Source-linked AI summary

Can We Read the Mind of an Audio LLM? A Verbalizable, Multilingual Middle-Layer Workspace

Jiajun Fan, Jingyuan Li, Prashanth Gurunath Shivakumar, Qi Luo, Jia-Hong Huang, M. Maruf, Roger Ren, Yile Gu, Rahul Pandey, Ge Liu, Ivan Bulyko

arXiv:2608.24958v1cs.SDcs.AIcs.CL

TL;DR

The paper investigates whether an audio LLM’s hidden, sound-derived reasoning can be read before it emits an answer, beyond what its prompt or transcription provides. It applies logit-lens probes, waveform swaps, paired audio-caption comparisons, and causal tests to Qwen3-Omni, finding a multilingual, paralinguistic middle-layer workspace whose content is used before output. The account is deliberately qualitative and scoped to one model family and limited data.

  • Problem

    Audio LLMs can answer spoken questions without writing chain-of-thought, leaving their intermediate, decision-relevant reasoning inaccessible to ordinary monitoring.

  • Method

    The study probes Qwen3-Omni’s audio-token positions with a logit lens and waveform-controlled comparisons, adding audio-versus-caption tests, activation patching, and layer deletion.

  • Results

    The readout exposes sound-driven, conceptual content before token emission, including inferred events, roles, sources, and affect, with the clearest separation from text priors in the middle workspace band.

  • Takeaways & Limitations

    Intermediate audio representations can provide a process-level interpretability signal for speech agents in cases where stated reasoning is absent.

  • Takeaways & Limitations

    The account is limited to one model family, a small sound-dominated slice, a cheap logit-lens proxy, and sparse readouts across audio positions.

Abstract

from arXiv · show

An audio language model is a black box in a specific way: we see what it says, never what it works out on the way there, and chain-of-thought monitoring helps only if the model writes its reasoning down. Reading a base Qwen3-Omni with a logit lens at the audio-token positions, we find that the answer to a spoken question becomes legible - in words - in the model's middle layers, before it emits any token. Five findings follow. (1) The readout carries concepts in neither the question, the options, nor the model's own transcription: on a clip whose verbatim transcription is empty garbling, it reconstructs Watergate and scandal, passes through the role president, and resolves to Nixon - a hidden multi-hop chain, read with no chain-of-thought. (2) The content is language-agnostic: one audio-inferred concept surfaces in several scripts at once, and 38% of top-1 readouts are Chinese on English inputs. (3) It is paralinguistic: given the same clip as audio and as the model's own emotion-free caption, the audio mind forms the sound source, speaker role, or affect that the caption discards, and answers correctly more often. (4) The audio-driven signal is absent at the input, turns on about a tenth of the way into the network, separates most cleanly from the text prior in the middle band (35-80% of depth), and activation patching shows it is causally used and committed before the last fifth of the layers. (5) Deleting single layers maps the pipeline: reading the sound in is localized to the entry layers and answer delivery to the output layer, while retrieval is distributed across the interior. Throughout, a waveform-swap control - identical text, only the sound changed - isolates the audio-driven signal from a prior over the printed options. This is a qualitative account of what an audio model works out before it speaks: the quantities are controls, not benchmark scores.

1 Introduction

The paper asks whether an audio LLM forms readable, task-relevant concepts before answering, rather than merely relying on the printed prompt. Using Qwen3-Omni and waveform-controlled intermediate readouts, it presents a qualitative account of an audio-driven middle-layer workspace.

  • Motivation and approach: The study reads Qwen3-Omni’s intermediate activations at audio-token positions before answer emission, exposing concepts that may be absent from the question, options, and transcription.The approach targets the hidden computation that chain-of-thought monitoring cannot observe when the model answers directly.
  • Motivation and approach: The waveform-swap control holds the question and options byte-identical while changing only the audio, separating sound-driven readouts from a text prior.The conditions use the real clip, a mismatched clip, or silence.
  • Five findings: The findings span hidden multi-hop reasoning, multilingual concepts, paralinguistic information, middle-layer formation, and distributed answer processing.The paper organizes its evidence into five findings rather than benchmark scores.
  • Scope: The authors scope the account as qualitative, based on one model family, a small sound-dominated slice, a logit lens, and sparse readouts across audio positions.Reported quantities function as controls against a text prior, not as benchmark performance scores.

2 Related Work

Prior work makes intermediate language-model representations readable and identifies a verbalizable middle-layer workspace, while audio studies probe semantic and auditory attributes. This paper extends that question to task-relevant audio reasoning with a waveform-swap control.

  • Reading transformer internals: The logit lens and tuned lens project intermediate states into vocabulary space, while circuit- and feature-level work studies how readable content is computed.These approaches motivate inspecting representations before final token production.
  • Reading transformer internals: Gurnee et al. describe a functional global workspace concentrated in middle layers, with sensory early layers and motor late layers surrounding it.The workspace contains verbalizable representations that can be reused, broadcast, and selectively recruited.
  • Latent language and audio lenses: Audio research includes a language-agnostic semantic hub and AudioLens analyses of surface auditory attributes such as gender and emotion.These lines establish that audio content and auditory attributes can be made lens-readable.
  • Novelty: The paper’s distinction is to test task-relevant global-workspace organization for audio and isolate sound-driven signals from printed-option priors with waveform swaps.It presents this control and the resulting localization as new to its knowledge.

3 Method

The method probes Qwen3-Omni’s 48-layer Thinker at audio-token positions with a logit lens, comparing controlled audio conditions and paired audio-versus-caption inputs. It combines rank-based workspace readouts with causal patching and layer deletion.

