Inherent's Faraday AI Outperforms OpenAI and Anthropic in Replicating Scientific Research

TL;DR

• British AI lab Inherent has unveiled Faraday, an AI teammate designed to autonomously replicate complex scientific papers from scratch, claiming it outperforms top models from OpenAI and Anthropic on research replication benchmarks.
• Faraday scored significantly higher on PaperBench and internal replication tests by combining code generation, experiment execution, and self-correction, rather than just producing text or code snippets.
• Experts see reliable paper replication as a critical stepping stone toward fully autonomous AI scientists, signaling a shift in the AI race from chatbots to agents capable of driving real-world discovery.

Who Is Inherent and What Is Faraday?

Inherent is a London-based AI lab founded by a team of former DeepMind researchers. While the company has operated largely in stealth since its founding, its mission has been clear: to build AI that can do science, not just talk about it.

That mission took a major step forward this week with the unveiling of Faraday, which the company describes not as a chatbot or copilot, but as an AI teammate. Unlike general-purpose large language models designed to answer questions or write code on demand, Faraday is built to take a scientific paper as input — including its methods, figures, and results — and autonomously attempt to reproduce it end-to-end.

This means reading the paper, writing the necessary code, gathering or synthesizing datasets, running experiments, debugging failures, and comparing its own results to those claimed in the original publication. The goal is to create an agent that can function like a skilled PhD student or postdoc.

How Faraday Actually Works

According to Inherent, Faraday's advantage comes from its agentic architecture rather than just raw model scale. The system is built on top of a powerful foundation model but wraps it in a framework designed for long-horizon scientific work.

The process starts with deep paper parsing, where Faraday extracts not just the text but the implied methodology, hyperparameters, and experimental logic that are often missing or ambiguous in published papers. It then moves to autonomous experiment planning, breaking the replication into a series of executable steps.

Crucially, Faraday can execute code in a sandboxed environment, run experiments, and observe the results. If an experiment fails or produces results that don't match the paper, it enters a self-correction loop — diagnosing bugs, searching for missing details, adjusting parameters, and re-running the work without human intervention. This closed-loop of reasoning, acting, and verifying is what Inherent says separates Faraday from standard LLMs that can generate plausible-looking code but cannot test if it actually works.

The company also emphasizes tool use, giving Faraday access to scientific libraries, data analysis tools, and the ability to browse documentation, mimicking how a human researcher would troubleshoot a replication.

Benchmark Results: Outperforming OpenAI and Anthropic

The headline claim from Inherent is performance. On PaperBench, a leading benchmark developed to test an AI's ability to replicate AI research papers from the ground up, Faraday has reportedly set a new state-of-the-art.

PaperBench tasks an AI with replicating 20 cutting-edge machine learning papers from scratch and grades it on whether the code runs, whether the experimental methodology is correct, and how closely the final results match the original paper's claims. It is considered one of the most difficult evaluations for AI agents because it requires sustained reasoning over many hours and thousands of lines of code.

Inherent reports that Faraday achieved a replication score of 42.5% on PaperBench, compared to 26.1% for Anthropic's Claude 4 Opus and 18.7% for OpenAI's o3 model under the same conditions. On a separate, more recent internal benchmark of 30 papers spanning biology, physics, and materials science, the company claims a similar lead.

While these results have not yet been independently verified by third parties, the margin is notable. Previous top models have struggled to get beyond 25% on PaperBench, often failing at the code execution and debugging stages. Inherent says Faraday's ability to iteratively fix its own errors was the key differentiator.

Why Replicating a Paper Is Such a Big Deal

At first glance, replicating a paper might sound less impressive than writing a new one. In reality, AI researchers consider it a far harder and more important test.

Reproducibility is the bedrock of science, but many published papers lack complete code, omit crucial implementation details, or contain small errors. A human expert often needs days or weeks to successfully replicate a single paper, filling in the gaps through intuition and trial-and-error.

For an AI to do this autonomously, it must demonstrate true scientific understanding, not just pattern matching. It has to infer unstated assumptions, handle ambiguous instructions, and ground its reasoning in empirical results. Success here suggests an AI can reliably follow the scientific method.

This capability is widely seen as a prerequisite for the next stage: AI-driven innovation. Before an AI can be trusted to design novel experiments, discover new materials, or propose new theories, it must first prove it can faithfully reproduce what humans have already done. Replication is the gateway to automation of the entire research cycle.

What This Means for Automated Science and the AI Race

If Faraday's performance holds up under independent scrutiny, it could accelerate the timeline for automated science. A reliable replication engine could be used to rapidly verify new research, audit published findings for errors, and serve as a foundation for AI systems that can then iterate and improve upon existing work.

For labs and universities, such a teammate could dramatically speed up R&D by handling the time-consuming work of reproducing baselines and running ablation studies. For industry, it points toward AI agents that can turn scientific literature directly into working code and products.

The announcement also intensifies the competitive landscape. While OpenAI, Anthropic, and Google DeepMind have focused heavily on general reasoning, coding, and multimodal chatbots, Inherent is betting on deep specialization in scientific agency. Its emergence highlights a growing trend of smaller, specialized labs in the UK and Europe challenging the dominance of US giants by targeting high-value scientific use cases rather than building ever-larger general models.

Inherent has not yet announced when Faraday will be widely available, saying it is currently being tested with a small group of academic and industry partners. The company plans to release a technical report and open-source a subset of its evaluation tasks in the coming weeks, which will allow the broader research community to stress-test its claims.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.