Breaking the Stage Barrier: A Novel Single-Stage Approach to Long Context Extension for Large Language Models

Recently, Large language models (LLMs) have revolutionized Natural Language Processing (NLP). Pretrained LLMs, due to limited training context size, struggle with handling long token sequences, limiting their performance on various downstream tasks. Current solutions toward long context modeling often employ multi-stage continual pertaining, which progressively increases the effective context length through several continual pretraining stages. However, those approaches require extensive manual tuning and human expertise. In this paper, we introduce a novel single-stage continual pretraining method, Head-Adaptive Rotary Position Encoding (HARPE), to equip LLMs with long context modeling capabilities while simplifying the training process. Our HARPE leverages different Rotary Position Encoding (RoPE) base frequency values across different attention heads and directly trains LLMs on the target context length. Extensive experiments on 4 language modeling benchmarks, including the latest RULER benchmark, demonstrate that HARPE excels in understanding and integrating long-context tasks with single-stage training, matching and even outperforming existing multi-stage methods. Our results highlight that HARPE successfully breaks the stage barrier for training LLMs with long context modeling capabilities.

LBPE: Long-token-first Tokenization to Improve Large Language Models

The prevalent use of Byte Pair Encoding (BPE) in Large Language Models (LLMs) facilitates robust handling of subword units and avoids issues of out-of-vocabulary words. Despite its success, a critical challenge persists: long tokens, rich in semantic information, have fewer occurrences in tokenized datasets compared to short tokens, which can result in imbalanced learning issue across different tokens. To address that, we propose LBPE, which prioritizes long tokens during the encoding process. LBPE generates tokens according to their reverse ranks of token length rather than their ranks in the vocabulary, granting longer tokens higher priority during the encoding process. Consequently, LBPE smooths the frequency differences between short and long tokens, and thus mitigates the learning imbalance. Extensive experiments across diverse language modeling tasks demonstrate that LBPE consistently outperforms the original BPE, well demonstrating its effectiveness.

CartesianMoE: Boosting Knowledge Sharing among Experts via Cartesian Product Routing in Mixture-of-Experts

Large language models (LLM) have been attracting much attention from the community recently, due to their remarkable performance in all kinds of downstream tasks. According to the well-known scaling law, scaling up a dense LLM enhances its capabilities, but also significantly increases the computational complexity. Mixture-of-Experts (MoE) models address that by allowing the model size to grow without substantially raising training or inference costs. Yet MoE models face challenges regarding knowledge sharing among experts, making their performance somehow sensitive to routing accuracy. To tackle that, previous works introduced shared experts and combined their outputs with those of the top $K$ routed experts in an ``addition'' manner. In this paper, inspired by collective matrix factorization to learn shared knowledge among data, we propose CartesianMoE, which implements more effective knowledge sharing among experts in more like a ``multiplication'' manner. Extensive experimental results indicate that CartesianMoE outperforms previous MoE models for building LLMs, in terms of both perplexity and downstream task performance. And we also find that CartesianMoE achieves better expert routing robustness.

MaskMoE: Boosting Token-Level Learning via Routing Mask in Mixture-of-Experts

Scaling model capacity enhances its capabilities but significantly increases computation. Mixture-of-Experts models (MoEs) address this by allowing model capacity to scale without substantially increasing training or inference costs. Despite their promising results, MoE models encounter several challenges. Primarily, the dispersion of training tokens across multiple experts can lead to underfitting, particularly for infrequent tokens. Additionally, while fixed routing mechanisms can mitigate this issue, they compromise on the diversity of representations. In this paper, we propose MaskMoE, a method designed to enhance token-level learning by employing a routing masking technique within the Mixture-of-Experts model. MaskMoE is capable of maintaining representation diversity while achieving more comprehensive training. Experimental results demonstrate that our method outperforms previous dominant Mixture-of-Experts models in both perplexity (PPL) and downstream tasks.

