Research

Paper

TESTING February 24, 2026

Test-Time Training with KV Binding Is Secretly Linear Attention

Authors

Junchen Liu, Sven Elflein, Or Litany, Zan Gojcic, Ruilong Li

Abstract

Test-time training (TTT) with KV binding as sequence modeling layer is commonly interpreted as a form of online meta-learning that memorizes a key-value mapping at test time. However, our analysis reveals multiple phenomena that contradict this memorization-based interpretation. Motivated by these findings, we revisit the formulation of TTT and show that a broad class of TTT architectures can be expressed as a form of learned linear attention operator. Beyond explaining previously puzzling model behaviors, this perspective yields multiple practical benefits: it enables principled architectural simplifications, admits fully parallel formulations that preserve performance while improving efficiency, and provides a systematic reduction of diverse TTT variants to a standard linear attention form. Overall, our results reframe TTT not as test-time memorization, but as learned linear attention with enhanced representational capacity.

Metadata

arXiv ID: 2602.21204
Provider: ARXIV
Primary Category: cs.LG
Published: 2026-02-24
Fetched: 2026-02-25 06:05

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.21204v1</id>\n    <title>Test-Time Training with KV Binding Is Secretly Linear Attention</title>\n    <updated>2026-02-24T18:59:30Z</updated>\n    <link href='https://arxiv.org/abs/2602.21204v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.21204v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Test-time training (TTT) with KV binding as sequence modeling layer is commonly interpreted as a form of online meta-learning that memorizes a key-value mapping at test time. However, our analysis reveals multiple phenomena that contradict this memorization-based interpretation. Motivated by these findings, we revisit the formulation of TTT and show that a broad class of TTT architectures can be expressed as a form of learned linear attention operator. Beyond explaining previously puzzling model behaviors, this perspective yields multiple practical benefits: it enables principled architectural simplifications, admits fully parallel formulations that preserve performance while improving efficiency, and provides a systematic reduction of diverse TTT variants to a standard linear attention form. Overall, our results reframe TTT not as test-time memorization, but as learned linear attention with enhanced representational capacity.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CV'/>\n    <published>2026-02-24T18:59:30Z</published>\n    <arxiv:comment>Webpage: https://research.nvidia.com/labs/sil/projects/tttla/</arxiv:comment>\n    <arxiv:primary_category term='cs.LG'/>\n    <author>\n      <name>Junchen Liu</name>\n    </author>\n    <author>\n      <name>Sven Elflein</name>\n    </author>\n    <author>\n      <name>Or Litany</name>\n    </author>\n    <author>\n      <name>Zan Gojcic</name>\n    </author>\n    <author>\n      <name>Ruilong Li</name>\n    </author>\n  </entry>"
}