pith. sign in

arxiv: 2606.22473 · v1 · pith:SIZMHBCAnew · submitted 2026-06-21 · 💻 cs.CL · cs.LG· cs.SD· eess.AS

Interleaved Speech Language Models Latently Work In Text

Talia Sternberg , Gallil Maimon , Yossi Adi This is my paper

Pith reviewed 2026-06-26 10:39 UTC · model grok-4.3

classification 💻 cs.CL cs.LGcs.SDeess.AS
keywords speech language modelsinterleaved traininglogit lensimplicit transcriptionspeech-text interactionmultimodal language models
0
0 comments X

The pith

Interleaved speech language models decode spoken words as text tokens in their intermediate layers.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines how speech language models trained on interleaved speech and text sequences process inputs internally. It finds that these models perform an implicit transcription step where the text token matching a spoken word becomes readable from middle-layer activations. This occurs without any explicit speech-to-text training and reaches top-candidate status for as much as 77 percent of examined cases. After transcription the model shifts to next-token prediction in text space and only later returns to speech output. The analysis traces this pattern across model families and links it to the benefits of interleaving data and text-LM initialization.

Core claim

These models go through an implicit transcription phase in which the text token of the spoken word becomes decodable in intermediate layers, despite not being trained for speech recognition. The transcription of the word appears as one of the top candidate words for as much as 77% of the data. Following this stage, the models proceed to predict the next word in the text space before transforming back to the speech domain.

What carries the argument

Logit lens applied to intermediate layers, which extracts the text-token predictions that become decodable during the transcription phase.

If this is right

  • Interleaving speech and text data during training elicits the internal transcription behavior.
  • Initializing from a text language model strengthens the emergence of the transcription phase.
  • The presence and strength of the transcription phase correlates with the model's spoken-knowledge performance.
  • After the transcription stage the model completes next-word prediction entirely within the text token space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Pure speech-only models without interleaving may lack access to this latent text pathway and therefore underperform on knowledge-intensive spoken tasks.
  • Training objectives could be designed to explicitly strengthen or control the duration of the transcription window.
  • Similar latent translation phases might appear when other modalities are interleaved with text.

Load-bearing premise

The logit lens applied to intermediate layers accurately reflects the model's actual internal computation and decision process rather than an artifact of the probing method itself.

What would settle it

A controlled experiment in which middle-layer activations are altered to suppress the observed text-token candidates and the model is then tested on whether its final speech output still matches the original behavior on the same inputs.

Figures

Figures reproduced from arXiv: 2606.22473 by Gallil Maimon, Talia Sternberg, Yossi Adi.

