Inherent's Faraday agent leaves competitors behind in research replication task

Summary

Inherent's AI agent Faraday, using a compact 27-billion-parameter model, independently reproduced scientific research papers and outperformed larger models from Anthropic and OpenAI.

Why it matters

The result shows that smaller, cost-effective models can match or even surpass the performance of large-scale systems on a critical scientific task. It also highlights Inherent's innovative use of reinforcement learning to develop 'research curiosity,' a step toward AI not only validating but potentially generating new scientific insights.

Highlights

• Faraday surpasses Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 in research replication.
• Operates on the relatively small Qwen 3.6 model (27B parameters), demonstrating efficiency.
• Uses reinforcement learning to instill 'research curiosity' for better experimental design.

Critical take

The achievement is limited to reproducing existing work; the transition from replication to original discovery remains untested, and reliance on external tools like OpenAI's Codex may constrain the agent's true independence.