pith. sign in

arxiv: 2606.04474 · v2 · pith:H3TEP2GWnew · submitted 2026-06-03 · 💻 cs.CL · eess.AS

Entity Binding Failures in Speech LLM Reasoning: Diagnosis and Chain-of-Thought Intervention

Ming-Hao Hsu , Xiaohai Tian , Jun Zhang , Zhizheng Wu This is my paper

Pith reviewed 2026-06-28 06:39 UTC · model grok-4.3

classification 💻 cs.CL eess.AS
keywords speech large language modelsentity bindingchain-of-thought promptingreasoning failuresSLLM evaluationlogical reasoningEA-CoT
0
0 comments X

The pith

Speech LLMs collapse on logical reasoning because continuous speech features blur entity-property bindings during implicit processing.

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

The paper evaluates two architecturally different speech LLMs and finds they match or exceed text models on spatial, syntactic, and factual tasks. Performance drops to chance only on logical tasks that require tracking which entity has which property. The authors diagnose the drop as an entity binding failure caused by the continuous nature of speech features. They introduce Entity-Aware Chain-of-Thought, an inference-time prompt that makes the model list entities and bind them to claims before reasoning. This intervention recovers up to 24.4 percentage points even when names are misrecognized, showing the gap is an elicitation problem rather than a missing capability.

Core claim

Speech-to-text accuracy collapses to chance on logical tasks requiring entity tracking while matching text-to-text on other categories. The cause is diagnosed as entity binding failure in which continuous speech features blur precise entity-property associations. Entity-Aware Chain-of-Thought forces explicit enumeration of entities and binding to claims before reasoning, closing the gap with up to 24.4 percentage-point gains even under name misrecognition. Ablations confirm the improvement comes from explicit semantic binding.

What carries the argument

Entity-Aware Chain-of-Thought (EA-CoT), an inference-time intervention that forces SLLMs to enumerate entities and bind them to claims before proceeding with reasoning.

If this is right

  • S2T performance equals T2T on spatial, syntactic, and factual tasks but collapses on entity-tracking logic.
  • EA-CoT recovers up to 24.4 percentage points on those logical tasks.
  • Gains persist even when spoken names are misrecognized.
  • Ablations show the benefit comes specifically from explicit semantic binding steps.
  • The performance gap is reframed as an elicitation failure rather than an absent capability.

Where Pith is reading between the lines

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

  • Training regimes that encourage explicit binding during pre-training could reduce the need for inference-time prompts.
  • The same binding issue may appear in other continuous-input modalities such as video or sensor streams.
  • Extending EA-CoT to multi-turn dialogues could test whether binding failures accumulate over longer contexts.

Load-bearing premise

The selected logical tasks isolate entity tracking as the main or only cause of failure rather than acoustic modeling limits or general speech-to-text degradation.

What would settle it

A controlled test in which the same logical problems are presented to the speech model with perfectly transcribed text versus spoken input; if the gap disappears, the binding-failure diagnosis is supported.

Figures

Figures reproduced from arXiv: 2606.04474 by Jun Zhang, Ming-Hao Hsu, Xiaohai Tian, Zhizheng Wu.

Figure 1
Figure 1. Figure 1: Method overview. The baseline generates a direct answer within 256 tokens, during which entity binding can fail for speech inputs. EA-CoT instructs the model to enumerate entities, record claims, and reason step by step within 1,024 tokens, converting implicit binding into explicit text. 2. Claim recording. Write down each statement linking a per￾son to a property. 3. Step-by-step reasoning. Resolve each c… view at source ↗
Figure 2
Figure 2. Figure 2: Speech recovery and gap reduction on web of lies. Left: speech accuracy rises from chance toward the text base￾line with EA-CoT. Right: the gap shrinks substantially across both models. generates responses directly without a separate reasoning mod￾ule. Both models are evaluated in their released configurations without fine-tuning. Four categories from VoiceBench BBH [5] are used, each containing 250 items … view at source ↗
read the original abstract

Speech Large Language Models (SLLMs) underperform their text counterparts on complex reasoning. We reveal that this gap is not a uniform cognitive deficit. Evaluating two architecturally diverse SLLMs, we show speech-to-text (S2T) matches or exceeds text-to-text (T2T) on spatial, syntactic, and factual tasks. Yet on logical tasks requiring entity tracking, S2T accuracy collapses to chance. We diagnose this as an entity binding failure: continuous speech features blur precise entity-property associations during implicit reasoning. To validate this diagnosis, we introduce Entity-Aware Chain-of-Thought (EA-CoT), a lightweight inference-time intervention forcing SLLMs to enumerate entities and bind them to claims before reasoning. EA-CoT bridges the gap, even when spoken names are misrecognized, yielding up to a 24.4 percentage-point accuracy gain. Ablations confirm the gains stem from explicit semantic binding, reframing the gap as an elicitation failure rather than a missing capability.

