Processing long input remains a significant challenge for large language models (LLMs) due to the scarcity of large-scale long-context training data and the high computational cost of training models for extended context windows. In this paper, we propose **Ada**ptive **Gro**uped **P**ositional **E**ncoding (AdaGroPE), a training-free, plug-and-play method to enhance long-context understanding in existing LLMs. AdaGroPE progressively increases the reuse count of relative positions as the distance grows and dynamically adapts the positional encoding mapping to sequence length, thereby fully exploiting the range of pre-trained position embeddings. Its design is consistent with the principles of rotary position embedding (RoPE) and aligns with human perception of relative distance, enabling robust performance in real-world settings with variable-length inputs. Extensive experiments across various benchmarks demonstrate that our AdaGroPE consistently achieves state-of-the-art performance, surpassing baseline methods and even outperforming LLMs inherently designed for long-context processing on certain tasks.

As large language models continue to scale, computational costs and resource consumption have emerged as significant challenges. While existing sparsification methods like pruning reduce computational overhead, they risk losing model knowledge through parameter removal. This paper proposes DSMoE (Dynamic Sparse Mixture-of-Experts), a novel approach that achieves sparsification by partitioning pre-trained FFN layers into computational blocks. We implement adaptive expert routing using sigmoid activation and straight-through estimators, enabling tokens to flexibly access different aspects of model knowledge based on input complexity. Additionally, we introduce a sparsity loss term to balance performance and computational efficiency. Extensive experiments on LLaMA models demonstrate that under equivalent computational constraints, DSMoE achieves superior performance compared to existing pruning and MoE approaches across language modeling and downstream tasks, particularly excelling in generation tasks. Analysis reveals that DSMoE learns distinctive layerwise activation patterns, providing new insights for future MoE architecture design.

Recently, Large Language Models (LLMs) have been 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 Scaling Laws, have discovered that the final test loss of LLMs scales as power-laws with model size, computational budget, and dataset size. However, the temporal change of the test loss of an LLM throughout its pretraining process remains unexplored, though it is valuable in many aspects, such as selecting better hyperparameters *directly* on the target LLM. In this paper, we propose the novel concept of Temporal Scaling Law, studying how the test loss of an LLM evolves as the training steps scale up. In contrast to modeling the test loss as a whole in a coarse-grained manner, we break it down and dive into the fine-grained test loss of each token position, and further develop a dynamic hyperbolic-law. Afterwards, we derive the much more precise temporal scaling law by studying the temporal patterns of the parameters in the dynamic hyperbolic-law. Results on both in-distribution (ID) and out-of-distribution (OOD) validation datasets demonstrate that our temporal scaling law accurately predicts the test loss of LLMs across training steps. Our temporal scaling law has broad practical applications. First, it enables direct and efficient hyperparameter selection on the target LLM, such as data mixture proportions. Secondly, viewing the LLM pretraining dynamics from the token position granularity provides some insights to enhance the understanding of LLM pretraining.

Multi-modal large language models (MLLMs) integrate the inherent text generation capabilities of large language models with an understanding of other modalities, promising wide applications in open-ended tasks. Despite their success, they often generate plausible but incorrect content. This phenomenon, known as hallucination, significantly impacts their practical deployment. In this paper, we delve into the intrinsic characteristics of hallucination from the perspective of interaction between input and output tokens. We find that the hallucination typically occurs with attention reduction of output tokens to image tokens. Based on this observation, we introduce image Token attention-guided Decoding (iTaD), a plug-and-play method which leverages MLLMs’ internal representations to mitigate their hallucinations. We first define an image token attention vector to measure the inter-layer differences in attention of output tokens to image tokens across different layers. Based on the vector, we design a novel layer selection strategy and conduct inter-layer contrastive decoding to highlight the progression in image understanding, thereby exploiting attention to image tokens to mitigate hallucinations. Extensive experiments well demonstrate iTaD’s effectiveness across different MLLMs and benchmarks.