Startup Founded by Former DeepMind Researchers Says Its Small Model Beat Claude and GPT-5.5 at Replicating Research

A four-person team that left Google DeepMind says it has built an AI agent smaller than the flagship models from Anthropic and OpenAI, yet more effective at a specific, practical scientific task: independently reproducing the results of published research papers without being given the answers in advance.

Inherent is a London startup based in King's Cross, a district that has become one of Europe's most important AI hubs in recent years, partly thanks to Google DeepMind's headquarters at 6 Pancras Square. The company was founded by Edward Hughes, who serves as chief scientist, along with Louis Kirsch, Kaloyan Aleksiev and Tantum Collins, all of whom gained their experience at DeepMind.

How the test works

Instead of measuring general language ability or math problem-solving, Inherent focused on a narrower but practical skill: reproducing experiments from published research papers. The agent is given the text of a paper but not its results in advance, and must independently design the code, run the experiment, and arrive at conclusions close to the original.

To evaluate this skill, the company built its own benchmark called Replica, covering 310 tasks derived from 100 peer-reviewed papers in fields such as natural language processing, meta-learning, structural biology, materials science and weather forecasting. According to Inherent, Faraday produced more faithful reproductions of results than its competitors across every category of paper tested.

Small model, big result

The size contrast is the key point. Faraday is built on Qwen 3.6, a 27-billion-parameter model, a fraction of what powers flagship systems like Anthropic's Claude Opus 4.8 or OpenAI's GPT-5.5. Inherent says the advantage came not from scale but from the training method: long-horizon reinforcement learning in which the agent writes and runs code as a tool, with its progress judged by an automatically generated rubric assigned to each task, rather than by another large language model acting as judge.

The company describes the effect of this training as building the model's "scientific intuition", an instinct for which experiments are worth running and how to design them rigorously. In its research post, Inherent also describes techniques for stabilizing training over long action sequences, referred to as aggregating multiple rollouts and assigning credit at the level of individual agent steps.

Faraday is learning the skills of a rigorous scientist, which is a step toward artificial intelligence capable of innovating across many fields - Inherent Labs, company research post.

Many PhD students start their research careers exactly this way - Edward Hughes, chief scientist at Inherent.

A note of caution

The comparison comes solely from Inherent's internal benchmark, not from an independent, reproducible third-party test. The company has not yet published detailed percentage scores for each model, and Anthropic and OpenAI have not publicly commented on the result. Until an independent party reproduces these comparisons on the same set of tasks, the announcement should be treated as a startup's claim rather than a confirmed industry fact.

Hughes himself stresses that research replication is a means to an end for Inherent, not the goal itself. The company aims to build systems capable not only of reproducing existing knowledge but, eventually, of generating new scientific discoveries, alongside parallel work on world models.

What it means for the market

Inherent's result fits into a broader trend in which smaller, specialized models trained for a specific task are starting to compete with general frontier models in narrow niches, despite having far fewer parameters and much smaller training budgets. It signals to companies investing in scientific agents that an edge doesn't necessarily come from model size, but from the quality of training data and the design of the reward.

For Polish research institutions and biotech companies considering automating parts of their research work, Inherent's example points to a direction for tools that support result reproducibility, a topic regularly criticized in the scientific community as a weak point of published research. An agent able to quickly verify whether a paper's results can be reproduced could become a practical tool in peer review and grant assessment, though for now it remains a demonstration by a single startup rather than a finished commercial product.

In the coming months, Inherent plans to grow its team from roughly 12 people today to 20-25 employees and continue developing further versions of Faraday, betting that reliable replication will become the foundation for more ambitious applications, including independently running experiments that lead to new discoveries.