Inherent's Faraday AI Agent Outperforms Larger Models in Scientific Paper Replication

Faraday's Performance in Scientific Replication

Inherent, an AI lab established by former Google DeepMind employees, has unveiled its AI agent, Faraday, which has demonstrated superior performance in replicating the findings of published scientific papers. This task involves independently reproducing research results without prior knowledge of the outcome, a standard exercise for human scientists.

Faraday's success in this area indicates its capability to handle complex scientific verification tasks, a foundational step towards Inherent's broader goal of developing AI for discovering new scientific knowledge.

Outperforming Larger AI Models

A notable aspect of Faraday's achievement is its ability to outperform significantly larger and more established models, specifically Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5. These competitor models are considered frontier-scale systems with a much greater number of parameters.

Faraday operates on a comparatively small model called Qwen 3.6, which has only 27 billion parameters. This efficiency in performance relative to its size suggests a different approach to AI development that could lead to more resource-effective solutions.

The Role of Reinforcement Learning

Inherent attributes Faraday's success to its training methodology, which heavily relies on reinforcement learning. This approach rewards the AI system for achieving positive outcomes rather than providing explicit rules or instructions.

This method was also used to instill "research taste" in Faraday, which refers to the AI's ability to discern which experiments are valuable and how to design them effectively. This intangible quality is crucial for scientific inquiry and represents a significant advancement in AI's capacity for nuanced decision-making.

Implications for AI Development

The performance of Faraday, particularly its efficiency compared to larger models, highlights a potential shift in AI development strategies. It suggests that effective AI solutions may not always require the largest models or most extensive computational resources.

Inherent's focus on reinforcement learning for scientific tasks could pave the way for AI agents that can contribute across various scientific fields, moving beyond mere verification to active knowledge discovery.