mlsanityData to 7 Oct 2026

The Cascade Equivalence Hypothesis: When Do Speech LLMs Behave Like ASRrightarrowLLM Pipelines?

arXiv 2602.17598 ↗First posted 19 Feb 2026cs.CL cs.AI eess.AScited by 5

Not on the map. Why?

Abstract (arXiv)

Speech LLMs are widely understood to be better than ASRrightarrowLLM cascades since they have access to the audio directly, and not just the transcript. In this paper, we present an evaluation methodology and a mechanistic interpretation of the observed behavior of speech LLMs. First, we introduce matched-backbone testing which separates out the behavior of the speech LLM from the reasoning capabilities of the underlying LLM. Second, we provide a mechanistic analysis of speech LLMs using logit lens and LEACE and show the literal transcript emerging from the LLM's hidden states and that text representations are causally necessary. We also show that in 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 0dB.

Key facts. Cited by 5 papers in the map (data through 7 Oct 2026). Builds on 23 of its 32 references that are in the map. First posted on arXiv 19 Feb 2026.

Builds on

23 of 32 references.

References · 23

  1. 2024The Llama 3 Herd of Modelscited by 14,192
  2. 2024Qwen2.5 Technical Reportcited by 4,713
  3. 2022Robust Speech Recognition via Large-Scale Weak Supervisioncited by 3,972
  4. 2020wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representationscited by 3,853
  5. 2017TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehensioncited by 2,895
  6. 2015Character-level Convolutional Networks for Text Classificationcited by 2,293
  7. 2016Understanding intermediate layers using linear classifier probescited by 1,279
  8. 2018MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversationscited by 636
  9. 2023Eliciting Latent Predictions from Transformers with the Tuned Lenscited by 579
  10. 2024Qwen2-Audio Technical Reportcited by 543
  11. 2023SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilitiescited by 518
  12. 2025Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAscited by 491
11 more
  1. 2015MUSAN: A Music, Speech, and Noise Corpuscited by 400
  2. 2023LEACE: Perfect linear concept erasure in closed formcited by 254
  3. 2024AIR-Bench: Benchmarking Large Audio-Language Models via Generative Comprehensioncited by 215
  4. 2019Towards Multimodal Sarcasm Detection (An _Obviously_ Perfect Paper)cited by 173
  5. 2024WavLLM: Towards Robust and Adaptive Speech Large Language Modelcited by 107
  6. 2023Dynamic-SUPERB: Towards A Dynamic, Collaborative, and Comprehensive Instruction-Tuning Benchmark for Speechcited by 99
  7. 2024Scaling Properties of Speech Language Modelscited by 30
  8. 2025Do Audio LLMs Really LISTEN, or Just Transcribe? Measuring Lexical vs. Acoustic Emotion Cues Reliancecited by 23
  9. 2025AudioLens: A Closer Look at Auditory Attribute Perception of Large Audio-Language Modelscited by 18
  10. 2025Closing the Gap Between Text and Speech Understanding in LLMscited by 14
  11. 2025Improving Causal Interventions in Amnesic Probing with Mean Projection or LEACEcited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 5.

5 papers.

  1. 2026Generating Synthetic Doctor-Patient Conversations for Long-form Audio Summarizationcited by 4
  2. 2026WASIL: In-the-Wild Arabic Spoken Interactions with LLMscited by 1
  3. 2026Interleaved Speech Language Models Latently Work In Textcited by 0
  4. 2026Inherited Heads: Audio language models track speakers with their text backbone's attention, and an attention-mass ranking retrieves a different setcited by 0
  5. 2026Transcribe, Then Reason: Two-Pass Decomposition for Multimodal Reviewcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Internalizing ASR with Implicit Chain of Thought for Efficient Speech-to-Speech Conversational LLM
  2. 2026Do Audio LLMs Listen or Read? Analyzing and Mitigating Paralinguistic Failures with VoxParadox
  3. 2023AudioChatLlama: Towards General-Purpose Speech Abilities for LLMs
  4. 2025Llasa: Scaling Train-Time and Inference-Time Compute for Llama-based Speech Synthesis
  5. 2026Rethinking Entropy Allocation in LLM-based ASR: Understanding the Dynamics between Speech Encoders and LLMs
  6. 2026Hearing Between the Lines: Unlocking the Reasoning Power of LLMs for Speech Evaluation
  7. 2026TRADE: Transducer-Augmented Decoder for Speech LLM
  8. 2026AOR-Bench: Do Large Audio Language Models Over-Refuse Pseudo-Harmful Queries?