Teaching Language Models to Forecast Research Success Through Comparative Idea Evaluation

In three linesResearchers train language models to forecast empirical success of research ideas before experimentation. On 11,488 idea pairs from PapersWithCode, an 8B model reaches 77.1% accuracy via SFT, outperforming GPT-5 (61.1%). RLVR approach generates interpretable justifications with 71.35% accuracy.

Summary generated by Claude — human-verified