Figure 1
Figure 1. Figure 1: Implicit transcription emerges without speech-recognition supervision. Logit-lens analysis of intermediate states for the spoken prompt “The capital of the United Kingdom is...”. Cells show textual tokens probability, from light yellow for zero to dark blue for high probability. The labels show the most probable relevant textual token at each position; notably, the model predicts “London” although it was n… view at source ↗
Figure 2
Figure 2. Figure 2: Speech LMs operate in text. Modality distri￾bution of inner state Logit Lens. (a) Sum of probabilities over the all speech tokens and text tokens respectively. (b) Same but only considering top 200 tokens. I-1/3, I-2/3, and I-5/6, corresponding to speech￾only, speech+text, and speech+text+interleaved training with increasing fractions of interleaved tokens. We use the prefixes P and R to indicate pretraine… view at source ↗
Figure 3
Figure 3. Figure 3: Implicit transcription and textual continuation emerge in speech hidden states. We apply the logit lens to speech-token hidden states and report Recall@k up to a given layer, for the current transcription word, the next word, and the final answer. Although the models are not explicitly trained for transcription, current-word transcription emerges reliably in intermediate layers across models, while next-wo… view at source ↗
Figure 4
Figure 4. Figure 4: Implicit transcription ability is positively correlated with factual knowledge retrieval. Each point represents a model. The x-axis reports the percentage of words for which the correct current-word transcription (left) or next-word transcription (right) appears in the top-10 logit-lens predictions at any aligned speech-token position and layer. The y-axis shows the binary accuracy on our commonsense factu… view at source ↗
Figure 5
Figure 5. Figure 5: Log Likelihood based evaluations for all models we used [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Implicit transcription and textual continuation emerge in speech hidden states for text pre-trained with interleaving. We apply the logit lens to speech-token hidden states and report Recall@k up to a given layer, for the current transcription word, the next word, and the final answer. Although the models are not explicitly trained for transcription, current-word transcription emerges reliably in intermedi… view at source ↗
Figure 7
Figure 7. Figure 7: Logit lens of intermediate states for the spoken input "lime", using the Llama-3.2 PI-1/3 (official) model. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Logit lens of intermediate states for the spoken input "white", using the Llama-3.2 PI-1/3 (official) [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Logit lens of intermediate states for the spoken input "Pakistan", using the Llama-3.2 PI-1/3 (official) [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Logit lens of intermediate states for the spoken input "teacher", using the Llama-3.2 PI-1/3 (official) [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Logit Lens of inner states of the spoken input: "The capital of United Kingdom is...", using Llama-3.2 [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Logit lens of inner states of the spoken input: "Paris is the capital of ...", using Llama-3.2 PI-1/3 (official). [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Logit Lens of inner states of the spoken input: "Paris is the capital of ...", using Llama-3.2 PI-1/3 (ours). [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Logit Lens of inner states of the spoken input: "Paris is the capital of ...", using Llama-3.2 PI-2/3. [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Logit Lens of inner states of the spoken input: "Paris is the capital of...", using Qwen2.5-3B PI-1/3 60k [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Logit Lens of inner states of the spoken input: "Paris is the capital of...", using Qwen2.5-1.5B PI-1/3 42K [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Logit Lens of inner states of the spoken input: "The capital of France is...", using Llama-3.2 PI-1/3 [PITH_FULL_IMAGE:figures/full_fig_p022_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Logit Lens of inner states of the spoken input: "The capital of France is...", using Llama-3.2 PI-2/3 [PITH_FULL_IMAGE:figures/full_fig_p022_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Logit Lens of inner states of the spoken input: "The capital of France is...", using Qwen2.5-3B PI-1/3 [PITH_FULL_IMAGE:figures/full_fig_p023_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Logit Lens of inner states of the spoken input: "Paris is the capital of ...", using Llama3.2-3B RST [PITH_FULL_IMAGE:figures/full_fig_p023_20.png] view at source ↗
read the original abstract

Speech language models (SLMs) have been extensively studied, with the common paradigm incorporating text data and pre-trained text LMs. A leading approach is speech-text interleaving in which models are trained over sequences containing both speech and text tokens, aiming to boost even speech-only capabilities. Yet the way these two modalities interact in the model latent space remains unclear. In this work, we analyze interleaved speech-text LMs from different model families and sizes through the scope of the logit lens to provide such insight. We reveal that these models go through an implicit transcription phase in which the text token of the spoken word becomes decodable in intermediate layers, despite not being trained for speech recognition. The transcription of the word appears as one of the top candidate words for as much as 77\% of the data. Following this stage, the models proceed to predict the next word in the text space before transforming back to the speech domain. We finally analyze the role of interleaving data, and initializing from text LMs in eliciting this behavior, as well as seeing how this correlates with spoken knowledge abilities. Our analysis sheds light on the internal mechanisms underlying the relationship between speech and text modalities and could shape SLM optimization.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper analyzes interleaved speech-text language models from multiple families and sizes using the logit lens. It claims these models undergo an implicit transcription phase in intermediate layers, where the text token of a spoken word becomes decodable and ranks among the top candidates for up to 77% of examples despite no explicit speech recognition training; models then predict the next token in text space before returning to the speech domain. The work further examines how interleaving data and text-LM initialization elicit this behavior and its correlation with spoken knowledge abilities.

