate that Aya Expanse 8B and 32B outperform leading open-weight models in their respective parameter classes, including Gemma 2, Qwen 2.5, and Llama 3.1, achieving up to a 76.6% win-rate. Notably, Aya Expanse 32B outperforms Llama 3.1 70B, a model with twice as many parameters, achieving a 54.0% win-rate. In this short technical report, we present extended evaluation results for the Aya Expanse model family and release their open-weights, together with a new multilingual evaluation dataset m-ArenaHard.

Team Cohere, Aakanksha, Arash Ahmadian et al. · mila

In this report we describe the development of Command A, a powerful large language model purpose-built to excel at real-world enterprise use cases. Command A is an agent-optimised and multilingual-capable model, with support for 23 languages of global business, and a novel hybrid architecture balancing efficiency with top of the range performance. It offers best-in-class Retrieval Augmented Generation (RAG) capabilities with grounding and tool use to automate sophisticated business processes. These abilities are achieved through a decentralised training approach, including self-refinement algorithms and model merging techniques. We also include results for Command R7B which shares capability and architectural similarities to Command A. Weights for both models have been released for research purposes. This technical report details our original training pipeline and presents an extensive evaluation of our models across a suite of enterprise-relevant tasks and public benchmarks, demonstrating excellent performance and efficiency.

Bowen Yang, Bharat Venkitesh, Dwarak Talupuru et al.

Long-context large language models (LLMs) have achieved remarkable advancements, driven by techniques like Rotary Position Embedding (RoPE) (Su et al., 2023) and its extensions (Chen et al., 2023; Liu et al., 2024c; Peng et al., 2023). By adjusting RoPE parameters and incorporating training data with extended contexts, we can train performant models with considerably longer input sequences. However, existing RoPE-based methods exhibit performance limitations when applied to extended context lengths. This paper presents a comprehensive analysis of various attention mechanisms, including RoPE, No Positional Embedding (NoPE), and Query-Key Normalization (QK-Norm), identifying their strengths and shortcomings in long-context modeling. Our investigation identifies distinctive attention patterns in these methods and highlights their impact on long-context performance, providing valuable insights for architectural design. Building on these findings, we propose a novel architecture featuring a hybrid attention mechanism that integrates global and local attention spans. This design not only surpasses conventional RoPE-based transformer models with full attention in both long and short context tasks but also delivers substantial efficiency gains during training and inference.

Diana Abagyan, Alejandro R. Salamanca, Andres Felipe Cruz-Salinas et al.

Pretraining massively multilingual Large Language Models (LLMs) for many languages at once is challenging due to limited model capacity, scarce high-quality data, and compute constraints. Moreover, the lack of language coverage of the tokenizer makes it harder to address the gap for new languages purely at the post-training stage. In this work, we study what relatively cheap interventions early on in training improve "language plasticity", or adaptation capabilities of the model post-training to new languages. We focus on tokenizer design and propose using a universal tokenizer that is trained for more languages than the primary pretraining languages to enable efficient adaptation in expanding language coverage after pretraining. Our systematic experiments across diverse groups of languages and different training strategies show that a universal tokenizer enables significantly higher language adaptation, with up to 20.2% increase in win rates compared to tokenizers specific to pretraining languages. Furthermore, a universal tokenizer also leads to better plasticity towards languages that are completely unseen in the tokenizer and pretraining, by up to 5% win rate gain. We achieve this adaptation to an expanded set of languages with minimal compromise in performance on the majority of languages included in pretraining.

Hangyu Lin, Chen Liu, Chengming Xu et al.

Cross-modality distillation arises as an important topic for data modalities containing limited knowledge such as depth maps and high-quality sketches. Such techniques are of great importance, especially for memory and privacy-restricted scenarios where labeled training data is generally unavailable. To solve the problem, existing label-free methods leverage a few pairwise unlabeled data to distill the knowledge by aligning features or statistics between the source and target modalities. For instance, one typically aims to minimize the L2 distance or contrastive loss between the learned features of pairs of samples in the source (e.g. image) and the target (e.g. sketch) modalities. However, most algorithms in this domain only focus on the experimental results but lack theoretical insight. To bridge the gap between the theory and practical method of cross-modality distillation, we first formulate a general framework of cross-modality contrastive distillation (CMCD), built upon contrastive learning that leverages both positive and negative correspondence, towards a better distillation of generalizable features. Furthermore, we establish a thorough convergence analysis that reveals that the distance between source and target modalities significantly impacts the test error on downstream tasks within the target modality which is also validated by the empirical results. Extensive experimental results show that our algorithm outperforms existing algorithms consistently by a margin of 2-3\% across diverse modalities and tasks, covering modalities of image, sketch, depth map, and audio and tasks of recognition and segmentation.

