Inherent Says Its Faraday AI Can Reproduce Research With a Smaller Model

London startup Inherent says its Faraday AI agent has outperformed larger frontier systems at a specific scientific task: reproducing results from published research without being given the answer in advance.

The company was founded by former Google DeepMind researchers and is building what it calls inventive AI for scientific discovery. Its longer-term goal is not simply to summarize papers, but to help researchers choose questions and design useful experiments.

A smaller model inside a larger research system

Inherent says Faraday uses Qwen 3.6, a 27-billion-parameter model, while competing against systems based on larger models from Anthropic and OpenAI. The agent can also call existing tools, including coding systems, rather than attempting to perform every part of the research process with one model.

The company describes research replication as a useful test because it requires more than retrieving an answer. An agent must interpret a paper, plan experiments, run them and decide whether the evidence supports the original result.

Inherent says it trains for what it calls research taste: choosing experiments that are worth running instead of blindly maximizing a narrow score. Its approach uses reinforcement learning to reward successful research behavior.

Promising, but still a company claim

The result should be interpreted carefully. Inherent reported the comparison, and independent researchers have not yet established that Faraday performs better across broader scientific domains or real laboratory settings. Replicating known work is also different from producing a genuinely new and useful discovery.

Still, the project illustrates an important direction for AI development. Scientific agents may depend less on simply scaling one language model and more on combining models, tools, experimental planning and reliable verification.

Inherent is structured as a public benefit corporation and says it wants AI to augment human researchers rather than remove them from discovery. Whether Faraday can turn a controlled replication result into dependable scientific collaboration will be the more meaningful test.