Significance. If the logit-lens observations hold and reflect genuine internal computation, the result would provide a concrete mechanistic account of modality interaction in interleaved SLMs, highlighting an emergent transcription-like stage that could guide training recipes. The comparative analysis across model scales and the correlation with spoken capabilities are useful observational contributions. The work employs an established probing technique rather than introducing new machinery, so its primary value lies in the reported patterns rather than in novel methodology.

major comments (3)
  1. [Abstract / Methods] Abstract and Methods: the headline quantitative result (77% top-candidate rate) is presented without dataset size, number of models or examples evaluated, statistical controls, or confirmation that the metric was pre-specified rather than selected post-hoc. These details are required to evaluate whether the central claim is supported.
  2. [Results (logit lens)] Results section on logit-lens analysis: the claim that the decoded text token reflects an implicit transcription phase that the model actually uses rests on the unembedding matrix applied to intermediate residual streams, yet no causal intervention (activation patching, head ablation, or counterfactual editing) is reported to test whether that token influences final predictions.
  3. [Discussion] Discussion of modality mixing: in speech-text models the residual stream interleaves modalities, so the assumption that the text-only unembedding matrix surfaces the model's actual internal computation (rather than a spurious correlation) requires explicit justification or controls; the current observational correlations with interleaving and text initialization do not address this.
minor comments (2)
  1. [Methods] Clarify the precise definition of 'top candidate words' (rank threshold, vocabulary size, handling of subword tokens) and report the exact evaluation protocol in a dedicated methods subsection.
  2. [Figures] Figure captions and axis labels should explicitly state the number of examples and models underlying each plotted percentage.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important aspects of clarity and evidential strength. We address each major comment below and will revise the manuscript to incorporate additional details, caveats, and justifications where appropriate. These changes will improve the paper without altering its core observational findings.

read point-by-point responses

  1. Referee: [Abstract / Methods] Abstract and Methods: the headline quantitative result (77% top-candidate rate) is presented without dataset size, number of models or examples evaluated, statistical controls, or confirmation that the metric was pre-specified rather than selected post-hoc. These details are required to evaluate whether the central claim is supported.

    Authors: We agree that these experimental details should be explicitly stated. The 77% figure represents the maximum top-candidate rate observed across the evaluated conditions. In the revised version we will report the dataset size (number of spoken utterances), the number of models and families analyzed, the total examples per condition, and any statistical measures used. We will also clarify that the top-candidate metric follows conventions from prior logit-lens studies rather than being chosen post-hoc. These additions will be made to both the abstract and methods sections. revision: yes

  2. Referee: [Results (logit lens)] Results section on logit-lens analysis: the claim that the decoded text token reflects an implicit transcription phase that the model actually uses rests on the unembedding matrix applied to intermediate residual streams, yet no causal intervention (activation patching, head ablation, or counterfactual editing) is reported to test whether that token influences final predictions.

    Authors: The analysis is observational and relies on the logit lens, an established but correlational technique. We do not present causal evidence that the surfaced text token is directly used in downstream computation. To address this, we will revise the results and discussion to use more precise language (e.g., “suggests a latent transcription-like stage” instead of implying direct usage) and will add an explicit limitations paragraph noting the absence of causal interventions such as patching. This revision clarifies the strength of the evidence without overstating it. revision: yes

  3. Referee: [Discussion] Discussion of modality mixing: in speech-text models the residual stream interleaves modalities, so the assumption that the text-only unembedding matrix surfaces the model's actual internal computation (rather than a spurious correlation) requires explicit justification or controls; the current observational correlations with interleaving and text initialization do not address this.

