TL;DR — An audio LLM is a black box in a specific way: we see what it says, never what it worked out on the way there, and chain-of-thought monitoring only helps if the model writes its reasoning down. Reading a base Qwen3-Omni with a logit lens at the audio-token positions, the answer to a spoken question becomes legible — in words — in the middle layers, before a single token is emitted.

WATCH THE FILM
What a logit lens at the audio-token positions reads A grid of input positions (rows: audio tokens from the spoken clip, text tokens, and an empty answer slot) against the layers of a base Qwen3-Omni (columns, left to right). The forward pass fills the grid layer by layer. At a middle layer, the lens takes the states at the audio-token positions and decodes them with the model's own output head, and a word appears: the Chinese token for US president, which is the top-1 word at those positions. The audio-token positions come before the text and cannot see it. The answer slot is still empty, because no token has been emitted yet. Base Qwen3-Omni one cell = one position at one layer middle layers layer 1 last layer audio tokens the spoken clip cannot see the text text tokens answer Output head the model’s own 美国总统 top-1 word= “US president” 1Forward passlayer by layer, left to right 2Lens at the audio positionsdecode one middle layer 3Answer slot: still emptyno token emitted yet Legible in words, mid-stack, before a single token is emitted.
How to read it. Each column is a layer, each row a position in the input; the audio-token positions come before the question and options and cannot see them. Before anything is emitted, the logit lens takes the hidden states at the audio-token positions of a middle layer and passes them through the model’s own output head, so they read out as a word — here 美国总统 (“US president”), the role that reaches rank 1 at the audio positions of the Nixon clip below. The answer slot is still empty: nothing has been said yet. Layers and positions are drawn for illustration. Schematic.

A hidden multi-hop chain, with no chain-of-thought

On a clip whose verbatim transcription is empty garbling, the readout reconstructs Watergate and scandal; then the person and the role surface together: Nixon and president both appear across the transition into the middle layers, the role persisting there as Chinese 总统 — an event-then-actor chain, not a strict word-by-word order. Watergate and scandal appear in neither the question, the options, nor the model's own transcription; Nixon is a printed option, but it is decoded over the audio positions, not merely read off the choice. The reasoning happened; it simply was never written down.

Reading an audio LLM before it answers A spoken clip that the model's own transcription renders as empty garbling enters the model. Reading each layer at the audio positions, the event (water plus gate, scandal) becomes readable in the early layers; Nixon and the role president both by about a third of the depth, the role persisting through the middle layers. The model then answers Richard Nixon. workspace band (35–80% depth) input output spoken clip ASR: garbled water + gatescandalNixonpresident · 总统 answer Richard Nixon
How to read it. The question asks which president was forced to resign; the model's own transcription of the clip is empty garbling. Before it answers, a logit lens at the audio positions already reads the event (water + gate, scandal) in the early layers; Nixon and the role president both by about a third of the depth, the role persisting through the middle layers (event, then actor; not a strict word-by-word order). Schematic of the paper's flagship clip; layer positions are approximate.
2026-10-12T20:45:14.452775 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 0 20 35 50 80 100 depth into the network (%) Dallas Bird Whip Lion Train Music Kidney assassination (L30, 62%) niǎolèi (birds) (L33, 69%) biān (whip) (L32, 67%) lions (L19, 40%) rail (L19, 40%) música (L18, 38%) kidney (L17, 35%) workspace band
Figure 1. Where the task-relevant concept becomes readable, for seven curated cases. The strongest readout lands in the middle band of the network — and the word that surfaces is often not in the language of the question.

What the workspace looks like

Four properties of the workspace Four properties of the middle-layer workspace: it reads out as words; one concept surfaces in several scripts; it holds speaker and affect information a caption drops; and activation patching shows it is used. 1 Verbalizable it reads out as words WatergatescandalNixon 2 Multilingual several scripts president总统38% of top-1 in Chinese 3 Paralinguistic what a caption drops speakeraffect 4 Causal used by the model cleancorruptpatchanswerrestored
How to read it. Reading an audio LLM with a logit lens at the audio positions, the middle-layer state is verbalizable, multilingual (38% of top-1 readouts are Chinese on English inputs), paralinguistic, and causal under activation patching. Schematic; words from the paper's Nixon case.

A waveform-swap control runs throughout: identical text, only the sound changed. That is what separates a genuinely audio-driven signal from a prior over the printed answer options.

The waveform-swap control Two inputs carry identical printed text, a question with answer options A to D, but two different sounds. Both go through the same base Qwen3-Omni and are read with a logit lens. Each output forks into two possible explanations. If the readout were only a prior over the printed options, it would come out the same both times, for example option B twice. If it is audio-driven, it changes with the sound, for example option B for sound 1 and option D for sound 2. The swap is what separates the two. identical text only the sound changed same model If only a prior…over the printed options If audio-driven…from the sound QuestionABCD sound 1 + Qwen3-Omni + logit lens option B option B QuestionABCD sound 2 + Qwen3-Omni + logit lens option B option D = ≠ same both times changes with the sound Identical text, only the sound changed: only an audio-driven readout can differ.
How to read it. The two columns on the right are two possible explanations, not two outputs of the model. The control keeps the printed text identical and changes only the sound. A readout that only reflects a prior over the printed answer options cannot tell the two inputs apart and would come out the same (dashed); a genuinely audio-driven readout differs between the sounds. In the paper the readout follows the sound: it is best with the real clip, weaker with a mismatched clip, and at chance with silence. Option letters are placeholders. Schematic.

Deleting layers maps the pipeline

Reading is correlational; deleting is causal. Bypassing one layer at a time gives a clean functional map: listening is localized at the entry, delivery at the output, and retrieval is spread across the interior — no single interior layer is essential.

2026-10-12T20:45:14.432212 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ L0 encoder to text workspace band (middle layers) any other interior layer 0 20 40 60 80 accuracy lost when the layer is deleted (pts) 65 pts at most 10 pts at most 10 pts
Figure 2. Accuracy lost when a single layer is deleted (identity skip, averaged over clips). Only the encoder-to-text entry layer is critical; every interior layer costs at most ten points.

The same tooling shows where hallucinations come from: in text mode the wrong concept usually takes over early, and is never revised afterwards.

Why it matters, and what it is not

If what a model is about to say is legible mid-stack, monitoring need not wait for it to speak, and need not trust it to narrate its own reasoning. But this is a deliberately qualitative account of one model family: a readable correlate is not a complete mechanism, and the paper is conservative about what the lens can and cannot establish.

BibTeX

@article{fan2026readmind,
  title   = {Can We Read the Mind of an Audio {LLM}? A Verbalizable,
             Multilingual Middle-Layer Workspace},
  author  = {Jiajun Fan and Jingyuan Li and Prashanth Gurunath Shivakumar
             and Qi Luo and Jia-Hong Huang and M. Maruf and Roger Ren
             and Yile Gu and Rahul Pandey and Ge Liu and Ivan Bulyko},
  journal = {arXiv preprint arXiv:2608.24958},
  year    = {2026}
}