Meta's research team, Meta FAIR, has introduced advancements in memory layer technology aimed at enhancing the capabilities of language models. The new scalable memory layer can increase a language model's capacity by up to 128 billion parameters, equivalent to one trillion tokens, without significantly increasing computational demands. This development is seen as a solution to the limitations of context length in existing models. Additionally, Meta AI has proposed a novel approach called Explicit Working Memory (EWE), which aims to improve factuality in long-form text generation by integrating a working memory system. These innovations are part of ongoing efforts to enhance the reasoning abilities of language models and address challenges in training and inference.
Meta AI Introduces EWE (Explicit Working Memory): A Novel Approach that Enhances Factuality in Long-Form Text Generation by Integrating a Working Memory A team of researchers from Meta FAIR has proposed EWE (Explicit Working Memory), an innovative AI approach that enhances… https://t.co/BajzJBzfHI
来自 @OpenBMB 最新的「高效低成本训练大模型」的研究进展! 使用PRIME(结合过程奖励的强化学习)方法训练了一个7B模型,不依赖任何蒸馏和模仿学习,就高效训练出了一个数学能力超过 GPT-4o、Llama-3.1-70B的 7B 模型 Eurus-2-7B-PRIME。 GitHub:https://t.co/aNRDu3vPex… https://t.co/j9QJ4hvXxA
Me likey extra memory for LLMs But I would have liked some studies on the generalization ability of those models https://t.co/tfedF4yFSl