    Authors: We acknowledge that applying a text-only unembedding to a mixed-modality residual stream requires careful interpretation. In the revision we will expand the discussion to justify the approach by highlighting that the text-token emergence is strongly modulated by text-LM initialization and interleaving data, and that it correlates with spoken-knowledge performance. We will also cite prior logit-lens applications to multimodal models and explicitly note the possibility of spurious correlations, recommending causal follow-up work. These additions provide the requested justification and controls discussion. revision: yes

Circularity Check

0 steps flagged

No circularity: observational logit-lens analysis on existing models

full rationale

The paper applies the established logit lens technique to probe intermediate layers of pre-trained interleaved SLMs. All claims (implicit transcription phase, top-candidate rates up to 77%, correlation with interleaving) are direct empirical observations from applying the final unembedding matrix to residual streams; no parameter is fitted to a subset and then renamed as a prediction, no self-citation chain justifies a uniqueness theorem, and no derivation reduces to its own inputs by construction. The work is self-contained against external benchmarks because the observations are falsifiable on the same models without requiring the paper's own fitted values.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an empirical probing study; the abstract introduces no new mathematical objects, free parameters, or postulated entities. The central claim rests on the validity of the logit lens as a faithful probe and on the representativeness of the tested models and data.

pith-pipeline@v0.9.1-grok · 5752 in / 1091 out tokens · 19692 ms · 2026-06-26T10:39:16.913879+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 21 canonical work pages · 7 internal anchors

  1. [1]

    On The Landscape of Spoken Language Models: A Comprehensive Survey

    On the landscape of spoken lan- guage models: A comprehensive survey.Preprint, arXiv:2504.08528. Jayadev Billa

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Jingyi Chen, Zhimeng Guo, Jiyun Chun, Pichao Wang, Andrew Perrault, and Micha Elsner

    The cascade equivalence hypoth- esis: When do speech llms behave like asr →llm pipelines?Preprint, arXiv:2602.17598. Jingyi Chen, Zhimeng Guo, Jiyun Chun, Pichao Wang, Andrew Perrault, and Micha Elsner

    work page arXiv

  3. [3]

    acoustic emotion cues reliance.Preprint, arXiv:2510.10444

    Do au- dio llms really listen, or just transcribe? measuring lexical vs. acoustic emotion cues reliance.Preprint, arXiv:2510.10444. Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T Tan, and Haizhou Li

    work page arXiv

  4. [4]

    VoiceBench: Benchmarking LLM-Based Voice Assistants

    V oicebench: Benchmarking llm-based voice assistants.arXiv preprint arXiv:2410.17196. Santiago Cuervo and Ricard Marxer

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    InProceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 351–361

    Scaling properties of speech language models. InProceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 351–361. Guy Dar, Mor Geva, Ankit Gupta, and Jonathan Berant

    2024

  6. [6]

    Moshi: a speech-text foundation model for real-time dialogue

    Moshi: a speech- text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, and 1 others

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    The Llama 3 Herd of Models

    The llama 3 herd of models.Preprint, arXiv:2407.21783. Daniel Galvez, Greg Diamos, Juan Ciro, Juan Felipe Cerón, Keith Achorn, Anjali Gopi, David Kanter, Maximilian Lam, Mark Mazumder, and Vijay Janapa Reddi

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Mor Geva, Avi Caciularu, Kevin Wang, and Yoav Gold- berg

    The people’s speech: A large-scale diverse english speech recognition dataset for com- mercial usage.Preprint, arXiv:2111.09344. Mor Geva, Avi Caciularu, Kevin Wang, and Yoav Gold- berg

    work page arXiv

  9. [9]

    InProceedings of the 2022 conference on empirical methods in natural language processing, pages 30–45

    Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. InProceedings of the 2022 conference on empirical methods in natural language processing, pages 30–45. Neta Glazer, Yael Segal-Feldman, Hilit Segev, Aviv Shamsian, Asaf Buchnick, Gill Hetz, Ethan Fe- taya, Joseph Keshet, and Aviv Navon

    2022

  10. [10]

