Tag
This paper investigates whether long-context language models exhibit episodic-like order memory similar to humans, using a temporal order memory task based on a novel. The authors find that models show the same distance effect and uncover a single attention head reinstating temporal context, supporting temporal context reinstatement as a mechanism for episodic-like memory in LLMs.
This paper proposes a training-inference consistent segmented execution framework for long-context LLMs to address the mismatch between full-context training and restricted inference regimes, achieving comparable performance with significantly reduced memory usage.
OjaKV introduces a context-aware online low-rank KV cache compression framework that uses hybrid storage and Oja's algorithm for incremental subspace adaptation to reduce GPU memory bottlenecks in long-context LLM inference without model fine-tuning.