The Open Source "Agentic Reasoning" Beats Google Gemini Deep Research

A deep dive into how the "Agentic Reasoning" framework works and the techniques behind it that make it outperform the most advanced reasoning LLMs today.

Image generated with DALL-E 3

Deep Research is getting popular.

With Google releasing an LLM agent that can impressively complete complex multi-step research tasks, this trend was soon followed by OpenAI and then Perplexity.

Now, there’s another exciting update, which surprisingly is open-source.

A team of researchers from the University of Oxford has just put out a framework called Agentic Reasoning that uses external tool-using LLM agents to solve multi-step logical problems that require deep research.

The results of this framework are surprisingly good, so much so that it outperforms leading RAG systems and closed source LLMs on PhD-level scientific reasoning (GPQA) and domain-specific deep research tasks.

The framework even beats Google Gemini Deep Research across multiple open-ended Q&A tasks in finance, medicine and law!