<?xml version="1.0" encoding="UTF-8"?>
<OAI-PMH xmlns="http://www.openarchives.org/OAI/2.0/" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://www.openarchives.org/OAI/2.0/ http://www.openarchives.org/OAI/2.0/OAI-PMH.xsd">
  <responseDate>2026-11-20T19:44:35Z</responseDate>
  <request verb="GetRecord" identifier="oai:lifeofresearch.com:229" metadataPrefix="oai_datacite">https://lifeofresearch.com/oai2</request>
  <GetRecord>
    <record>
      <header>
        <identifier>oai:lifeofresearch.com:229</identifier>
        <datestamp>2026-11-15T11:05:38Z</datestamp>
        <setSpec>type:preprint</setSpec>
        <setSpec>openaire</setSpec>
      </header>
      <metadata>
        <resource xmlns="http://datacite.org/schema/kernel-4" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://datacite.org/schema/kernel-4 http://schema.datacite.org/meta/kernel-4.4/metadata.xsd">
          <identifier identifierType="DOI">10.5281/zenodo.24235210</identifier>
          <alternateIdentifiers><alternateIdentifier alternateIdentifierType="URL">https://lifeofresearch.com/preprints/229</alternateIdentifier></alternateIdentifiers>
          <creators>
            <creator>
              <creatorName>Pranay Mahendrakar</creatorName>
              <nameIdentifier nameIdentifierScheme="ORCID" schemeURI="https://orcid.org">0009-0003-7224-029X</nameIdentifier>
              <affiliation>SONYTECH</affiliation>
            </creator>
          </creators>
          <titles><title>A Relocated Dependency Is Not a Vanished One: What the Neural-Cellular-Automaton Result on Integer Multiplication Shows About Long-Range Dependency, Why the Neural GPU Line Already Bounds It, and the Controls That Would Separate Locality, Supervision and Carry Coverage</title></titles>
          <publisher>Life of Research</publisher>
          <publicationYear>2026</publicationYear>
          <subjects>
            <subject>integer multiplication</subject>
            <subject>length generalization</subject>
            <subject>long-range dependency</subject>
            <subject>neural cellular automata</subject>
            <subject>algorithm learning</subject>
            <subject>carry propagation</subject>
            <subject>transformers</subject>
            <subject>circuit complexity</subject>
          </subjects>
          <resourceType resourceTypeGeneral="Preprint">Preprint</resourceType>
          <rightsList><rights rightsURI="https://creativecommons.org/licenses/by/4.0/">Creative Commons Attribution 4.0</rights></rightsList>
          <descriptions><description descriptionType="Abstract">(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0.
Life of Research
Multi-digit multiplication is a standing failure case for neural sequence models, and a 2025 reverse-engineering study traced the failure of standard fine-tuning to long-range dependencies that the model never learns. A 2026 paper rejects that diagnosis: it lays two binary integers out as an outer-product grid, trains a 321-parameter neural cellular automaton on single update steps, and reports exact products up to 4,096 bits after training on 2 to 6 bits, which it reads as proof that the long-range dependency was a mirage of the chosen representation and architecture. In this analysis the author reviews and synthesises published results to show that the two papers are mostly not in conflict, because &quot;long-range dependency&quot; names three different things. The functional dependency of a high product bit on many input bits is a property of multiplication under any encoding; the reach of a single update step is a property of representation and architecture, and the grid removes it; the number of sequential steps is the price of that removal, and the automaton pays it, taking about n steps for n-bit operands. The paper then argues that the published record already bounds the mirage claim. Local, iterated networks learned binary multiplication from input-output pairs alone in 2015, including a one-dimensional variant that reached 2,000 bits from 20; the same line failed on long carry chains that uniformly random test inputs almost never contain; and on both sides of the dispute, most of the reliable successes supervise intermediate states. The paper consolidates the measured values in Table 1 and Figure 1, states an attribution procedure as Algorithm 1, and specifies six controls, including a carry-adversarial test set and an end-to-end-trained automaton, that would separate locality, supervision and carry coverage.</description></descriptions>
        </resource>
      </metadata>
    </record>
  </GetRecord>
</OAI-PMH>