Research

Paper

AI LLM February 19, 2026

The Cascade Equivalence Hypothesis: When Do Speech LLMs Behave Like ASR$\rightarrow$LLM Pipelines?

Authors

Jayadev Billa

Abstract

Current speech LLMs largely perform implicit ASR: on tasks solvable from a transcript, they are behaviorally and mechanistically equivalent to simple Whisper$\to$LLM cascades. We show this through matched-backbone testing across four speech LLMs and six tasks, controlling for the LLM backbone for the first time. Ultravox is statistically indistinguishable from its matched cascade ($κ{=}0.93$); logit lens reveals literal text emerging in hidden states; LEACE concept erasure confirms text representations are causally necessary in both architectures tested, collapsing accuracy to near-zero. Qwen2-Audio genuinely diverges, revealing cascade equivalence is architecture-dependent, not universal. For most deployed use cases, current speech LLMs are expensive cascades, and under noise, they are worse ones, with clean-condition advantages reversing by up to 7.6% at 0 dB.

Metadata

arXiv ID: 2602.17598
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-02-19
Fetched: 2026-02-21 18:51

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.17598v1</id>\n    <title>The Cascade Equivalence Hypothesis: When Do Speech LLMs Behave Like ASR$\\rightarrow$LLM Pipelines?</title>\n    <updated>2026-02-19T18:22:39Z</updated>\n    <link href='https://arxiv.org/abs/2602.17598v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.17598v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Current speech LLMs largely perform implicit ASR: on tasks solvable from a transcript, they are behaviorally and mechanistically equivalent to simple Whisper$\\to$LLM cascades. We show this through matched-backbone testing across four speech LLMs and six tasks, controlling for the LLM backbone for the first time. Ultravox is statistically indistinguishable from its matched cascade ($κ{=}0.93$); logit lens reveals literal text emerging in hidden states; LEACE concept erasure confirms text representations are causally necessary in both architectures tested, collapsing accuracy to near-zero. Qwen2-Audio genuinely diverges, revealing cascade equivalence is architecture-dependent, not universal. For most deployed use cases, current speech LLMs are expensive cascades, and under noise, they are worse ones, with clean-condition advantages reversing by up to 7.6% at 0 dB.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='eess.AS'/>\n    <published>2026-02-19T18:22:39Z</published>\n    <arxiv:comment>10 pages, 6 figures, 7 tables</arxiv:comment>\n    <arxiv:primary_category term='cs.CL'/>\n    <author>\n      <name>Jayadev Billa</name>\n    </author>\n  </entry>"
}