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.

This paper describes our TemporalTeller system for SemEval Task 1: Unsupervised Lexical Semantic Change Detection. We develop a unified framework for the common semantic change detection pipelines including preprocessing, learning word embeddings, calculating vector distances and determining threshold. We also propose Gamma Quantile Threshold to distinguish between changed and stable words. Based on our system, we conduct a comprehensive comparison among BERT, Skip-gram, Temporal Referencing and alignment-based methods. Evaluation results show that Skip-gram with Temporal Referencing achieves the best performance of 66.5% classification accuracy and 51.8% Spearman’s Ranking Correlation.