Hangyu Lin, Yanwei Fu, Yu-Gang Jiang et al.

Previous researches of sketches often considered sketches in pixel format and leveraged CNN based models in the sketch understanding. Fundamentally, a sketch is stored as a sequence of data points, a vector format representation, rather than the photo-realistic image of pixels. SketchRNN studied a generative neural representation for sketches of vector format by Long Short Term Memory networks (LSTM). Unfortunately, the representation learned by SketchRNN is primarily for the generation tasks, rather than the other tasks of recognition and retrieval of sketches. To this end and inspired by the recent BERT model, we present a model of learning Sketch Bidirectional Encoder Representation from Transformer (Sketch-BERT). We generalize BERT to sketch domain, with the novel proposed components and pre-training algorithms, including the newly designed sketch embedding networks, and the self-supervised learning of sketch gestalt. Particularly, towards the pre-training task, we present a novel Sketch Gestalt Model (SGM) to help train the Sketch-BERT. Experimentally, we show that the learned representation of Sketch-BERT can help and improve the performance of the downstream tasks of sketch recognition, sketch retrieval, and sketch gestalt.

Peng Lu, Gao Huang, Hangyu Lin et al.

This paper proposes a novel approach for Sketch-Based Image Retrieval (SBIR), for which the key is to bridge the gap between sketches and photos in terms of the data representation. Inspired by channel-wise attention explored in recent years, we present a Domain-Aware Squeeze-and-Excitation (DASE) network, which seamlessly incorporates the prior knowledge of sample sketch or photo into SE module and make the SE module capable of emphasizing appropriate channels according to domain signal. Accordingly, the proposed network can switch its mode to achieve a better domain feature with lower intra-class discrepancy. Moreover, while previous works simply focus on minimizing intra-class distance and maximizing inter-class distance, we introduce a loss function, named Multiplicative Euclidean Margin Softmax (MEMS), which introduces multiplicative Euclidean margin into feature space and ensure that the maximum intra-class distance is smaller than the minimum inter-class distance. This facilitates learning a highly discriminative feature space and ensures a more accurate image retrieval result. Extensive experiments are conducted on two widely used SBIR benchmark datasets. Our approach achieves better results on both datasets, surpassing the state-of-the-art methods by a large margin.

Peng Lu, Hangyu Lin, Yanwei Fu et al.

Sketch has been employed as an effective communicative tool to express the abstract and intuitive meanings of object. Recognizing the free-hand sketch drawing is extremely useful in many real-world applications. While content-based sketch recognition has been studied for several decades, the instance-level Sketch-Based Image Retrieval (SBIR) tasks have attracted significant research attention recently. The existing datasets such as QMUL-Chair and QMUL-Shoe, focus on the retrieval tasks of chairs and shoes. However, there are several key limitations in previous instance-level SBIR works. The state-of-the-art works have to heavily rely on the pre-training process, quality of edge maps, multi-cropping testing strategy, and augmenting sketch images. To efficiently solve the instance-level SBIR, we propose a new Deep Triplet Classification Siamese Network (DeepTCNet) which employs DenseNet-169 as the basic feature extractor and is optimized by the triplet loss and classification loss. Critically, our proposed DeepTCNet can break the limitations existed in previous works. The extensive experiments on five benchmark sketch datasets validate the effectiveness of the proposed model. Additionally, to study the tasks of sketch-based hairstyle retrieval, this paper contributes a new instance-level photo-sketch dataset - Hairstyle Photo-Sketch dataset, which is composed of 3600 sketches and photos, and 2400 sketch-photo pairs.

Wanyun Cui, Xiyou Zhou, Hangyu Lin et al.

Verbs are important in semantic understanding of natural language. Traditional verb representations, such as FrameNet, PropBank, VerbNet, focus on verbs' roles. These roles are too coarse to represent verbs' semantics. In this paper, we introduce verb patterns to represent verbs' semantics, such that each pattern corresponds to a single semantic of the verb. First we analyze the principles for verb patterns: generality and specificity. Then we propose a nonparametric model based on description length. Experimental results prove the high effectiveness of verb patterns. We further apply verb patterns to context-aware conceptualization, to show that verb patterns are helpful in semantic-related tasks.