Scientific progress relies on replication, the process of reproducing experiments to verify their validity . While general-purpose AI models have long been used for data analysis, a new specialized agent is attempting to automate the actual process of scientific discovery.
Inherent, a London-based startup founded by DeepMind alumni, recently emerged from stealth mode with a $50 million seed round [S4, S5]. The company has introduced Faraday, an AI agent designed to independently reproduce the findings of published scientific papers without knowing the expected answers in advance [S5, S6].
How Faraday Differs from Frontier Models
Most leading AI systems rely on massive scale to achieve high performance. However, Inherent claims that Faraday outperforms much larger frontier models, specifically OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8, at replicating research [S1, S4, S6].
This performance gap is notable because of the difference in architecture. Faraday runs on Qwen 3.6, a model with 27 billion parameters [S4, S5]. This is significantly smaller than the architectures used by its rivals [4].
To achieve these results, Inherent used a specific benchmark called Replica, which consists of 310 tasks drawn from 100 AI-for-science and machine-learning papers [7]. While the company reports superior performance, these results have not yet been independently verified [S2, S5].
The Concept of “Research Taste”
Inherent is not just optimizing for the correct final answer. Co-founder Edward Hughes explains that the goal is to develop “research taste,” which is the ability to identify which experiments are worth conducting and how to design them effectively [S4, S5].
Why does this matter? A system that simply follows a fixed procedure can verify a result, but a true scientific collaborator must decide which variables to test and when the evidence is strong enough to justify a conclusion [5].
To cultivate this instinct, Inherent used reinforcement learning [S4, S6]. Instead of training the AI on data about how scientists typically work, the system is rewarded for achieving useful outcomes [5]. This approach encourages the AI to prioritize experiments with high potential for scientific impact rather than following rigid, predefined rules [4].
A Modular Approach to AI Agency
Faraday does not attempt to be a monolithic tool that does everything. Instead, it follows a collaborative architecture that mirrors how human researchers operate [5].
For example, Inherent did not build a dedicated coding tool for Faraday. Instead, the agent utilizes OpenAI’s GPT-5.5 Codex for its coding tasks [S4, S5]. This allows the agent to leverage existing high-performance tools for specific functions while focusing its own reasoning capabilities on the scientific workflow [5].
This strategy suggests a shift in AI development. Rather than building a single massive model, the future of technical agents may lie in combining a smaller reasoning model with specialized tools, experiment environments, and evaluation loops [5].
Practical Implications for R&D Teams
For enterprises in drug discovery, materials research, or engineering, the emergence of specialized agents like Faraday introduces a new tradeoff between model size and specialized utility [5].
Smaller models can be cheaper to deploy, which is critical when a workflow requires repeated calls to a model or thousands of simulated experiments [5]. However, the total cost of an agent is not just the model size; it also includes the cost of tool use, data access, and human review [5].
For R&D teams, the primary value of such a system is not just speed, but the creation of a verifiable audit trail [5]. In sensitive scientific fields, an attractive result is less valuable than one that can be checked and reproduced [5].
As Inherent expands its team in London, the industry will be watching to see if this reinforcement-learning approach can scale across multiple scientific disciplines [5].
Sources
- Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just …
- DeepMind Alumni Startup Inherent Claims AI Teammate Beats OpenAI, Anthropic
- Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just …
- Inherent’s AI Outperforms Anthropic, OpenAI
- Inherent says its Faraday AI agent beat Anthropic and OpenAI at …
- DeepMind alumni’s AI beats OpenAI and Anthropic at research replication …
- Inherent Says Faraday Tops Claude, GPT-5.5 at Paper Replication - AI Weekly
