A 27B AI scientist trained to replicate research

Faraday is a 27B AI Scientist trained with RL to replicate scientific research and direct coding agents across long-horizon research tasks.

Faraday is a 27-billion-parameter agent trained to replicate results from scientific papers. Rather than simply answering questions about a paper, it must reconstruct experiments and reproduce a given figure within limited time and compute budgets, without access to the original plot.

For this training, Inherent developed Replica, a set of 310 tasks drawn from 100 machine learning and AI-for-science papers. The domains range from natural language processing and materials science to structural biology and weather forecasting. Each task requires the agent to reconstruct an experimental process that is only partially described in the paper.

This difficulty is central to the project. Replicating a result is not simply about producing a similar-looking figure: the agent must choose the right hypotheses, adapt the experiment to available constraints, evaluate intermediate results, and abandon certain approaches. Inherent describes this set of decisions as a form of “research taste,” which Faraday learns through an LLM judge validated with human experts, task-specific rubrics, and a credit assignment system applied across different stages of the trajectory.

Faraday also uses coding agents as tools. Inherent says it can direct GPT-5.5 Codex to carry out part of the technical work, while the 27B agent retains the role of scientific supervision. The model can also generalize to directing a more capable coding agent at inference time after being trained with GPT-5.4-mini.

According to evaluations published by Inherent, Faraday produces more faithful replications than Claude Opus 4.8 and GPT-5.5 across every category in Replica, with particularly pronounced gains in meta-learning, structural biology, and materials science. The baselines were run in Claude Code and Codex respectively, with high reasoning effort.

The team is also testing whether these learned skills can extend beyond replication. To do this, paper variants include results that were not present in the original research. Inherent reports that Faraday outperforms GPT-5.5 on these “imagined” tasks without receiving a reward at test time or using a harness specifically designed to explore new hypotheses. However, the experiment still takes place in an environment constructed by the team and, on its own, does not establish a general capability for autonomous scientific innovation.