Continual pretraining enables Large Language Models (LLMs) to adapt to specialized domains like medicine and law. However, we observe a consistent phenomenon across different model sizes and domains: a temporary performance drop at the start of the continual pretraining process, followed by a performance recovery phase. To gain a deeper understanding of this issue, we use the stability gap— a concept adapted from the visual domain—which explains this initial drop arises from instability in the model’s general abilities. We validate this hypothesis through a series of experiments. To address this initial instability and enhance LLM performance within a fixed compute budget, we propose a training strategy that mitigates instability by increasing the number of epochs, alongside two data sampling strategies targeting data domain relevance and corpus distribution. We conduct experiments on Llama-family models to validate the effectiveness of our strategies for continual pretraining and instruction tuning in medical and legal domains. Our strategies improve the average medical task performance of the OpenLlama-3B model from 36.2% to 40.7% using only 40% of the original training budget, while also enhancing general task performance without causing forgetting. Furthermore, we aPPLy our strategies to continually pre-train and instruction-tune the Llama-3-8B model. The resulting model, Llama-3-Physician, achieves the best medical performance among open-source models on several benchmarks and rivals GPT-4 on specific tasks. We release our models at https://huggingface.co/YiDuo1999/Llama-3-Physician-8B-Instruct.

We introduce AdamS, a simple yet effective alternative to Adam for large language model (LLM) pretraining and post-training. By leveraging a novel denominator, i.e., the root of weighted sum of squares of the momentum and the current gradient, AdamS eliminates the need for second-moment estimates. Hence, AdamS is efficient, matching the memory and compute footprint of SGD with momentum while delivering superior optimization performance. Moreover, AdamS is easy to adopt: it can directly inherit hyperparameters of AdamW, and is entirely model-agnostic, integrating seamlessly into existing pipelines without modifications to optimizer APIs or architectures. The motivation behind AdamS stems from the observed smoothness properties in transformer objectives, where local smoothness is governed by gradient magnitudes that can be further approximated by momentum magnitudes. We establish rigorous theoretical convergence guarantees and provide practical guidelines for hyperparameter selection. Empirically, AdamS demonstrates strong performance in various tasks, including pre-training runs on GPT-2 and Llama2 (up to 13B parameters) and reinforcement learning in post-training regimes. With its efficiency, simplicity, and theoretical grounding, AdamS stands as a compelling alternative to existing optimizers.

Large language models (LLMs) excel in many tasks, but their safety guarantees vary by language, e.g., responses in English tend to be safer than those in low-resource languages. This inconsistency creates a vulnerability, since an attacker can circumvent safety measures by using a less-supported language as an intermediary, even without fluency in that language. Traditional solutions rely on multilingual safety alignment, which demands vast, per-language datasets and introduces significant trade-offs between usefulness and safety (the so-called “alignment tax”). To overcome these limitations, we introduce English as Defense Proxy (E-Proxy), a unified approach that leverages English, usually the advantage language of LLMs, as a universal safety anchor. During multilingual training, E-Proxy uses English jailbreak prompts to extract the model’s existing safety knowledge, then applies simple language-mapping prompts (e.g., “Please answer in target language”) to transfer that knowledge across languages. Our analysis shows that formulating prompts in a high-resource language preserves the model’s utility, while enforcing responses in the target language significantly enhances safety. We evaluate E-Proxy on extensive benchmarks of both attack resistance and task performance. On the MultiJail benchmark, E-Proxy blocks over 99 % of jailbreak attempts while retaining 95 % of average task performance, all with a simply constructed multilingual alignment data.

Recent researches on video large language models (VideoLLM) predominantly focus on model architectures and training datasets, leaving the interaction format between the user and the model under-explored. In existing works, users often interact with VideoLLMs by using the entire video and a query as input, after which the model generates a response. This interaction format constrains the application of VideoLLMs in scenarios such as live-streaming comprehension where videos do not end and responses are required in a real-time manner, and also results in unsatisfactory performance on time-sensitive tasks that requires localizing video segments. In this paper, we focus on a video-text duet interaction format. This interaction format is characterized by the continuous playback of the video, and both the user and the model can insert their text messages at any position during the video playback. When a text message ends, the video continues to play, akin to the alternative of two performers in a duet. We construct MMDuetIT, a video-text training dataset designed to adapt VideoLLMs to video-text duet interaction format. We also introduce the Multi-Answer Grounded Video Question Answering (MAGQA) task to benchmark the real-time response ability of VideoLLMs. Trained on MMDuetIT, MMDuet demonstrates that adopting the video-text duet interaction format enables the model to achieve significant improvements in various time-sensitive tasks (76% CIDEr on YouCook2 dense video captioning, 90% mAP on QVHighlights highlight detection and 25% R@0.5 on Charades-STA temporal video grounding) with minimal training efforts, and also enable VideoLLMs to reply in a real-time manner as the video plays.

While the Mamba architecture demonstrates superior inference efficiency and competitive performance on short-context natural language processing (NLP) tasks, empirical evidence suggests its capacity to comprehend long contexts is limited compared to transformer-based models. In this study, we investigate the long-context efficiency issues of the Mamba models and propose ReMamba, which enhances Mamba’s ability to comprehend long contexts. ReMamba incorporates selective compression and adaptation techniques within a two-stage re-forward process, incurring minimal additional inference costs overhead. Experimental results on the LongBench and L-Eval benchmarks demonstrate ReMamba’s efficacy, improving over the baselines by 3.2 and 1.6 points, respectively, and attaining performance almost on par with same-size transformer models.

Is it always necessary to compute tokens from shallow to deep layers in Transformers? The continued success of vanilla Transformers and their variants suggests an undoubted “yes”. In this work, however, we attempt to break the depth-ordered convention by proposing a novel architecture dubbed mixture-of-modules (MoM), which is motivated by an intuition that any layer, regardless of its position, can be used to compute a token as long as it possesses the needed processing capabilities. The construction of MoM starts from a finite set of modules defined by multi-head attention and feed-forward networks, each distinguished by its unique parameterization. Two routers then iteratively select attention modules and feed-forward modules from the set to process a token. The selection dynamically expands the computation graph in the forward pass of the token, culminating in an assembly of modules. We show that MoM provides not only a unified framework for Transformers and their numerous variants but also a flexible and learnable approach for reducing redundancy in Transformer parameterization. We pre-train various MoMs using OpenWebText. Empirical results demonstrate that MoMs, of different sizes, consistently outperform vanilla transformers. More interestingly, after removing 50% of the multi-head attention modules and 25% of the feed-forward modules, an MoM model still holds comparable performance. Additionally, by properly adjusting the number of modules and compressing the model depth, one can have an MoM that achieves comparable performance to GPT-2 (774M) while saving 16% TFLOPs and 42% memory usage during forward computation.