Inherent trains 27B Faraday agent to replicate research

Inherent has introduced Faraday, a 27-billion-parameter AI scientist agent trained to reproduce results from research papers rather than simply summarize them. The accompanying research paper and Inherent research post describe Replica, a benchmark and reinforcement-learning task space built around a difficult scientific skill: reconstructing an experiment when important implementation details are missing.

Replica contains 310 tasks drawn from 100 machine-learning and AI-for-science papers. In each task, a result figure is removed and the agent must recover it under limited time and compute. That forces the system to do more than write code. It has to interpret an underspecified method, choose experiments, diagnose failures and decide whether a result is faithful to the original paper.

Faraday is based on a 27B model and is post-trained with long-horizon reinforcement learning. Crucially, it can use coding agents as tools. Inherent says the training setup teaches the smaller model to direct a more capable coding system while adding a layer of scientific judgment about what to try next and how to evaluate the result. The company reports that Faraday scores above Claude Opus 4.8 and GPT-5.5 on its held-out paper-replication tasks.

That headline needs an important qualification. Faraday is not a self-contained 27B model defeating frontier systems in a clean model-versus-model comparison. Inherent explicitly says Faraday can employ GPT-5.5 Codex as a tool. The claimed advantage is therefore about orchestration and learned scientific decision-making inside the Replica setup, not proof that a 27B base model has broader capability than GPT-5.5 or Claude Opus 4.8.

The evaluation also depends on a rubric-based automated judge. Replicating a scientific figure is not perfectly verifiable from a single numeric answer, so the researchers generate task-specific rubrics and use a judge designed to score experimental design, scientific practice and faithfulness to the source paper. They report that the rubric judge is more consistent and aligns better with human ratings than a simpler LLM-judge baseline. Even so, the benchmark and training framework are author-produced, so independent reproduction will matter.

The broader significance is the choice of replication as a training target. Scientific papers often omit failed experiments, intermediate decisions and low-level implementation details. A system that can recover those missing steps must explore hypotheses rather than merely follow an explicit procedure. Inherent argues that this makes replication a useful curriculum for agents that may later perform more open-ended research.

Faraday is therefore interesting less as a benchmark upset and more as a systems pattern for AI science: a relatively small learned controller, long-horizon reinforcement learning, coding agents as tools, and an evaluation process that rewards scientific judgment. If the approach generalizes beyond the current task suite, it could shift part of AI-for-science development away from larger monolithic models and toward trained orchestration layers that know how to use powerful tools rigorously.

## Sources

- Inherent: Training AI Scientists to Replicate Research
- arXiv: Training AI Scientists to Replicate Research