Brief

Researchers have introduced GROW, a novel reinforcement learning framework designed to enhance the capabilities of vision-language model (VLM) agents in open-world tasks. Unlike previous methods that relied heavily on supervised fine-tuning, GROW adapts the Group Relative Policy Optimization (GRPO) algorithm by decomposing trajectories into state-action samples. This approach mitigates issues with long contexts and noise inherent in standard GRPO, enabling more effective multi-turn learning. Experiments on over 800 Minecraft tasks demonstrated that GROW achieves state-of-the-art performance, showcasing its potential for advancing VLM agents. AI

A dedicated group of Minecraft players has completed a massive digital archaeology project, archiving 24 terabytes of data from the game's notorious 2b2t anarchy server. This extensive archive preserves a million-square-block region of the server's world, which has been running for 16 years with minimal rules. The team, which included organizers like Crayne and contributors such as Fuch and Mahan, spent years and significant personal resources on this endeavor, carefully working to avoid detection by other players who might disrupt their efforts. The archived data, covering various dimensions of the server, will be shared via torrent in the coming weeks, with open-source tools also being made available. AI

Replit has introduced Replcraft, a new library that enables users to automate various aspects of the Minecraft game using code directly within the Replit environment. This tool allows for actions such as placing and retrieving blocks, managing items, and interacting with game entities. Developers can set up Replcraft by forking a template, joining a compatible Minecraft server, and configuring a special in-game structure with a unique token. AI