Alejandro H. Artiles and colleagues built a sampler that searches for plausible research ideas outside familiar research patterns. It breaks 16,068 LLM papers into 273 reusable concepts, called “idea atoms.” Two small models score whether combinations fit together and whether existing author communities seem positioned to propose them. An LLM writes up the selected combinations.

The authors report:

  • Broader portfolios: among 300 proposals, Alien scores 225.2 on an overlap-based measure of effective diversity. Claude Opus 4.7 scores 56.5 when prompted separately and 197.9 when generating the whole batch together.
  • Comparable human ratings: Alien and Claude tie at 4.03/5 overall, based on just 10 ideas per method and 40 blinded reviews total.
  • A novelty trade-off: models built only from pre-2025 papers recover 29 strictly new 2025 concept combinations using plausibility alone. Adding moderate pressure toward unfamiliar combinations raises that to 40, while reducing total recovered combinations from 762 to 453.

The often-quoted 35.9-times-chance recovery comes from plausibility alone. Searching further from familiar research communities changes what gets found; it does not improve every measure. Most larger quality comparisons use model judges, and batch-prompted baselines inherit quality ratings from separate runs rather than receiving their own ratings.

This is evidence for diversifying research proposals, with small early experiments behind some of them. Whether the extra directions become productive science remains an open question. The captured version is v3, revised October 21; its corpus and method differ substantially from the March preprint.

The 44-comment thread on Hacker News adds a concrete warning about applying the idea beyond its evaluated field.

What the thread adds

  • negura — says biology proposals in the related Lacuna explorer looked nonsensical to them. That is criticism of explorer output, separate from this paper’s LLM-research evaluation.
  • martincsweiss — identifying as a co-author, acknowledges worse results in sparse areas such as biology. They argue full-paper access matters because abstracts leave out key concepts.
  • jkhdigital — warns that a dense cluster of incremental publications can indicate a field stuck repeating itself. Publication volume alone is a weak guide to what deserves further work.

HN handles are pseudonymous, no per-comment scores are public, and ordering is HN’s ranking. This is a slice of the thread, not a consensus.