  • Model and lens: The model has 48 layers, width d=2048, vocabulary |V|=152,064, and untied output embedding WU; intermediate residual states are decoded with the terminal readout.The lens is applied prematurely to intermediate states.
  • Model and lens: Ranks measure each target token’s position in the decoded vocabulary distribution, with rank 1 denoting the top token and minimum rank reported across audio positions.Workspace-band layers are Thinker layers 17–38, approximately 35–80% of depth, selected a priori from the text-model account.
  • Assumptions: The analysis treats the logit lens as a cheap proxy rather than the more faithful Jacobian lens and therefore focuses claims on the middle band because early readouts are noisy.The four-way prediction is the option with minimum rank.
  • Controls: The audio-swap control keeps the written question and every option byte-identical while substituting real, mismatched, or silent waveforms.This design isolates audio-driven readouts from priors over the printed options.
  • Controls: The two-mind comparison presents the same clip as either the real waveform or the model’s emotion-free caption, with identical questions and options.This tests what information the sound contributes beyond its words.
  • Causal tests and data: Activation patching tests causal use by restoring clean residual streams at audio positions and selected layer bands, while layer deletion measures accuracy drops from identity skips.The core controls use a curated 140-clip MMAU set, with additional corpus-level pools and 500 spoken TriviaQA questions.
  • Figure 1: Figure 1 plots each concept’s logit-lens rank across layer depth, with audio positions followed by question and options on the input axis and darkness indicating better rank.The shaded band marks the workspace; event concepts strengthen earlier than actor concepts in the Nixon example.

4 Findings

The model forms readable, audio-derived concepts before answering, including inferred multi-hop content, multilingual representations, and paralinguistic information absent from captions. Controls and causal tests place this signal in the middle layers, while calibration cases and methodological scope constrain the interpretation.

  • Finding 1: We Can Read the Mind—No Chain-of-Thought Needed: The workspace reconstructs Watergate from water and gate, then surfaces scandal, president, and Nixon over audio positions without chain-of-thought.Nixon is decoded over audio positions rather than merely read from the printed options; failed ASR does not prove the words were never spoken.
  • Finding 1: We Can Read the Mind—No Chain-of-Thought Needed: Across seven curated cases, the workspace reads intermediate concepts before output, spanning multi-hop inference, speech-content recovery, and sound-source identification.The cases include excavation before Howard Carter, negotiation before Egypt and Israel, a spoken kidney detail, and direct identifications of lion, bird, music, train, and whip.
  • Finding 1: We Can Read the Mind—No Chain-of-Thought Needed: True answer concepts occupy 1–34% of workspace cells versus 0.12% for ten unrelated placebos, while only 4–30% of audio positions carry the concept.The sparse temporal signal motivates aggregation across audio positions; trumpet reaches 1.3% on the whip clip as a genuine acoustic resemblance.
  • Finding 2: The Thinking Space Is Multilingual: The workspace renders one audio-inferred concept in multiple languages: 38.5% of top-1 real-word readouts are Chinese on English inputs.Music appears simultaneously as English, Chinese, Spanish, German, and Italian; frequency controls show Chinese forms are clip-specific rather than globally frequent.
  • Finding 3: When Speech Carries Extra Information, the Audio Mind Registers It: Audio preserves sound source, speaker role, and affect that captions discard: audio answers correctly on roar and priest examples, and reads 33 of 39 affects absent from captions.On disagreement clips, correct speaker roles appear in 88.9% of audio readouts versus 70.4% for captions.
  • Finding 3: When Speech Carries Extra Information, the Audio Mind Registers It: A spoken TriviaQA boundary shows that audio does not always help when the clip repeats the written question, with text-side concepts outperforming speech-side concepts on discordant items.Among 87 items correct from text but wrong from speech, the answer concept appears in the text workspace on 83% versus 62% for speech; the authors treat this as a boundary, not a clean mechanism.
  • Finding 4: The Thought Lives in the Middle, and Forms in Depth Order: The audio-driven signal is strongest in the workspace band: real audio reaches 40.0% balanced accuracy versus 21.8% for silence, with a significant +18.2 gap there.The real−silence gap is negligible in the sensory band and intermediate in the motor band; layerwise significance begins at about 12% depth and remains open.

5 Conclusion

The study finds evidence for a verbalizable middle-layer workspace in an audio LLM, while deliberately scoping its claims to a qualitative, proxy-based account. It also identifies potential uses for workspace readouts in training and monitoring, pending broader tests.

  • 5 Conclusion: The audio LLM forms answer-relevant concepts before emitting tokens, with content that is multilingual, paralinguistic, and inferred from sound rather than merely retranscribed.The reported evidence includes waveform-driven concepts, causal use before output, and a functional map in which listening and delivery occur at opposite ends of the stack.
  • 5 Conclusion: The account is limited to a logit-lens proxy, one model family, a small sound-dominated slice, and sparse readouts across audio positions.The authors characterize the quantities as controls for ruling out a text prior, not benchmark scores.
  • 5 Conclusion: Activation patching localizes commitment before the motor band but does not establish that the workspace alone contains the causal mechanism.The authors call for corpus-scale replication under a faithful lens, larger forced-alignment sets, and tests on safety-relevant decisions before speech-agent actions.
  • 5 Conclusion: The workspace readout could provide a process-level training signal and a low-cost monitor for multimodal systems.The proposed training use complements scalar rewards over final answers, while the monitoring use relies on roughly one matrix product per read position.
Loading 2608.24958v1…