    Be- yond Transcription: Mechanistic Interpretability in ASR,

    Be- yond transcription: Mechanistic interpretability in asr.Preprint, arXiv:2508.15882. Danny Halawi, Jean-Stanislas Denain, and Jacob Stein- hardt

    work page arXiv

  11. [11]

    InInternational Conference on Learning Representa- tions, volume 2024, pages 42749–42787

    Overthinking the truth: Understanding how language models process false demonstrations. InInternational Conference on Learning Representa- tions, volume 2024, pages 42749–42787. Michael Hassid, Tal Remez, Tu Anh Nguyen, Itai Gat, Alexis Conneau, Felix Kreuk, Jade Copet, Alexan- dre Defossez, Gabriel Synnaeve, Emmanuel Dupoux, and 1 others

    2024

  12. [12]

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdel- rahman Mohamed

    Anatomy of the modality gap: Dissecting the internal states of end- to-end speech llms.Preprint, arXiv:2603.01502. Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdel- rahman Mohamed

    work page arXiv

  13. [13]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units.Preprint, arXiv:2106.07447. J. Kahn, M. Riviere, W. Zheng, E. Kharitonov, Q. Xu, P.E. Mazare, J. Karadayi, V . Liptchinsky, R. Col- lobert, C. Fuegen, T. Likhomanenko, G. Synnaeve, A. Joulin, A. Mohamed, and E. Dupoux

    work page arXiv

  14. [14]

    InICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), page 7669–7673

    Libri- light: A benchmark for asr with limited or no super- vision. InICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), page 7669–7673. IEEE. Arne Köhn, Florian Stegen, and Timo Baumann

    2020

  15. [15]

    InProceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016), Paris, France

    Mining the spoken wikipedia for speech data and beyond. InProceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016), Paris, France. European Language Re- sources Association (ELRA). Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, A...

    2016

  16. [16]

    10 Omar Mahmoud, Buddhika Laknath Semage, Thom- men George Karimpanal, and Santu Rana

    Looking beyond the top-1: Transformers determine top tokens in order.Preprint, arXiv:2410.20210. 10 Omar Mahmoud, Buddhika Laknath Semage, Thom- men George Karimpanal, and Santu Rana

    work page arXiv

  17. [17]

    Gallil Maimon, Avishai Elmakies, and Yossi Adi

    Improving multilingual language models by align- ing representations through steering.Preprint, arXiv:2505.12584. Gallil Maimon, Avishai Elmakies, and Yossi Adi. 2025a. Slamming: Training a speech language model on one GPU in a day. InFindings of the Association for Computational Linguistics: ACL 2025, pages 12201– 12216, Vienna, Austria. Association for ...

    work page arXiv 2025

  18. [18]

    Scaling open discrete audio founda- tion models with interleaved semantic, acoustic, and text tokens.arXiv preprint arXiv:2602.16687. Pooneh Mousavi, Gallil Maimon, Adel Moumen, Dar- ius Petermann, Jiatong Shi, Haibin Wu, Haici Yang, Anastasia Kuznetsova, Artem Ploujnikov, Ricard Marxer, Bhuvana Ramabhadran, Benjamin Elizalde, Loren Lugosch, Jinyu Li, Cem...

    work page arXiv

  19. [19]

    InInternational Conference on Learn- ing Representations, volume 2024, pages 51883– 51898

    Spoken question an- swering and speech continuation using spectrogram- powered llm. InInternational Conference on Learn- ing Representations, volume 2024, pages 51883– 51898. Clement Neo, Luke Ong, Philip Torr, Mor Geva, David Krueger, and Fazl Barez

    2024

  20. [20]

    InInternational Conference on Learning Representations, volume 2025, pages 57172–57189

    Towards interpret- ing visual information processing in vision-language models. InInternational Conference on Learning Representations, volume 2025, pages 57172–57189. Tu Anh Nguyen, Benjamin Muller, Bokai Yu, Marta R. Costa-jussa, Maha Elbayad, Sravya Pop- uri, Christophe Ropers, Paul-Ambroise Duquenne, Robin Algayres, Ruslan Mavlyutov, Itai Gat, Mary Wi...

