arXiv q-bio.NC: ApproxBrain argues neural sufficiency depends on task and tolerance

arXiv q-bio.NC: ApproxBrain argues neural sufficiency depends on task and tolerance

The authors of ApproxBrain argue that a connectome—the available neural substrate—does not by itself show how much of it a computation needs. They replace a single reported minimum circuit with cost-indexed sets of neuron subsets, conditioned on task and tolerance, and test candidates on held-out parameter draws. On a released walking simulator, they report that a published 148-neuron circuit is sufficient with margin, while no 146- or 147-neuron subset of it is sufficient; removing any one of its three interneurons destroys adequacy. The neurons needed also vary with the task: the authors’ best-found solutions for two tasks share only 3 of 11 searchable neurons. The authors report that relaxing tolerance yields alternative sufficient circuits; the whole-CNS interface alone fails every replicate, while a search-chosen set of 93 neurons yields a certified solution. These results make sufficiency a set of possibilities shaped by task and tolerance, rather than a single universal circuit. The authors emphasize that every result is model-relative and that a frontier is interpretable only alongside its interface-only baseline, tolerance, and search budget.

Publication: 6e0e7f48-2b63-48ea-86a6-a827898d922f

attribution — pass: The reading attributes the framing, results and conclusions to the ApproxBrain authors in an arXiv q-bio.NC paper, consistent with the abstract page. The source's caveats, including model-relativity, remain attributed to the authors.

Authors:Ali Mahdavi, Azadeh Zamanifar, Amirfarhad Farhadi, Omid Kashefi

summary — pass: The title, summary and question faithfully convey the argument that sufficiency is set-valued and depends on task and tolerance, along with the main results. The line 'These results make sufficiency a set of possibilities' follows the authors' own conclusion. It could be attributed to them explicitly. 'Tolerance' (permitted degradation and failure fraction) and 'whole-CNS interface' are not explained in the text. They should get term-library tooltips. This is not an accuracy error. The omission of the search-strategy comparison does not distort the reading.

Sufficiency is set-valued and conditioned jointly on task and tolerance, and no single search strategy dominates across regimes.

date — pass: The originalPublished date of 2026-09-22 matches the v1 submission date on arXiv. The later feed date of 2026-11-06 is correctly not used.

[Submitted on 22 Sep 2026]

support — pass: Every claim's excerpt is verbatim from the abstract and supports its claim. The figures for 148, 146/147, three interneurons, 3 of 11, and 93 neurons match the source. The limitations correctly note that the results are model-relative and that only the abstract is available. The question follows from the stated dependence on task and tolerance without implying an answer. The reading adds no benefits or costs beyond the source.

the whole-CNS interface alone fails every replicate while 93 search-chosen neurons give a certified solution

media — pass: The selected media is the verified direct link to the original arXiv abstract page. No other source media was found, so this is the most direct access to the source.