Reed Bender, Karina Kofman, Blaise Agüera y Arcas, Michael Levin (2025), arXiv preprint — ARXIV-2505.15849
Key takeaways
- When 68 expert definitions of life are scored pairwise by three LLMs and clustered, they form a continuous landscape with eight clusters, two thirds of them in overlapping “Cognitive Autonomy” and “Dissipative Self-Organizing Systems” clusters. The authors argue this makes “what is life?” a question about a multidimensional space rather than a binary demarcation.
Seeded from the abstract and a skim, before the reading in NOTE-109. What follows is what the work says about itself.
No definition of life commands universal agreement, and synthetic biology, AI and astrobiology make the question more pressing. The authors collected definitions of life from a curated set of cross-disciplinary experts and used large language models to score how similar each pair of definitions is. They then applied agglomerative clustering, LLM-driven analysis of the definitions within each cluster, and t-SNE projection. The result is a continuous landscape of themes, not a set of discrete camps. On that basis they recommend treating the definition of life as a set of perspectives within one conceptual space, and offer the method for other contested definitions.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 5 seconds of active reading over 1 session. It was filed Deferred, unread. NOTE-109 is the close reading of 2026-09-26, and it placed the work: Active — worth reading as a dataset: 68 signed, three-sentence definitions of life from named researchers, with a manual thematic coding (Table 1). Its computational conclusions are weakly supported and are better read as illustrations than as findings.
When seeded, before the reading in NOTE-109, the entry gave a priority for a deeper reading: low — t=5 s. The skim captures the result, and its evidential weight is limited by the LLM-judgment method and the curated sample.
What the seeding asked a deeper reading to check (NOTE-109 is that reading):
- A data point for nucleation notes on life–mind continuity and on agency-based definitions of life. It is relevant alongside LIT-167 (Baedke), LIT-120 (autopoiesis) and LIT-195 (Safron, Levin et al.).
- The method is the weak point to check. The “correlations” are LLM similarity judgments, not measurements. Model agreement shows the LLMs agree with each other, not that the clusters are real. The respondent pool was curated by the authors, many of them from Levin’s network, so the “emerging consensus” on cognitive autonomy may reflect who was asked.
- Check whether the published version (if any) changed the analysis or the number of respondents from v1.
Access when seeded: arXiv abstract page (v1 submitted Mon 19 May 2025; a v2 exists) and the full PDF (54 pp.). Read the title page and keywords, abstract, §1 Introduction, the structure of §2 Background and §3 Methods, §4 Results (pairwise correlations, Table 2 cluster consensus definitions), the opening of §5 Discussion, §7 Conclusion, and Table 1 (manual categorization). Journal publication not checked; treated as a preprint.