The Age of Reasoning [pdf] (metisreasoning.ai)

A new paper titled "The Age of Reasoning," authored by Thore Graepel and colleagues, critiques the reasoning limitations of large language models (LLMs) and proposes a novel architecture inspired by AlphaGo's strategic reasoning. This architecture emphasizes learned reasoning strategies, where claims are only accepted if supported by evidence, aiming to enhance both trustworthiness and efficiency of AI systems. By employing Hintikka’s conception of reasoning, the paper advocates for a system that can learn from experience while maintaining an independent knowledge base that is auditable and verifiable. The significance of this work lies in addressing fundamental issues faced by current LLMs, such as hallucinations, opaque reasoning, and lack of manageable knowledge representation. These shortcomings hinder LLMs from being reliably employed in high-stakes environments like healthcare or finance. The proposed architecture's ability to reason from first principles and progressively improve as knowledge accrues could bridge the "trust gap" hindering AI deployment in critical domains. This work is a step toward building AI systems that can reason more transparently and effectively, ultimately unlocking their full potential across various applications.