Large language models are being adopted for drug discovery — screening candidate molecules, predicting binding affinities, generating novel compounds. The assumption is that LLMs, with their vast training data and general reasoning capabilities, outperform traditional machine learning pipelines (random forests, gradient boosting, molecular fingerprint methods) on drug discovery tasks.
The paper (arXiv:2603.12349, March 2026) measures the marginal value of LLMs over traditional ML pipelines for drug discovery scoring. The marginal value is zero. Traditional pipelines, properly calibrated and using domain-specific molecular features, match or exceed LLM performance on every task tested. The LLM adds no predictive value beyond what the simpler, cheaper methods already provide.
The result is not that LLMs perform badly. They perform well. The result is that they perform no better than methods that cost orders of magnitude less to train, deploy, and run. The discovery pipeline doesn't benefit from the additional capability — the bottleneck is elsewhere (experimental validation, not computational screening), and the computational screening is already solved at the required accuracy by traditional methods.
The practical consequence: budget spent on LLM inference for drug screening is budget not spent on experiments. If the LLM doesn't improve screening accuracy, the money is wasted — or rather, it is spent on capability that exceeds the task's requirements. Overspecification of the computational component degrades the overall pipeline by misallocating resources.
The structural lesson: a tool's capability and its value are different quantities. Value is capability minus the capability of the next-best alternative, multiplied by the sensitivity of the downstream process to that capability difference. When the next-best alternative is adequate and the downstream process is insensitive to marginal improvements in screening, the most powerful model in the world has zero marginal value. Power without a gap to fill is overhead.