    2025

  21. [21]

    Qwen2.5 Technical Report

    Qwen2.5 technical report.Preprint, arXiv:2412.15115. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock- man, Christine McLeavey, and Ilya Sutskever

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Scaling Language Models: Methods, Analysis & Insights from Training Gopher

    Scaling language models: Methods, analysis & insights from training gopher.arXiv preprint arXiv:2112.11446. Sakshi Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ramaneswaran Selvakumar, Oriol Nieto, Ra- mani Duraiswami, Sreyan Ghosh, and Dinesh Manocha

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    InInter- national Conference on Learning Representations, volume 2025, pages 84929–84964

    Mmau: A massive multi-task audio understanding and reasoning benchmark. InInter- national Conference on Learning Representations, volume 2025, pages 84929–84964. Changhan Wang, Morgane Rivière, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux

    2025

  24. [24]

    Maurice Weber, Daniel Y

    V oxpop- uli: A large-scale multilingual speech corpus for rep- resentation learning, semi-supervised learning and interpretation.Preprint, arXiv:2101.00390. Maurice Weber, Daniel Y . Fu, Quentin Anthony, Yonatan Oren, Shane Adams, Anton Alexandrov, Xiaozhong Lyu, Huu Nguyen, Xiaozhe Yao, Vir- ginia Adams, Ben Athiwaratkun, Rahul Chalamala, Kezhen Chen, M...

    work page arXiv

  25. [25]

    Bajian Xiang, Shuaijiang Zhao, Tingwei Guo, and Wei Zou

    The semantic hub hypothesis: Language models share semantic repre- sentations across languages and modalities.Preprint, arXiv:2411.04986. Bajian Xiang, Shuaijiang Zhao, Tingwei Guo, and Wei Zou

    work page arXiv

  26. [26]

    Zhifei Xie and Changqiao Wu

    Understanding the modality gap: An empirical study on the speech-text alignment mech- anism of large speech language models.Preprint, arXiv:2510.12116. Zhifei Xie and Changqiao Wu

    work page arXiv

  27. [27]

    Sohee Yang, Elena Gribovskaya, Nora Kassner, Mor Geva, and Sebastian Riedel

    Mini-omni: Lan- guage models can hear, talk while thinking in stream- ing.arXiv preprint arXiv:2408.16725. Sohee Yang, Elena Gribovskaya, Nora Kassner, Mor Geva, and Sebastian Riedel

    work page arXiv

  28. [28]

    StressTest: Can YOUR Speech LM Handle the Stress?

    Stresstest: Can your speech lm handle the stress? Preprint, arXiv:2505.22765. Aohan Zeng, Zhengxiao Du, Mingdao Liu, Lei Zhang, Yuxiao Dong, Jie Tang, and 1 others

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    InInternational Conference on Learning Rep- resentations, volume 2025, pages 49396–49419

    Scaling speech-text pre-training with synthetic interleaved data. InInternational Conference on Learning Rep- resentations, volume 2025, pages 49396–49419. Yiran Zhao, Wenxuan Zhang, Guizhen Chen, Kenji Kawaguchi, and Lidong Bing

    2025

  30. [30]

    12 A Appendix A.1 Common Sense Dataset Table 2 reports statistics for the different subsets of our evaluation dataset, along with one representa- tive example from each subset

    How do large language models handle multilingualism?Preprint, arXiv:2402.18815. 12 A Appendix A.1 Common Sense Dataset Table 2 reports statistics for the different subsets of our evaluation dataset, along with one representa- tive example from each subset. A.2 Experimental Setup Interleaving.We follow Zeng et al. (2025), sam- pling speech-segment lengths ...

    work page arXiv 2025