DAST: Context-Aware Compression in LLMs via Dynamic Allocation of Soft Tokens

Large Language Models (LLMs) face computational inefficiencies and redundant processing when handling long context inputs, prompting a focus on compression techniques. While existing semantic vector-based compression methods achieve promising performance, these methods fail to account for the intrinsic information density variations between context chunks, instead allocating soft tokens uniformly across context chunks. This uniform distribution inevitably diminishes allocation to information-critical regions. To address this, we propose Dynamic Allocation of Soft Tokens (DAST), a simple yet effective method that leverages the LLM's intrinsic understanding of contextual relevance to guide compression. DAST combines perplexity-based local information with attention-driven global information to dynamically allocate soft tokens to the informative-rich chunks, enabling effective, context-aware compression. Experimental results across multiple benchmarks demonstrate that DAST surpasses state-of-the-art methods.

Establishing a stronger baseline for lightweight contrastive models

Recent research has reported a performance degradation in self-supervised contrastive learning for specially designed efficient networks, such as MobileNet and EfficientNet. A common practice to address this problem is to introduce a pretrained contrastive teacher model and train the lightweight networks with distillation signals generated by the teacher. However, it is time and resource consuming to pretrain a teacher model when it is not available. In this work, we aim to establish a stronger baseline for lightweight contrastive models without using a pretrained teacher model. Specifically, we show that the optimal recipe for efficient models is different from that of larger models, and using the same training settings as ResNet50, as previous research does, is inappropriate. Additionally, we observe a common issu e in contrastive learning where either the positive or negative views can be noisy, and propose a smoothed version of InfoNCE loss to alleviate this problem. As a result, we successfully improve the linear evaluation results from 36.3\% to 62.3\% for MobileNet-V3-Large and from 42.2\% to 65.8\% for EfficientNet-B0 on ImageNet, closing the accuracy gap to ResNet50 with $5\times$ fewer parameters. We hope our research will facilitate the usage of lightweight contrastive models.

Efficient Sub-structured Knowledge Distillation

Structured prediction models aim at solving a type of problem where the output is a complex structure, rather than a single variable. Performing knowledge distillation for such models is not trivial due to their exponentially large output space. In this work, we propose an approach that is much simpler in its formulation and far more efficient for training than existing approaches. Specifically, we transfer the knowledge from a teacher model to its student model by locally matching their predictions on all sub-structures, instead of the whole output space. In this manner, we avoid adopting some time-consuming techniques like dynamic programming (DP) for decoding output structures, which permits parallel computation and makes the training process even faster in practice. Besides, it encourages the student model to better mimic the internal behavior of the teacher model. Experiments on two structured prediction tasks demonstrate that our approach outperforms previous methods and halves the time cost for one training epoch.