Blind Benchmark Catches Frontier AI at Just Three Percent on Research Idea Recovery

A new benchmark tests frontier AI large language models on their ability to reconstruct research ideas using only reference lists from scientific papers, revealing uniform low performance between 3% and 15%. A multi-agent pipeline improves recovery rates to 23-42%, indicating current models lack reliable abductive reasoning but that collaborative approaches can enhance hypothesis generation.

This Wire brief sits within Fusion42's coverage of AI Frontier Models.