Nicola Angius, Pietro Perconti, Alessio Plebe, Alessandro Acciai (2025), Frontiers in Artificial Intelligence, vol. 8 (Hypothesis and Theory) — DOI-10.3389/frai.2025.1509338
Key takeaways
- Current explanations of how transformer LMs achieve linguistic competence use the same three strategies cognitive science uses for humans — functional analysis, mechanistic explanation, and brain co-simulation — so the “explanatory gap” for NLMs is of a familiar kind.
Seeded from the abstract and a skim, before the reading in NOTE-106. What follows is what the work says about itself.
This is an epistemological analysis of attempts to explain how the fairly simple algorithmic parts of neural language models produce real linguistic competence. After introducing the Transformer, the authors argue that philosophy of AI has moved from Turing’s “can machines think?” to “how can machines think?”, which exposes an explanatory gap. They argue that the explanatory strategies now used on NLMs match those cognitive science uses on human intelligence. Behavioural probes of theory of mind, discourse-entity tracking and property induction count as functional analysis. The copying algorithm and induction heads give a mechanistic explanation of in-context learning. Work predicting brain activation from NLMs is described as “co-simulation”, in which model and brain are each used to understand the other.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 35 seconds of active reading over 1 session. It was filed Deferred, unread. NOTE-106 is the close reading of 2026-09-26, and it placed the work: Active — worth reading only as a philosopher’s map of interpretability methods sorted by explanation type, and for the “co-simulation” label; the argument is thin, the Transformer primer contains a technical error, and a reader wanting the underlying results should go to the cited primary work (Elhage et al. 2021, Olsson et al. 2022, Caucheteux et al.).
When seeded, before the reading in NOTE-106, the entry gave a priority for a deeper reading: medium — 35 s dwell. It is directly about explaining transformer behaviour, but it surveys rather than argues deeply, and the skim captures its taxonomy.
What the seeding asked a deeper reading to check (NOTE-106 is that reading):
- A philosophy-of-science map of interpretability practice (behavioural, mechanistic, neuro-alignment), useful for placing interpretability work in the anthology’s explanation topics.
- Check whether the paper argues that any one strategy is sufficient or privileged, or only that all three mirror cognitive science.
- Check how “co-simulation” is defined and whether it answers the worry that NLM–brain correlations are predictive but not explanatory.
Access when seeded: Fetched the open-access Frontiers HTML full text with curl (CC BY). It gives received 2024-10-10, accepted 2025-03-17 and published 2025-04-01. Read the abstract, keywords, the section list, and the openings of §§3–7 including the concluding remarks.