Training AI Scientists to Replicate Research · inherent

inherent

Research

Training AI Scientists to Replicate Research

14th August 2026

Read the paper↗

In 1821, Richard Phillips asked his friend Michael Faraday to write a review of the emerging field of electromagnetism. Faraday was a strange choice for the task. The British scientist had little formal education and limited experience in the field. But he was a well-known experimentalist.

To write his review, Faraday decided to replicate past results manually. In the course of his experiments from the candlelit basement of the Royal Institution, Faraday found he could make a wire carrying current move around a magnet. “Very satisfactory”, he wrote in his journal. Faraday had invented the electric motor.

Inspired by Michael, we introduce Faraday, a 27B-parameter “AI Scientist” agent that outperforms Claude Opus 4.8 and GPT-5.5 on the task of replicating research. Trained via long-horizon RL with coding agents as a tool, Faraday learns the skills of a rigorous scientist, a step towards AI Scientists capable of innovation across domains.

Faraday outperforms Claude Opus 4.8 (“Claude”) and GPT-5.5 (“Codex”) on paper replication tasks.

From replication to innovation

To train Faraday, we introduce Replica, a scalable space of RL tasks. Each task requires an agent to replicate a figure from a research paper with a limited time and compute budget, and without access to the original plot. The initial suite comprises 310 tasks from 100 ML and AI for science papers, in domains as diverse as natural language processing, materials science and weather forecasting.

We run recent agents from the leading labs on Replica and find that they do not saturate the task space. For Claude Opus 4.8 and GPT-5.5 baselines, we run the model in the Claude Code and Codex harnesses respectively, with thinking effort set to extra high. Faraday produces more faithful replications for every category of paper in the task suite, and struggles less with recent research, effectively applying its scientific skills to work unseen by the base model at pre-training.

Faraday replicates more effectively than baseline agents across domains, with a particularly pronounced advantage in meta-learning, structural biology and materials science.

Replicating a figure may not seem like an especially innovative task. But looking at the process, rather than the output, replication becomes a stepping stone towards innovation. Research papers describe what the authors found that worked, not the negative results that got them there. To succeed, agents must recover the “99% perspiration” that does not appear on the page. This requires the hypothesis-driven exploration characteristic of open-ended research.

Replication forms the basis for a curriculum of underspecified tasks. Features beyond single plots can be removed from papers given to agents to replicate. Resource constraints can be further tightened or relaxed. Papers can even be imagined, leading the same Faraday model to innovate without knowing it.

RL for non-verifiable domains

Perfectly reproducing a plot is not the same as a successful replication, which also requires strong experimental design, good scientific practice, faithfulness to the claims of the original paper, and effective use of available resources. Ultimately, replication requires “research taste”.

We design an LLM judge and run a human study to validate that it captures the research taste of experts. But training on LLM judges remains challenging, because the stochasticity of the underlying model creates a noisy reward signal. We use per-task rubrics to solve this problem, achieving greater consistency and lower noise than an LLM baseline.

Our auto-generated rubric judge is more consistent and agrees more with humans.

To address the instability typical of long-horizon RL training, we make two further train-time modifications to our judge: multi-sample aggregation and turn-level credit assignment. Trained using this recipe, Faraday learns to become a more rigorous scientist.

Scalable scientific oversight

Faraday employs GPT-5.5 Codex as a tool, much like human scientists use coding agents. Remarkably, Faraday directs the work of a model several orders of magnitude larger in a way that improves replication performance.

Moreover, Faraday can generalise to directing a more capable agent at test-time, adapting to GPT-5.5 Codex after training with GPT-5.4-mini. As frontier coding agents continue to advance, we expect that the value of scientific judgement will only increase.

Faraday builds on previous discoveries to find new insights at test-time, similarly to existing AI Scientist agents. But unlike these agents, Faraday requires no hand-coded evolutionary harness, and has no test-time reward. In other words, Faraday learns to value discoveries intrinsically.

Faraday uncovers insights without the use...