Inherent releases Faraday, an AI agent trained to reproduce research papers
The London startup's 27-billion-parameter model, trained via reinforcement learning, reportedly beat Claude Opus 4.8 and GPT-5.5 on Replica, its own new 310-task reproduction benchmark.
- Models & capabilities
- Benchmarks & progress
- Notable
Inherent, a London-based startup, published a paper describing Faraday, a 27-billion-parameter “AI Scientist” agent trained through long-horizon reinforcement learning to reproduce the results of published machine-learning and AI-for-science papers. Alongside the model the company introduced Replica, a benchmark of 310 tasks drawn from 100 papers spanning natural-language processing, materials science and weather forecasting. Each task asks an agent to recreate a specific figure from a paper’s reported results without access to the original plot, under limited time and compute budgets.
Inherent reported that Faraday produced more faithful replications than Claude Opus 4.8 and GPT-5.5 — both much larger models — across every category in the Replica suite, with what the company described as a particularly pronounced advantage in meta-learning, structural biology and materials science. The system pairs Faraday, which plans experiments and judges progress, with a separate coding agent it calls as a tool to write and run the implementation code. The paper credits much of the training’s tractability to an automated, rubric-based judge tuned to agree with human assessors, which supplied the reward signal for reinforcement learning on open-ended replication tasks. Authors on the paper include Damon Falck and Edward Hughes.
The outperformance claim is self-reported: it comes from a single paper, with no independent benchmarking, and Replica itself was designed, built and scored by the same team reporting the result. The release drew some pickup from technology-industry news aggregators, but no outside group had independently verified the reproduction results at the time of publication.
The paper appeared the same day Hugging Face published results from a separate, crowdsourced effort using general-purpose coding agents to check whether thousands of ICML papers reproduce — a different approach to the same underlying question of whether AI systems can verify published science at a scale humans cannot.