Found first: a primary source the press has not covered yet.
Researchers at Carnegie Mellon University introduced IdeaTaste, a benchmark that scores LLM-generated research proposals against papers the same communities actually chose to publish in 2026. Across more than 90,000 proposals from six models in four research areas, models reached about 26% of human performance on citation matching and about 18% on idea matching. The paper is on arXiv.
What the source says
IdeaTaste freezes a literature snapshot at 2025, asks six contemporary LLMs to generate research proposals, then scores those proposals against what the same communities published in 2026, using citation matching and idea matching as metrics. The dataset covers 90,000-plus proposals across four research areas spanning established to emerging fields. Models scored about 26% of human level on citation matching and about 18% on idea matching. In blind expert review, AI proposals ranked lowest on "expected success" and highest on "presentation," pointing to a gap in research direction rather than in writing quality. The authors report that neither expanded context nor larger model populations closed the gap.
Why it matters
Benchmarks that use human raters to judge research proposals reward fluency and plausibility, which LLMs produce readily regardless of whether the direction is productive. IdeaTaste replaces that criterion with what research communities actually chose to publish, a target that well-structured writing cannot game. The roughly one-in-five idea-matching rate means models are systematically missing the directions a field will pursue, and expert reviewers attribute this to weaker directions rather than overlooked opportunities. The continuously updated leaderboard will track whether future model generations close the gap against live publication cycles.