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

2 major / 0 minor

Summary. The paper claims that the performance gap between speech-to-text (S2T) and text-to-text (T2T) in SLLMs is not uniform but specific to logical tasks requiring entity tracking, where S2T collapses to chance while matching or exceeding T2T on spatial, syntactic, and factual tasks. This is diagnosed as an 'entity binding failure' caused by continuous speech features blurring entity-property associations during implicit reasoning. The authors introduce Entity-Aware Chain-of-Thought (EA-CoT), an inference-time intervention that forces explicit enumeration and binding of entities to claims, which yields up to 24.4 percentage-point gains even with misrecognized names, reframing the gap as an elicitation failure.

Significance. If the diagnosis is supported after controlling for confounds, the work is significant in isolating a mechanism-specific failure mode in SLLMs rather than a general capability deficit, and in demonstrating a lightweight, inference-only intervention that improves reasoning without retraining. The EA-CoT approach and the 'even when names misrecognized' result, if rigorously validated, offer a practical reframing with potential applicability to other multimodal reasoning settings.

major comments (2)
  1. [Abstract] Abstract: The diagnosis that the accuracy collapse on logical tasks is caused specifically by entity binding failure in speech features (rather than generic S2T degradation) requires evidence that the task categories were matched on surface features including utterance length, number of entities, acoustic variability, and baseline S2T error rates. No such matching or controls are indicated, so the reported pattern (S2T matches T2T on non-logical tasks but collapses on logical ones) remains consistent with length or acoustic confounds instead of the proposed binding mechanism.
  2. [Abstract] Abstract / Methods: Dataset construction, statistical controls, baseline comparisons, and error analysis are not described, which are load-bearing for validating that the collapse is isolated to entity tracking and that EA-CoT gains stem from explicit semantic binding rather than other factors such as general enumeration benefits.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback emphasizing the need for explicit controls and methodological detail. We address each major comment below and will revise the manuscript accordingly to strengthen the claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The diagnosis that the accuracy collapse on logical tasks is caused specifically by entity binding failure in speech features (rather than generic S2T degradation) requires evidence that the task categories were matched on surface features including utterance length, number of entities, acoustic variability, and baseline S2T error rates. No such matching or controls are indicated, so the reported pattern (S2T matches T2T on non-logical tasks but collapses on logical ones) remains consistent with length or acoustic confounds instead of the proposed binding mechanism.

    Authors: We agree that surface-feature matching is essential to isolate entity binding as the mechanism. Although not detailed in the abstract, the dataset was constructed with explicit balancing: utterance lengths averaged 12-15 tokens across categories, entity counts were fixed at 2-4 per example, acoustic variability was controlled via identical TTS synthesis, and baseline S2T WER was measured at 8-12% with no category-level differences. We will add a new 'Dataset Controls' subsection with summary statistics and a table confirming balance, plus error analysis showing that WER does not predict logical-task collapse. This supports the binding diagnosis over confounds. revision: yes

  2. Referee: [Abstract] Abstract / Methods: Dataset construction, statistical controls, baseline comparisons, and error analysis are not described, which are load-bearing for validating that the collapse is isolated to entity tracking and that EA-CoT gains stem from explicit semantic binding rather than other factors such as general enumeration benefits.

    Authors: We acknowledge that the current manuscript provides insufficient detail on these elements. Dataset construction used category-specific templates with controlled variables; statistical comparisons employed paired t-tests and bootstrap resampling; baselines included direct prompting and standard CoT; and error analysis decomposed failures into binding vs. other errors. EA-CoT ablations already include an enumeration-only variant showing smaller gains. We will expand the Methods and Results sections with full procedural descriptions, additional tables, figures, and the requested ablations to demonstrate that gains derive from explicit binding. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical diagnosis and intervention measured directly

full rationale

The paper's central claims rest on direct empirical comparisons of SLLM vs. text performance across task categories, followed by measurement of accuracy gains from an explicitly described inference-time intervention (EA-CoT). No equations, parameter fits, self-citations, or uniqueness theorems are used as load-bearing steps in the provided text. The diagnosis is framed as an interpretation of observed patterns rather than a derivation that reduces to its inputs by construction, and the intervention benefit is reported as a measured outcome on held-out tasks. This is the most common honest non-finding for evaluation-plus-intervention papers.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Review performed on abstract only; no equations, parameters, or explicit background assumptions are stated in the provided text.

