Faraday: An integrated model AI agent outperforms giant models in reproducing scientific research

British AI lab “Inherit”, founded by former Google DeepMind researchers, announced that its new intelligent agent “Faraday” was able to outperform globally leading models such as “Claude Opus 4.8” from Anthropic and “GPT 5.5” from OpenAI in a task of independently reproducing results from published scientific papers without being provided with prior answers. This announcement comes a few weeks after the London lab emerged from a secretive phase with a seed funding round of fifty million dollars.

What is interesting about this experiment is that the agent “Faraday” does not rely on a massive, ultra-large model, but runs on the “Qwen 3.6” model, which has only 27 billion parameters, a very small fraction of the size and training costs of competing models. Co-founder and chief scientist at the lab, Edward Hughes, explains that reproducing scientific research results is merely a first step that mirrors what human PhD students begin with, within a broader goal of building an intelligent agent capable of discovering new scientific knowledge rather than merely verifying existing results.

Building research acumen through reinforcement learning instead of imposing strict rulesPure accuracy was not the only benchmark for the lab’s success; the team focused on giving the agent what they called research acumen, the ability to identify experiments worth pursuing and to design them efficiently. To achieve this, the lab employed a reinforcement-learning approach that rewards good outcomes rather than providing explicit instructions for scientific inquiry, betting on this method’s capacity to generalise across various scientific disciplines in the future.

Simulating the behavior of a curious scientific colleague and leveraging existing software toolsThe lab did not aim to build bespoke software tools for coding; instead it directed the agent “Faraday” to use the “GPT-5.5 Codex” model from OpenAI to perform programming tasks, just as human scientists rely on available software to conduct their research. Hughes stresses that the goal is to avoid creating agents that merely echo what a user wants to hear, and to develop an agent that acts as a team colleague who initiates experiments out of scientific curiosity, then presents the results for discussion and evaluation.

The lab’s team consists of twelve on-site staff working from an office in King’s Cross, London, with plans to expand to around twenty to twenty-five employees before the end of the year. Hughes also touched on talent-acquisition challenges, calling for an end to the so-called garden-leave system in Britain, which imposes restrictions that prevent departing researchers from joining competing companies for several months, giving US startups an advantage in attracting talent and hiring quickly.