When LLMs Know but Fail to Reason: Injecting Memory for Reasoning Enhancement
Published in NeurIPS, 2026
Recommended citation: Wang, T., Hu, S., Qin, R., Zhang, H., Wang, Q., Yu, S., Liu, D., Wei, Z., & Shen, W. When LLMs Know but Fail to Reason: Injecting Memory for Reasoning Enhancement. In NeurIPS 2026. https://openreview.net/pdf?id=4I2mGv36HD
Abstract. This paper introduces Memory-to-Reasoning Alignment (MRA) to measure memory engagement during reasoning and proposes training-free Memory-Guided Activation Injection (MGAI) to improve reasoning by injecting internal memory information into reasoning representations.
Authors: Tian Wang*, Shiyu Hu*, Ruiyang Qin, Hanyue Zhang, Qingzhuo Wang, Shuhan Yu, Dongrui Liu, Zhihua Wei, Wen Shen†.
* Equal contribution; † Corresponding author.