axioms (1)
  • domain assumption The selected logical tasks measure entity tracking independently of other speech-processing difficulties.
    The diagnosis that the gap is specifically a binding failure rests on this interpretation of task performance.
invented entities (1)
  • Entity binding failure no independent evidence
    purpose: Explains why speech features cause collapse on logical tasks while other task types remain intact.
    New diagnostic construct introduced to account for the observed pattern; no independent falsifiable prediction supplied in the abstract.

pith-pipeline@v0.9.1-grok · 5713 in / 1344 out tokens · 35792 ms · 2026-06-28T06:39:08.650679+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

43 extracted references · 9 linked inside Pith

  1. [1]

    Ka” to “Cass

    Introduction Speech Large Language Models (SLLMs) can directly process spoken input, enabling natural conversational interactions with- out relying on cascaded text transcription [1, 2, 3, 4]. However, recent benchmarking indicates that these models consistently underperform their text-based counterparts on complex reason- ing tasks [5, 6, 7, 8]. Previous...

  2. [2]

    binding prob- lem

    Related Work The Speech-Text Modality Gap.Recent end-to-end Speech Large Language Models (SLLMs) achieve strong conversa- arXiv:2606.04474v2 [cs.CL] 11 Jun 2026 tional abilities but consistently underperform their text coun- terparts on complex reasoning benchmarks [5, 6, 7, 19]. Prior studies have approached this modality gap primarily from a macroscopic...

    Pith/arXiv arXiv 2026

  3. [3]

    Per-Task Gap Analysis Rather than reporting a single aggregate modality gap, we eval- uate speech and text inputs separately on each BBH category

    Method 3.1. Per-Task Gap Analysis Rather than reporting a single aggregate modality gap, we eval- uate speech and text inputs separately on each BBH category. We use four categories from the V oiceBench BBH split [5, 34]: hyperbaton,navigate,sports understanding, andweb of lies, to- taling 1,000 items (250 per category). Each item is presented as both syn...

  4. [4]

    Setup Two openly released, reproducible speech LLMs spanning dif- ferent architectures are evaluated

    Experiments 4.1. Setup Two openly released, reproducible speech LLMs spanning dif- ferent architectures are evaluated. Qwen2.5-Omni-7B [13] em- ploys a dedicated thinker module that generates internal reason- ing tokens before the visible response. Phi-4-Multimodal [14] Qwen Phi-40 20 40 60 80 100Accuracy (%) S2T Baseline S2T + Best CoT T2T Baseline Qwen ...

  5. [5]

    Ka” to “Cass

    Analysis Semantic Binding vs. Acoustic Recognition.If the speech- text gap simply arose from mishearing names (ASR errors), corrupting names in text inputs should trigger a similar perfor- mance collapse. However, replacing 100% of person names with random strings in T2Tweb of lies(Table 3) reduces Qwen’s accuracy by only 3.6 pp, accounting for barely 11%...

  6. [6]

    Conclusion Entity binding failure is the primary driver of the S2T/T2T modality gap in complex reasoning. We demonstrate that SLLMs maintain robust performance on spatial and factual tasks, but tasks requiring continuous entity tracking collapse to chance-level accuracy on speech inputs. To causally validate this diagnosis, we introduce Entity-Aware Chain...

  7. [7]

    All experimental design, implementation, analysis, and scientific claims are the work of the authors

    Disclosure of AI Tool Use A generative AI assistant was used for editing and polishing the manuscript. All experimental design, implementation, analysis, and scientific claims are the work of the authors

  8. [8]

    Audiopalm: A large language model that can speak and listen,

    P. K. Rubensteinet al., “Audiopalm: A large language model that can speak and listen,”CoRR, vol. abs/2306.12925, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    SALMONN: towards generic hearing abilities for large language models,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “SALMONN: towards generic hearing abilities for large language models,” inICLR. OpenReview.net, 2024

    2024

  10. [10]

    Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,

    D. Zhang, S. Li, X. Zhang, J. Zhan, P. Wang, Y . Zhou, and X. Qiu, “Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,” inEMNLP (Findings), ser. Findings of ACL, vol. EMNLP 2023. Association for Computa- tional Linguistics, 2023, pp. 15 757–15 773

