Locks and Keys: Learning Scientific Inspiration Retrieval as Pattern Matching at Scale

A new hypothesis often arises when a research problem meets complementary knowledge not previously associated with it. Making such discoveries thus begins with inspiration retrieval: finding, among the vast body of existing knowledge, the piece that fits the problem. The task looks unlearnable, since discoveries are new connections by definition and the ones a model must find at test time are absent from its training data. Yet it is learnable, and we propose why: inspiration retrieval is essentially pattern matching. A research problem is a lock and an inspiration is a key; individual pairs are new, but the way a problem's structure fits an inspiration's structure recurs, so the rule can be learned from records of past discoveries by ordinary supervised fine-tuning. Experiments bear this out: what a trained model learns is the fit, not a memory of similar pairs and not resemblance between problem and paper. The view also says how to learn the task: as a score on each problem–paper pair, computed by embedding and reranking models, rather than as a selection over a candidate set by a generative reasoning model, as in prior large-scale training. Given the same base model and annotations, scoring pairs is far more accurate than generative selection, at a fraction of the cost, and stays ahead as accuracy grows log-linearly over hundreds of thousands of training pairs. Finally, the learned fit uses only a small fraction of the embedding dimensions available, and the fraction does not increase with more training data: the task is limited not by representation capacity but by how much of the fit has been learned.