Scaffold-BPE: Enhancing Byte Pair Encoding with Simple and Effective Scaffold Token Removal

Byte Pair Encoding (BPE) serves as a foundation method for text tokenization in the Natural Language Processing (NLP) field. Despite its wide adoption, the original BPE algorithm harbors an inherent flaw: it inadvertently introduces a frequency imbalance for tokens in the text corpus. Since BPE iteratively merges the most frequent token pair in the text corpus while keeping all tokens that have been merged in the vocabulary, it unavoidably holds tokens that primarily represent subwords of complete words and appear infrequently on their own in the text corpus. We term such tokens as Scaffold Tokens. Due to their infrequent appearance in the text corpus, Scaffold Tokens pose a learning imbalance issue for language models. To address that issue, we propose Scaffold-BPE, which incorporates a dynamic scaffold token removal mechanism by parameter-free, computation-light, and easy-to-implement modifications to the original BPE. This novel approach ensures the exclusion of low-frequency Scaffold Tokens from the token representations for the given texts, thereby mitigating the issue of frequency imbalance and facilitating model training. On extensive experiments across language modeling tasks and machine translation tasks, Scaffold-BPE consistently outperforms the original BPE, well demonstrating its effectiveness and superiority.

Temporal Scaling Law for Large Language Models

Recently, Large Language Models (LLMs) are widely adopted in a wide range of tasks, leading to increasing attention towards the research on how scaling LLMs affects their performance. Existing works, termed as Scaling Laws, have discovered that the loss of LLMs scales as power laws with model size, computational budget, and dataset size. However, the performance of LLMs throughout the training process remains untouched. In this paper, we propose the novel concept of Temporal Scaling Law and study the loss of LLMs from the temporal dimension. We first investigate the imbalance of loss on each token positions and develop a reciprocal-law across model scales and training stages. We then derive the temporal scaling law by studying the temporal patterns of the reciprocal-law parameters. Results on both in-distribution (IID) data and out-of-distribution (OOD) data demonstrate that our temporal scaling law accurately predicts the performance of LLMs in future training stages. Moreover, the temporal scaling law reveals that LLMs learn uniformly on different token positions, despite the loss imbalance. Experiments on pre-training LLMs in various scales show that this phenomenon verifies the default training paradigm for generative language models, in which no re-weighting strategies are attached during training. Overall, the temporal scaling law provides deeper insight into LLM pre-training.

PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation

Recently, the scale of transformers has grown rapidly, which introduces considerable challenges in terms of training overhead and inference efficiency in the scope of task adaptation. Existing works, namely Parameter-Efficient Fine-Tuning (PEFT) and model compression, have separately investigated the challenges. However, PEFT cannot guarantee the inference efficiency of the original backbone, especially for large-scale models. Model compression requires significant training costs for structure searching and re-training. Consequently, a simple combination of them cannot guarantee accomplishing both training efficiency and inference efficiency with minimal costs. In this paper, we propose a novel Parallel Yielding Re-Activation (PYRA) method for such a challenge of training-inference efficient task adaptation. PYRA first utilizes parallel yielding adaptive weights to comprehensively perceive the data distribution in downstream tasks. A re-activation strategy for token modulation is then applied for tokens to be merged, leading to calibrated token features. Extensive experiments demonstrate that PYRA outperforms all competing methods under both low compression rate and high compression rate, demonstrating its effectiveness and superiority in maintaining both training efficiency and inference efficiency for large-scale foundation models. Our code will be released to the public.

Confidence-based Visual Dispersal for Few-shot Unsupervised Domain Adaptation

Unsupervised domain adaptation aims to transfer knowledge from a fully-labeled source domain to an unlabeled target domain. However, in real-world scenarios, providing abundant labeled data even in the source domain can be infeasible due to the difficulty and high expense of annotation. To address this issue, recent works consider the Few-shot Unsupervised Domain Adaptation (FUDA) where only a few source samples are labeled, and conduct knowledge transfer via self-supervised learning methods. Yet existing methods generally overlook that the sparse label setting hinders learning reliable source knowledge for transfer. Additionally, the learning difficulty difference in target samples is different but ignored, leaving hard target samples poorly classified. To tackle both deficiencies, in this paper, we propose a novel Confidence-based Visual Dispersal Transfer learning method (C-VisDiT) for FUDA. Specifically, C-VisDiT consists of a cross-domain visual dispersal strategy that transfers only high-confidence source knowledge for model adaptation and an intra-domain visual dispersal strategy that guides the learning of hard target samples with easy ones. We conduct extensive experiments on Office-31, Office-Home, VisDA-C, and DomainNet benchmark datasets and the results demonstrate that the proposed C-VisDiT significantly outperforms state-of-the-art FUDA methods. Our code is available at https://github.com/Bostoncake/C-VisDiT.