    2023

  11. [11]

    Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,” 2023. [Online]. Available: https://arxiv.org/abs/2311.07919

    Pith/arXiv arXiv 2023

  12. [12]

    V oicebench: Benchmarking llm-based voice assistants,

    Y . Chen, X. Yue, C. Zhang, X. Gao, R. T. Tan, and H. Li, “V oicebench: Benchmarking llm-based voice assistants,”CoRR, vol. abs/2410.17196, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    SAKURA: on the multi- hop reasoning of large audio-language models based on speech and audio information,

    C. Yang, N. Ho, Y . Piao, and H. Lee, “SAKURA: on the multi- hop reasoning of large audio-language models based on speech and audio information,” inINTERSPEECH. ISCA, 2025

    2025

  14. [14]

    V oice evaluation of reason- ing ability: Diagnosing the modality-induced performance gap,

    Y . Lin, Z. Hu, Q. Wang, Y . Liu, H. Zhang, J. Subramanian, N. Vlassis, H. H. Li, and Y . Chen, “V oice evaluation of reason- ing ability: Diagnosing the modality-induced performance gap,” CoRR, vol. abs/2509.26542, 2025

    arXiv 2025

  15. [15]

    Audiobench: A universal benchmark for audio large language models,

    B. Wang, X. Zou, G. Lin, S. Sun, Z. Liu, W. Zhang, Z. Liu, A. Aw, and N. F. Chen, “Audiobench: A universal benchmark for audio large language models,” inNAACL (Long Papers). Association for Computational Linguistics, 2025, pp. 4297–4316

    2025

  16. [16]

    Anatomy of the modality gap: Dissecting the internal states of end-to-end speech llms,

    M.-H. Hsu, X. Zhang, X. Tian, J. Zhang, and Z. Wu, “Anatomy of the modality gap: Dissecting the internal states of end-to-end speech llms,” 2026. [Online]. Available: https://arxiv.org/abs/2603.01502

    arXiv 2026

  17. [17]

    Understanding the modality gap: An empirical study on the speech-text alignment mechanism of large speech language models,

    B. Xiang, S. Zhao, T. Guo, and W. Zou, “Understanding the modality gap: An empirical study on the speech-text alignment mechanism of large speech language models,” inEMNLP. As- sociation for Computational Linguistics, 2025, pp. 5187–5202

    2025

  18. [18]

    Listen, think, and understand,

    Y . Gong, H. Luo, A. H. Liu, L. Karlinsky, and J. R. Glass, “Listen, think, and understand,” inICLR. OpenReview.net, 2024

    2024

  19. [19]

    Prompting large language models with speech recog- nition abilities,

    Y . Fathullah, C. Wu, E. Lakomkin, J. Jia, Y . Shangguan, K. Li, J. Guo, W. Xiong, J. Mahadeokar, O. Kalinli, C. Fuegen, and M. Seltzer, “Prompting large language models with speech recog- nition abilities,” inICASSP. IEEE, 2024, pp. 13 351–13 355

    2024

  20. [20]

    Qwen2.5-omni technical report,

    J. Xu, Z. Guo, J. He, H. Hu, T. He, S. Bai, K. Chen, J. Wang, Y . Fan, K. Dang, B. Zhang, X. Wang, Y . Chu, and J. Lin, “Qwen2.5-omni technical report,”CoRR, vol. abs/2503.20215, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras,

    A. Aboueleninet al., “Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras,” CoRR, vol. abs/2503.01743, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” inICML, ser. Proceedings of Machine Learning Re- search, vol. 202. PMLR, 2023, pp. 28 492–28 518

    2023

  23. [23]

    High fidelity neural audio compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,”Trans. Mach. Learn. Res., vol. 2023, 2023

    2023

  24. [24]

    wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” inNeurIPS, 2020

    2020

  25. [25]

    HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhut- dinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 29, pp. 3451–3460, 2021

    2021

  26. [26]

    Seamlessm4t-massively multilingual & multimodal machine translation,

    S. Communicationet al., “Seamlessm4t-massively multilingual & multimodal machine translation,”CoRR, vol. abs/2308.11596, 2023

    arXiv 2023

  27. [27]

    CORD: bridging the audio- text reasoning gap via weighted on-policy cross-modal distilla- tion,

    J. Hu, D. Zhu, X. Luo, D. Zhang, S. He, Y . Lei, H. Zheng, S. Feng, J. He, Y . Sun, H. Wu, and H. Wang, “CORD: bridging the audio- text reasoning gap via weighted on-policy cross-modal distilla- tion,”CoRR, vol. abs/2601.16547, 2026

    arXiv 2026

  28. [28]

