Advancements in Large Language Model Memory and Reasoning

The field of large language models (LLMs) is witnessing significant developments in memory and reasoning capabilities. Recent research has focused on enhancing LLMs' ability to store, retrieve, and utilize knowledge effectively, leading to improved performance in various tasks. Notably, the integration of human-inspired cognitive architectures, such as dual-process theory, has shown promise in boosting LLMs' reasoning abilities. Additionally, innovative memory management systems and self-evolving frameworks have been proposed to enable LLMs to learn from experience and adapt to new tasks. These advancements have the potential to revolutionize the field of natural language processing and enable more sophisticated and human-like language understanding.

Noteworthy papers include: PRIME, which introduces a multi-agent reasoning framework that dynamically integrates fast and slow thinking, enabling open-source LLMs to perform competitively with state-of-the-art closed-source models. MemGen proposes a dynamic generative memory framework that equips agents with a human-esque cognitive faculty, enabling them to recall and augment latent memory throughout reasoning. LatentEvolve presents a self-evolving latent test-time scaling framework that enables LLMs to learn how to scale test-time computation effectively, leading to improved performance on various benchmarks.

Sources

PRIME: Planning and Retrieval-Integrated Memory for Enhanced Reasoning

Distributed Associative Memory via Online Convex Optimization

MemGen: Weaving Generative Latent Memory for Self-Evolving Agents

LatentEvolve: Self-Evolving Test-Time Scaling in Latent Space

ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory

Memory Management and Contextual Consistency for Long-Running Low-Code Agents

ID-RAG: Identity Retrieval-Augmented Generation for Long-Horizon Persona Coherence in Generative Agents

Mem-{\alpha}: Learning Memory Construction via Reinforcement Learning

Memory-Driven Self-Improvement for Decision Making with Large Language Models

TokMem: Tokenized Procedural Memory for Large Language Models

Exploring System 1 and 2 communication for latent reasoning in LLMs

Redundancy-as-Masking: Formalizing the Artificial Age Score (AAS) to Model Memory Aging in Generative AI

Built with on top of