Faraday - Inherent's AI Scientist Agent

What is Faraday?

Faraday, an AI scientist agent developed by the London-based AI lab Inherent, is trained using reinforcement learning based on the Qwen 3.6 model with 27 billion parameters. Like a human PhD student, Faraday can autonomously read papers, design experiments, and reproduce research results without prior knowledge of the answers. Faraday outperforms Claude Opus 4.8 and GPT-5.5 on the paper reproduction benchmark Replica.

Faraday's main functions

• Reproduction of self-written papersRead academic papers with obscured results graphs, infer experimental methods, and independently design and execute a reproducible experiment within 60 minutes or more of limited GPU resources.
• Scientific Experiment DesignWhen faced with experiments that cannot be fully reproduced, the ability to scale them down proportionally while remaining faithful to the core arguments of the original paper demonstrates "research sophistication."
• Tool Call ProgrammingUse Codex GPT-5.5 as an external coding tool, take responsibility for scientific decision-making, and let a professional programming agent handle code execution.
• Cross-domain generalizationTrained on ML papers and tested on AI-for-Science papers, it has cross-domain transfer capabilities.

Faraday's technical principles

• Base model selectionFaraday is built on Qwen 3.6-27B with only 27 billion parameters. The team trained it with reinforcement learning to give it scientific decision-making capabilities, and the code implementation was completed by external tools, which verifies that scientific intelligence and engineering intelligence can be decoupled.
• Coding Agent as a Tool (CAT) architectureFaraday uses a shell tool to call Codex GPT-5.5 as an external programming assistant; Faraday focuses on high-level scientific decision-making, including understanding papers, designing experimental schemes, and allocating computing resources, while Codex is responsible for the specific code writing, debugging, and execution.
• Replica Task SpaceThe training data comes from automatically generated replica benchmarks, covering 100 well-known papers from 1990 to 2026 with a total of 310 reproducible tasks; the task is to automatically identify and mask the result graphs in the papers using Gemini 2.5 Pro, and the agent needs to autonomously reproduce the results within a strict limit of 60 minutes and 1/7 of an H200 GPU.
• Rubric-based Judge Reward System:To address the challenge of reproducing tasks with no definitive answer, the system first uses Claude Opus 4.7 to automatically generate a unique scoring standard for each task (hiding the original gold plot to prevent overfitting). Then, Codex GPT-5.5 acts as the execution judge, reviewing the code repository, git history, and interaction records within 10 minutes, and providing a reward signal of 0–1 points based on five dimensions: visual matching, support for scientific claims, experimental fidelity, resource utilization efficiency, and scientific integrity.
• Long-term reinforcement learning trainingFaraday employs a modified GRPO algorithm with LoRA fine-tuning (rank=128), and performs post-training in a 128K token context window with a learning rate of 6×10⁻⁶. Each batch is uniformly sampled to cover papers from different eras, ensuring that training is not dominated by a single scientific paradigm and achieving stable optimization of long-term non-verifiable tasks.

How to use Faraday

Faraday is currently in the research and release phase and has not yet launched any public APIs, product interfaces, or user experience portals.

Faraday's core advantages

• Small Model, Big CapabilitiesQwen 3.6, based on 27 billion parameters, outperforms Claude Opus 4.8 and GPT-5.5 in the paper reproduction task, proving that scientific intelligence does not rely on ultra-large-scale models.
• Tool decoupling architectureBy adopting the Coding Agent as Tool (CAT) paradigm, Faraday focuses on scientific decision-making and delegates the code implementation to Codex GPT-5.5, achieving an efficient division of labor between scientific intelligence and engineering intelligence.
• Study on taste cultivationBy learning "research taste" through reinforcement learning, it can still design scaled-down experiments that are faithful to the original paper when resources are limited, rejecting hard-coding and cheating shortcuts.
• Cross-domain generalizationIt maintains a task win rate of over 60% when trained on ML papers and tested on AI-for-Science papers, demonstrating strong cross-domain transfer capabilities.
• Scalable benchmarkBased on Replica, a task space is automatically generated, expanding from 100 papers to 310 reproducible tasks, providing a scalable data foundation for AI scientists to train.
• Stable long-range RL
  Successfully extended GRPO reinforcement learning to long-term, non-verifiable research tasks, achieving stable training through per-task rubric judge and turn-level credit assignment.

Faraday's project address

• Project official website：https://inherentlabs.ai/research/training-to-replicate
• arXiv technical paper：https://arxiv.org/pdf/2608.13331

Comparison of Faraday's similar products

Faraday's application scenarios

• Reproduction and verification of academic papersResearch institutions can use Faraday to batch verify the reproducibility of published papers, solving the "reproducibility crisis" in the field of machine learning.
• Doctoral research trainingAs a "virtual research assistant," it helps doctoral students cultivate experimental design and scientific taste by reproducing classic papers.
• AI-for-Science Experimental ValidationAutomatically reproduce experimental results from interdisciplinary papers in fields such as materials science, weather forecasting, and bioinformatics.
• Research benchmarksProvides a standardized replica evaluation framework for AI scientist agents to measure the capabilities of different models in scientific research tasks.
• Research Methodology AuditJournal reviewers or grant review agencies can use Faraday to quickly verify the authenticity and reproducibility of experimental results in submitted papers.