With each upgrade, AI models become more advanced. Take our latest Command R models, for example; they offer significant improvements in coding, math, and latency. With every new update, it makes you wonder: how good are large language models (LLMs) at reasoning?

For businesses, LLM reasoning is essential because it allows AI to tackle more than just basic tasks. With advanced reasoning, models can manage complex processes that involve decision-making like automating workflows, analyzing data trends, and supporting strategic decisions. This makes them especially useful tools in fields like finance, healthcare, and legal, where getting things right really matters.

What is LLM reasoning?

LLM reasoning is a skill that models are increasingly being trained to perform. It refers to the ability of AI models to logically process information, break down complex problems, and make decisions based on context and patterns. This is a step-change from traditional LLM retrieval and response.

Cohere Co-Founder and CEO Aidan Gomez describes reasoning as, “crucial to intelligence … It's not a discrete 'does it have this capability or not'; it's a continuum of how robust the reasoning engine inside these models is.”

Modern LLMs are becoming more adept and can now perform mathematical reasoning, predictive logic, and autonomous multi-agent reasoning. One of the most exciting advancements is the ability to simulate a kind of inner monologue, where models can walk through their decision-making processes. This is a key step toward robust reasoning, although training data for this is rare and can be difficult to create. Overall, different types of reasoning are essential to LLM intelligence, and ongoing research is continually expanding our understanding of their potential.

Why is LLM reasoning important?

LLMs with strong reasoning skills are incredibly valuable for businesses. Right now, everyone’s talking about how to make them work for their specific needs. We’re stepping into an era where AI can take on specialized, complex tasks in more personalized and domain-specific ways.

Reasoning also helps build trust with users. Just like how citations add transparency to retrieval-augmented generation (RAG) systems, as models get better at explaining their thought process through complex problems, people can rely on them for more accurate and context-aware results.

What is Cohere’s approach?

Cohere's approach to LLM development is deeply rooted in real-world practicality, with a strong focus on creating models that solve immediate business challenges. Recent improvements in model design and training methods are changing the way LLMs tackle reasoning tasks.

One breakthrough is to train the model to break down complex problems and elicit reasoning, for example through chain-of-thought (CoT). There is growing evidence that this works on human-created data, but also on synthetically-created data, which enables LLMs to learn how to reason through a broader array of problems. This way, LLMs solve a problem in a step-by-step fashion.

What’s next?

Over the next year, we’re excited to see LLM reasoning make big strides, helping to create specialized AI models that businesses can really put to use. At the same time, we’ll be tackling important challenges like data security, fairness, and reliability, so that industries like finance, healthcare, and legal can fully benefit from these advancements.

LLM reasoning sets new benchmarks for AI