Jumbly Grindrod, J. D. Porter, Nat Hansen (2024), arXiv preprint (cs.CL); also a chapter in Cappelen & Sterken (eds.), Communicating with AI: Philosophical Perspectives, Oxford University Press — ARXIV-2405.12084
Key takeaways
- The holism of distributional/LLM meaning does not succumb to the instability objection, because what matters for communication is “differential” instability (changes in relative distances), which is constrained by frequent, consistent use and turns out to be a feature, not a bug.
Seeded from the abstract and a skim, before the reading in NOTE-158. What follows is what the work says about itself.
LLMs rest on distributional semantics, and the distributional hypothesis carries a holistic view of word meaning: a word’s meaning depends on its relations to the other words in the model. Holism faces the classic instability objection, that any change in a system’s meaning properties ripples through the whole system. The authors separate absolute instability (shifts in where points sit in a vector space) from differential instability (shifts in the relative distances between points), and argue that only the second matters for communication. They build two small language models, show how each kind of instability appears as the training corpus grows, and argue that a word’s resistance to change scales roughly with how frequently and consistently it is used. Differential instability then allows productive meaning change without the damage the objection alleges.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 0 seconds of active reading over 0 sessions. It was filed Deferred, unread. NOTE-158 is the close reading of 2026-09-26, and it placed the work: Proposed — promising as a philosophical reply, but its empirical half rests on top-10 neighbour tables for four hand-picked words, with no stability metric and no seed replication. It would be settled by a systematic measurement of neighbour-set stability against word frequency across corpus increments and random seeds, which Antoniak & Mimno (2018) partly supply and the paper cites but does not reproduce.
When seeded, before the reading in NOTE-158, the entry gave a priority for a deeper reading: medium — never opened (t = 0), but it is a well-grounded philosophical treatment of LLM meaning with small, reproducible experiments. The skim captures the thesis; the experimental details in §4 would repay a closer look.
What the seeding asked a deeper reading to check (NOTE-158 is that reading):
- Bears directly on the anthology’s representation topics. Relative distances as the invariant corresponds to the ML practice of comparing embedding spaces up to rotation or alignment (e.g. Antoniak & Mimno 2018 on embedding stability, which the paper cites).
- Check whether the argument carries over from static word embeddings to contextual transformer representations; the experiments use count and Word2vec models only.
- Check how §5 handles the sufficiency objection, i.e. whether distributional relations suffice for meaning at all.
Access when seeded: philpapers.org is behind Cloudflare. Found the work on arXiv (v1 submitted 2024-05-20, v2 2025-04-03). Read the abs page and the v2 PDF (25 pp): abstract, section heads, §4 opening and method, §5 opening, §6 conclusion. The OUP chapter DOI comes from Crossref, which gives the chapter’s issued date as 2026-09-29. I have not checked the chapter text against the preprint.