Inherent: Small-Scale AI Surpasses Giants in Paper Replication Task

AFFILIATE_PRODUCTS:

AI Lab Founded by DeepMind Alumni

AI research lab Inherent, based in London, UK, was established by graduates of Google DeepMind. On August 22, 2026, the company announced that its self-developed AI agent, “Faraday,” has surpassed large models from Anthropic and OpenAI in a specific task: independently replicating the content of academic papers.

Small-Scale Model Beats Frontier Models

At the heart of the announcement is the disparity in model scale. Faraday operates on a relatively small model called “Qwen 3.6,” which has 27 billion (27B) parameters. In contrast, competing systems like Anthropic’s “Claude Opus 4.8” and OpenAI’s “GPT-5.5” are far more massive, frontier-scale systems. According to a report by Anna Heim of TechCrunch AI, Inherent’s achievement is notable given the significant difference in parameter count.

Reinforcement Learning to Cultivate

“Research Taste”

The challenge was not merely about competing on replication accuracy. Inherent aimed to imbue AI with “research taste” — the intuition human scientists possess for identifying valuable experiments and designing them effectively. Edward Hughes, the company’s co-founder and chief scientist, identifies reinforcement learning as the key to reaching this goal.

“What we found most interesting was not just the result of beating a frontier agent, but actually the method by which we built this agent.”

Reinforcement learning is a method where an AI system receives rewards for producing correct outcomes. Instead of being explicitly taught rules, it acquires general capabilities through trial and error. Inherent judges that this reward-based approach, rather than learning scientific methodology itself, can generalize to building agents that contribute to diverse fields of science in the future.

Design Philosophy Leveraging Existing Tools

Notably, Inherent does not develop its own general-purpose coding tools, instead using OpenAI’s GPT-5.5 Codex for Faraday. This is based on the philosophy that, just as human scientists do not build all software from scratch but use existing tools, the agent should do the same. This case illustrates the delineation between “what to build” and “what to use” in agent development.

Editorial Opinion

Short-Term Impact This announcement poses a question mark against the “scale supremacy” of models in AI agent development. The fact that a 27B parameter model outperforms frontier models on specific tasks could accelerate the development race toward more efficient and cost-effective task-specialized agents within the next six months. Investors and researchers will likely begin to re-focus not only on models with large parameter counts but also on architectures and training methods specialized for specific purposes.

Long-Term Perspective Over a 1- to 3-year span, the critical question will be whether Inherent’s vision of “scientific knowledge discovery” directly translates into commercial success. Paper replication is a prerequisite; true value lies in hypothesis generation and experiment design. The ability to generalistically apply the acquisition of “research taste” via reinforcement learning across diverse scientific fields will likely become the watershed for long-term success. The automation of science by AI is not merely an agent technology problem; it is also a sociotechnical challenge that includes building trust with the scientific community.

References

• “Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research”, by Anna Heim — TechCrunch AI, 2026-08-22T19:00:00.000Z (ARR)
• Source URL: https://techcrunch.com/2026/08/22/inherent-founded-by-deepmind-alumni-says-its-ai-teammate-just-outperformed-anthropic-and-openai-at-replicating-research/