Meta’s DreamGym framework trains AI agents in a simulated world to cut reinforcement learning costs

Yann LeCun, a prominent figure in artificial intelligence and head of Meta's AI research lab, has confirmed his departure from Meta to establish a new startup focused on advanced machine intelligence. Meta will continue to partner with LeCun's new venture, indicating ongoing collaboration in AI research.

The paper introduces Group Turn Policy Optimization (GTPO), a novel reinforcement learning algorithm aimed at enhancing the training of Large Language Models (LLMs) for multi-turn Tool-Integrated Reasoning (TIR). GTPO addresses limitations of existing methods like Group Relative Policy Optimization (GRPO) by implementing turn-level reward assignments, return-based advantage estimation, and self-supervised reward shaping, which collectively improve learning signals for complex interactions.

Recent advancements in large language models (LLMs) have enhanced their reasoning and decision-making capabilities as agents. However, these improvements often lead to longer chains of thought, reducing interaction efficiency. To address this, a new method called DEPO (Dual-Efficiency Preference Optimization) has been introduced, focusing on step-level and trajectory-level efficiencies. Experiments indicate DEPO can reduce token usage by up to 60.9% and steps by 26.9%, while improving performance by 29.3%.

The paper presents Group Relative Policy Optimization for Representation Model (GRPO-RM), a reinforcement learning method aimed at fine-tuning large language models (LLMs). It establishes a predefined output set to replace token sequence sampling, facilitating the generation of an output group essential for GRPO's optimization. A specialized reward function is also introduced to cater to representation models, with extensive experiments validating the method's effectiveness across various real-world datasets.

Dr. Yann LeCun, a leading figure in artificial intelligence, has announced his departure from Meta following a restructuring of the company's AI initiatives. This move comes as CEO Mark Zuckerberg aims to enhance Meta's position in the competitive tech landscape. LeCun has previously expressed skepticism about the potential of large language models achieving superintelligence.

Meta's Chief AI Scientist, Dr. Yann LeCun, is leaving the company after 12 years. His departure follows a restructuring of Meta's AI initiatives, which aims to enhance the company's competitive position in the technology sector. LeCun has been a pivotal figure in advancing artificial intelligence at Meta.

Yann LeCun, a prominent figure in artificial intelligence and head of Meta's AI research lab, has announced his departure from Meta at the end of 2025. He plans to establish a new startup focused on advanced machine intelligence research, with Meta remaining as a partner. This move marks a significant transition for both LeCun and the company as it navigates changes in its AI leadership.

Meta has secured a significant legal victory against the US Federal Trade Commission (FTC) in an antitrust case, which could allow the company to resume acquiring startups after a period of regulatory caution. This ruling may also impact other major tech firms like Google and Microsoft, who have been hesitant to pursue acquisitions due to increased scrutiny under FTC Chair Lina Khan.