Researchers develop a novel evaluation method using Future Alignment Score to assess whether AI-generated research proposals predict actual future scientific directions.

New approach reframes research proposal generation as a time-sliced scientific forecasting problem, solving the challenge of automatically measuring novelty and soundness

Future Alignment Score (FAS) evaluates proposals by checking if they anticipate research directions that actually appear in papers published after a cutoff date

Dataset includes 17,771 papers with time-consistent citations and synthesized reasoning traces to train models on gap identification

Method provides a verifiable, scalable alternative to costly large-scale human evaluation of LLM-generated research proposals