Scientific ideation — the process of generating research questions and hypotheses — is increasingly mediated by large language models. But according to a new arXiv preprint, the systems built for this task are usually trained and evaluated on what it calls "immediately judgeable proxies": novelty, clarity, and feasibility.
The preprint, "Learning to Ideate for Scientific Impact" (arXiv:2609.29802), states that this focus leaves something open. The abstract in the announcement is truncated at exactly that point, so the authors' proposed direction is not visible from the excerpt alone.
Even so, the significance is clear: if ideation systems are optimised for what is easy to judge, they may be optimising for the wrong target. The paper's title signals a move toward scientific impact as the ultimate criterion, but the details of that move are not yet available in the abstract.