Inherent Says Faraday Beats Larger AI Models at Replicating Research

According to TechCrunch, Inherent, a London AI lab founded by Google DeepMind alumni, says its Faraday agent outperformed larger Anthropic and OpenAI systems on a task: reproducing findings from published scientific papers. The result tests whether an agent can choose experiments, use tools and persist through a research workflow—not whether it is universally superior.

A Benchmark Built Around Reproduction

Inherent describes Faraday and its Replica research programme as a test of whether an AI system can recover a paper’s results without seeing the original plot or being handed the answer. The initial suite contains 310 tasks drawn from 100 machine-learning and AI-for-science papers across natural-language processing, materials science and weather forecasting. The accompanying arXiv paper provides the formal account of the benchmark and training approach.

That setup is more demanding than asking a model to summarize a paper, but it is not the same as independently discovering a new scientific result. It measures experimental discipline and coding, not whether a system can generate valuable hypotheses with no predefined answer.

Small Model, Larger Supervisory Role

Faraday runs on Qwen 3.6 with 27 billion parameters, according to Inherent. The company says the model was trained with long-horizon reinforcement learning, using coding agents as tools rather than relying only on a conventional chat response. In practice, the agent’s performance depends on the interaction between the model, its tool-use harness, the reward signal and the available compute budget. The Qwen project materials offer useful context for the base-model family, but do not independently validate Inherent’s Faraday results.

Inherent compared Faraday with Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, run through their respective Claude Code and Codex harnesses. The company says Faraday produced more faithful replications across the task categories. That is a claim about this benchmark and configuration, not evidence that a 27B model is generally more capable than frontier systems.

Teaching an Agent to Have Research Taste

The more interesting design choice is Inherent’s attempt to train “research taste”: knowing which experiments are worth running and how to use limited time and compute. Its research notes say the system uses an LLM judge and a human study to test whether the evaluation captures more than visual similarity. That matters because a plot can look close while the experiment is poorly designed, the method is unfaithful or the conclusion is unsupported.

Reinforcement learning is a logical fit for that problem because the desired behavior is difficult to express as a fixed checklist. But reward design can also narrow an agent toward what an evaluator recognizes. Researchers will want to see held-out tasks, reproducible code, compute accounting and failure analysis before treating the result as evidence of general scientific reasoning. OpenAI’s reinforcement-learning primer explains why reward signals shape behavior without removing the need for careful evaluation.

From Paper Replication to Discovery

Inherent says replication is a stepping stone toward AI scientists that can work across domains. The logic is plausible: papers describe successful results, while the path to those results includes failed hypotheses, discarded experiments and resource trade-offs. An agent that can reconstruct some of that hidden process may be more useful than one that only produces a polished explanation.

The gap between a benchmark and scientific usefulness remains substantial. Reproducibility, data access, laboratory constraints, attribution, safety review and human accountability all matter outside a controlled task suite. The company’s public research agenda is ambitious, while the reported $50 million seed round gives it room to pursue the longer experiment. Its next proof point will be whether external researchers can reproduce the benchmark and find value beyond it.

That question connects to real-world agent evaluation, agent-building methods, enterprise AI agents moving from conversation to action, security controls for tool-using systems and models for physical-world environments. Faraday’s reported result is notable because it shifts attention from parameter count to process. It will be durable only if that process survives independent scrutiny.