The interest in linear complexity models for large language models is on the rise, although their scaling capacity remains uncertain. In this study, we present the scaling laws for linear complexity language models to establish a foundation for their scalability. Specifically, we examine the scaling behaviors of three efficient linear architectures. These include TNL, a linear attention model with data-independent decay; HGRN2, a linear RNN with data-dependent decay; and cosFormer2, a linear attention model without decay. We also include LLaMA as a baseline architecture for comparison with softmax attention. These models were trained with six variants, ranging from 70M to 7B parameters on a 300B-token corpus, and evaluated with a total of 1,376 intermediate checkpoints on various downstream tasks. These tasks include validation loss, commonsense reasoning, and information retrieval and generation. The study reveals that existing linear complexity language models exhibit similar scaling capabilities as conventional transformer-based models while also demonstrating superior linguistic proficiency and knowledge retention.

Improving the efficiency of inference in Large Language Models (LLMs) is a critical area of research. Post-training Quantization (PTQ) is a popular technique, but it often faces challenges at low-bit levels, particularly in downstream tasks. Quantization-aware Training (QAT) can alleviate this problem, but it requires significantly more computational resources. To tackle this, we introduced Weight-Decomposed Low-Rank Quantization-Aware Training (DL-QAT), which merges the advantages of QAT while training only less than 1% of the total parameters. Specifically, we introduce a group-specific quantization magnitude to adjust the overall scale of each quantization group. Within each quantization group, we use LoRA matrices to update the weight size and direction in the quantization space. We validated the effectiveness of our method on the LLaMA and LLaMA2 model families. The results show significant improvements over our baseline method across different quantization granularities. For instance, for LLaMA-7B, our approach outperforms the previous state-of-the-art method by 4.2% in MMLU on 3-bit LLaMA-7B. Additionally, our quantization results on pre-trained models also surpass previous QAT methods, demonstrating the superior performance and efficiency of our approach.

Despite the promising performance of state space models (SSMs) in long sequence modeling, limitations still exist. Advanced SSMs like S5 and S6 (Mamba) in addressing non-uniform sampling, their recursive structures impede efficient SSM computation via convolution. To overcome compatibility limitations in parallel convolutional computation, this paper proposes a novel non-recursive non-uniform sample processing strategy. Theoretical analysis of SSMs through the lens of Event-Triggered Control (ETC) theory reveals the Non-Stable State (NSS) problem, where deviations from sampling point requirements lead to error transmission and accumulation, causing the divergence of the SSM’s hidden state. Our analysis further reveals that adjustments of input sequences with early memories can mitigate the NSS problem, achieving Sampling Step Adaptation (SSA).Building on this insight, we introduce a simple yet effective plug-and-play mechanism, State Memory Replay (SMR), which utilizes learnable memories to adjust the current state with multi-step information for generalization at sampling points different from those in the training data. This enables SSMs to stably model varying sampling points. Experiments on long-range modeling tasks in autoregressive language modeling and Long Range Arena demonstrate the general effectiveness of the SMR mechanism for a series of SSM models.

Pretrained vision-language (VL) models have shown impressive results on various multi-modal downstream tasks recently. Many of the benchmark models build on pretrained causal language models (LMs), leveraging the original few-shot learning and generalization capability of the LMs trained with large text corpora. However, these models are often gigantic and require large-scale image and text data with high computational cost to train. This paper introduces a moderate-size model called MAP for efficient VL transfer learning through adapter-based pretraining and prompting. We aim to answer the question of how much we can complete through VL pretraining within the low-data regime while maximizing efficiency in transferring knowledge of a moderate-size frozen LM. Our experiments demonstrate that MAP achieves substantially better zero-shot and few-shot performance on downstream VL tasks with only 10% the size of pretraining data and a 30x lighter pretrained LM backbone compared to Frozen. MAP also outperforms fully trained models of comparable size at retaining its transfer learning ability when the amount of training data reduces.

Multi-Document Scientific Summarization (MDSS) aims to produce coherent and concise summaries for clusters of topic-relevant scientific papers. This task requires precise understanding of paper content and accurate modeling of cross-paper relationships. Knowledge graphs convey compact and interpretable structured information for documents, which makes them ideal for content modeling and relationship modeling. In this paper, we present KGSum, an MDSS model centred on knowledge graphs during both the encoding and decoding process. Specifically, in the encoding process, two graph-based modules are proposed to incorporate knowledge graph information into paper encoding, while in the decoding process, we propose a two-stage decoder by first generating knowledge graph information of summary in the form of descriptive sentences, followed by generating the final summary. Empirical results show that the proposed architecture brings substantial improvements over baselines on the Multi-Xscience dataset.