The field of Large Language Model (LLM) agents is rapidly evolving, with a focus on improving their ability to learn from interactions, reason about complex procedures, and adapt to novel environments. Recent developments have centered around enhancing the sample efficiency of LLM agents, enabling them to learn from fewer interactions and generalize to new situations. This has been achieved through the introduction of new frameworks and techniques, such as hindsight trajectory rewriting and graph-based procedure representations. Additionally, researchers have been exploring the use of LLMs as world models, allowing agents to simulate future states and predict action outcomes. However, this capability is limited by LLMs' tendency toward hallucination and their reliance on static training knowledge. To address these limitations, researchers have proposed retrieval-augmented world models and context-folding frameworks, which have shown promising results in improving the performance of LLM agents. Noteworthy papers include: PADME, which introduces a graph-based representation of procedures to improve the execution of long-horizon procedures. R-WoM, which proposes a retrieval-augmented world model to ground LLM simulations in factual, up-to-date knowledge. OneLife, which presents a framework for inferring symbolic world models from unguided exploration in complex, stochastic environments.