The generative AI future beckons bright for financial institutions

In the rapidly evolving landscape of artificial intelligence, we are witnessing a transformational shift from Large Language Models (LLMs) to Large Reasoning Models (LRMs) to agentic technology leveraging compositionality. This transition signifies a revolutionary change in how businesses, especially in sectors like banking, can achieve enhanced decision-making capabilities through advanced analytical processes.

The evolution of AI reasoning

The journey from LLMs to LRMs is marked by a significant advancement in reasoning. While traditional LLMs excel in language understanding, LRMs leverage Type 2 reasoning during inference to heighten analytical depth without necessitating retraining of existing models. Pioneering systems like GPT-4-03 and Alibaba’s Marco-01 are steppingstones in this evolution, utilising advanced analytical techniques like chain-of-thought reasoning and Monte Carlo tree search to outperform earlier models.

The quadrumvirate: Enhancing AI performance

At the heart of this shift is a redefined framework comprising four elements: compute power, data, algorithms, and crucially, reasoning methods at inference time. This “quadrumvirate” moves beyond LLMs to LRMs and agentic technology to engage in context-driven reasoning, significantly enhancing their performance. By optimising the existing capabilities of trained LLMs through advanced reasoning techniques that include compositionality at inference time using models that have specialised capabilities that are brought together through agentic technology, businesses can harness deeper analytical insights for improved decision-making.

Balancing fast and slow thinking

Cognitive psychology describes two modes of reasoning: Type 1 (fast, intuitive) and Type 2 (slow, analytical). LRMs bridge these modes by dynamically adjusting their approach, allowing for swift actions when necessary while applying thorough analysis in more complex situations. This balance enhances decision-making processes, particularly in sectors like banking, which require agility and depth.

Compositionality: The foundation of agentic technology

The evolution from LLMs to LRMs is further accelerated by the concept of compositionality—enabling systems to deconstruct tasks into manageable components for systematic resolution. Advanced LRM systems utilise techniques like chain-of-thought reasoning and search algorithms, integrating specialised models trained in distinct domains. Unlike classic LLMs, which struggle with compositionality and adaptability, LRMs can effectively navigate complex challenges that extend beyond their initial training data.

Note that this is exactly how IBM’s approach to optimising banking through its proprietary Component Business Modeling (CBM) and industry standards like BIAN operate. These standard industry models break down a bank into its constituent parts thereby enabling specialisation.