Yuzhe Gu

Large language models (LLMs) with extended context windows enable powerful downstream applications but impose significant memory overhead, as caching all key-value (KV) states scales linearly with sequence length and batch size. Existing cache eviction methods address this by exploiting attention sparsity, yet they typically rank tokens heuristically using accumulated attention weights without considering their true impact on attention outputs. We propose Optimal Brain Cache (OBCache), a principled framework that formulates cache eviction as a layer-wise structured pruning problem. Building upon the Optimal Brain Damage (OBD) theory, OBCache quantifies token saliency by measuring the perturbation in attention outputs induced by pruning tokens, with closed-form scores derived for isolated keys, isolated values, and joint key-value pairs. Our scores account not only for attention weights but also for information from value states and attention outputs, thereby enhancing existing eviction strategies with output-aware signals. Experiments on LLaMA and Qwen models demonstrate that replacing the heuristic scores in existing works, which estimate token saliency across different query positions, with OBCache's output-aware scores consistently improves long-context accuracy.

Throughout a conversation, the way participants interact with each other is in constant flux: their tones may change, they may resort to different strategies to convey their points, or they might alter their interaction patterns. An understanding of these dynamics can complement that of the actual facts and opinions discussed, offering a more holistic view of the trajectory of the conversation: how it arrived at its current state and where it is likely heading. In this work, we introduce the task of summarizing the dynamics of conversations, by constructing a dataset of human-written summaries, and exploring several automated baselines. We evaluate whether such summaries can capture the trajectory of conversations via an established downstream task: forecasting whether an ongoing conversation will eventually derail into toxic behavior. We show that they help both humans and automated systems with this forecasting task. Humans make predictions three times faster, and with greater confidence, when reading the summaries than when reading the transcripts. Furthermore, automated forecasting systems are more accurate when constructing, and then predicting based on, summaries of conversation dynamics, compared to directly predicting on the transcripts.

Yuzhe Gu, Enmao Diao. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Yilun Hua, Nicholas Chernogor, Yuzhe Gu, Seoyeon Jeong, Miranda Luo, Cristian Danescu-Niculescu-Mizil. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.