Researchers have developed MEMAUDIT, a novel evaluation protocol for assessing the long-term memory writing capabilities of LLM agents. This protocol separates memory writing performance from retrieval and reasoning, allowing for a more precise analysis of how agents compress past interactions into persistent memory under budget constraints. Concurrently, a new hybrid processing-using-memory architecture called DARTH-PUM has been proposed, which integrates analog and digital processing-using-memory techniques to enable general-purpose computation within memory arrays. This architecture shows significant speedups for applications including large language models, AES encryption, and convolutional neural networks. AI
Summary written by gemini-2.5-flash-lite from 4 sources. How we write summaries →
IMPACT New evaluation protocols and hardware architectures could accelerate the development of more capable and efficient LLM agents.
RANK_REASON This cluster contains multiple research papers detailing new evaluation protocols and hardware architectures for LLMs.