    Clos- ing the modality reasoning gap for speech large language models,

    C. Wang, H. Lu, X. Zhang, S. Liu, Y . Lu, J. Li, and Z. Wu, “Clos- ing the modality reasoning gap for speech large language models,” CoRR, vol. abs/2601.05543, 2026

    Pith/arXiv arXiv 2026

  29. [29]

    Sparks of artificial general intelligence: Early experiments with GPT-4,

    S. Bubeck, V . Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y . T. Lee, Y . Li, S. M. Lundberg, H. Nori, H. Palangi, M. T. Ribeiro, and Y . Zhang, “Sparks of artificial general intelligence: Early experiments with GPT-4,”CoRR, vol. abs/2303.12712, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    How do language models bind entities in context?

    J. Feng and J. Steinhardt, “How do language models bind entities in context?” inICLR. OpenReview.net, 2024

    2024

  31. [31]

    Entity tracking in language models,

    N. Kim and S. Schuster, “Entity tracking in language models,” in ACL (1). Association for Computational Linguistics, 2023, pp. 3835–3855

    2023

  32. [32]

    Lost in the middle: How language mod- els use long contexts,

    N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang, “Lost in the middle: How language mod- els use long contexts,”Trans. Assoc. Comput. Linguistics, vol. 12, pp. 157–173, 2024

    2024

  33. [33]

    Understanding the limits of vision language models through the lens of the binding problem,

    D. Campbell, S. Rane, T. Giallanza, N. D. Sabbata, K. Ghods, A. Joshi, A. Ku, S. Frankland, T. Griffiths, J. D. Cohen, and T. W. Webb, “Understanding the limits of vision language models through the lens of the binding problem,” inNeurIPS, 2024

    2024

  34. [34]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” inNeurIPS, 2022

    2022

  35. [35]

    Ddcot: Duty- distinct chain-of-thought prompting for multimodal reasoning in language models,

    G. Zheng, B. Yang, J. Tang, H. Zhou, and S. Yang, “Ddcot: Duty- distinct chain-of-thought prompting for multimodal reasoning in language models,” inNeurIPS, 2023

    2023

  36. [36]

    Large language models are zero-shot reasoners,

    T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large language models are zero-shot reasoners,” inNeurIPS, 2022

    2022

  37. [37]

    Self-consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. V . Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” inICLR. OpenRe- view.net, 2023

    2023

  38. [38]

    Tree of thoughts: Deliberate problem solving with large language models,

    S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y . Cao, and K. Narasimhan, “Tree of thoughts: Deliberate problem solving with large language models,” inNeurIPS, 2023

    2023

  39. [39]

    Thinking with sound: Audio chain-of-thought enables multimodal reasoning in large audio-language models,

    Z. Xiong, Y . Cai, Z. Li, J. Yuan, and Y . Wang, “Thinking with sound: Audio chain-of-thought enables multimodal reasoning in large audio-language models,”CoRR, vol. abs/2509.21749, 2025

    arXiv 2025

  40. [40]

    Internalizing ASR with im- plicit chain of thought for efficient speech-to-speech conversa- tional LLM,

    R. S. Yuen, T. T. Tse, and J. Zhu, “Internalizing ASR with im- plicit chain of thought for efficient speech-to-speech conversa- tional LLM,”CoRR, vol. abs/2409.17353, 2024

    arXiv 2024

  41. [41]

    Challenging big-bench tasks and whether chain-of-thought can solve them,

    M. Suzgun, N. Scales, N. Sch ¨arli, S. Gehrmann, Y . Tay, H. W. Chung, A. Chowdhery, Q. V . Le, E. H. Chi, D. Zhou, and J. Wei, “Challenging big-bench tasks and whether chain-of-thought can solve them,” inACL (Findings), ser. Findings of ACL, vol. ACL

  42. [42]

    13 003–13 051

    Association for Computational Linguistics, 2023, pp. 13 003–13 051

    2023

  43. [43]

    MMSU: A massive multi-task spoken language understanding and reasoning benchmark,

    D. Wang, J. Wu, J. Li, D. Yang, X. Chen, T. Zhang, and H. Meng, “MMSU: A massive multi-task spoken language understanding and reasoning benchmark,”CoRR, vol. abs/2506.04779, 2025

    Pith/arXiv arXiv 2025