Pith. sign in

REVIEW 3 major objections 5 minor 25 references

Complex audio problems can be solved by an agent that hears, invokes external skills, and reasons over returned evidence; training on complete interaction trajectories yields 15-point gains over fixed-input audio models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An audio agent trained with trajectory-based SFT and multi-turn GRPO improves tool-use and reasoning on a new AI-generated audio agent benchmark, including tasks with unseen tools and workflows.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Solid training recipe and a useful new benchmark, but the OOD generalization claim is undercut by the shared LLM generator and co-tuned evaluation weights. the 3 major comments →

arxiv 2608.01881 v1 pith:BCXIHRX3 submitted 2026-08-03 cs.MM cs.AI

Hear, Invoke, and Understand: A Skill-Calling Multimodal Agent for Large Audio Language Models

Yuwen Wang , Tian-Hao Zhang , Minghao Cai , Yilin Ren , Ziyang Jiang , Xin Wang , Zhichao Wang , Pan Zhou
show 2 more authors
Kun Zhan Xinyuan Qian
This is my paper
classification cs.MM cs.AI
keywords tool-interactive audio reasoningaudio agentskill callingtrajectory-based supervised fine-tuningmulti-turn reinforcement learningout-of-distribution generalizationlarge audio language modelsHIU-Bench
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper argues that many hard audio problems cannot be solved from a fixed input; the model must actively acquire evidence by invoking acoustic skills and tools, then reason over the text or processed audio those tools return. To make this concrete, the authors build a large corpus of interaction trajectories, HIU-Corpus, and train SpeechAgent-R in two stages: supervised fine-tuning on full interaction trajectories, then multi-turn reinforcement learning with tool-aware rewards. They introduce HIU-Bench with in-distribution and out-of-distribution splits, and report that SpeechAgent-R reaches 84.17 in-domain and 70.94 out-of-domain, outperforming the same base model under the same agent harness by roughly 15 points in both settings. A sympathetic reader would take away that tool coordination is a learnable skill in audio agents, and that trajectory-level supervision plus reinforcement learning, not answer-only supervision, is what makes the skill transfer.

Core claim

The paper's central claim is that large audio language models should be cast as interactive agents that hear the input, invoke external skills, and understand the returned evidence, rather than answer directly from a fixed clip. The authors establish this by constructing a corpus of 65,492 verified interaction trajectories spanning 24 tasks, 8 skills, and 9 tools, then train SpeechAgent-R to select skill documents, execute tools, and integrate observations. On the companion benchmark, the trained agent improves the base model under the same harness from 65.01 to 80.05 overall, with the largest gains on tool-related scores, and the authors attribute the improvement to learning skill-tool coor

What carries the argument

The load-bearing object is the Hear-Invoke-Understand loop: at each step the agent maintains a context of query, audio, skill documents, past actions, and tool observations; it either retrieves the description of a selected skill, invokes one of that skill's tools, or emits a final response. The training machinery is trajectory-based supervised fine-tuning, which masks user and tool tokens while weighting agent actions, followed by multi-turn GRPO with a composite reward combining format validity, tool interaction quality, and task accuracy. The HIU-Corpus supplies the verified trajectories and skill documents that make this loop learnable.

Load-bearing premise

The out-of-distribution split in HIU-Bench is generated by the same LLM-driven pipeline used to create the training corpus, so the claim that SpeechAgent-R generalizes to unseen tool configurations rests on this synthetic distribution shift standing in for real-world novelty.

What would settle it

Replace the OOD split with human-authored tasks and workflows that use no training tools, run SpeechAgent-R under the same harness, and compare against the reported 70.94 OOD score; if tool-call accuracy collapses to near chance or the total score falls to the base-harness level, the generalization claim would be falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single audio model can learn to coordinate external acoustic tools for multi-step problems, so fixed-input audio benchmarks understate what large audio language models can do.
  • Answer-only supervised fine-tuning gives only small gains; trajectory-level supervision is what establishes reliable skill selection and tool use.
  • Multi-turn reinforcement learning refines learned tool behavior, reducing incorrect tool-selection failures by roughly a third and tool-result-utilization failures by about a sixth relative to SFT.
  • When ground-truth tool outputs replace actual tool responses, the ID score rises from 84.2 to 91.4, indicating that after tool selection is learned, further progress depends on tool output quality and the model's own reasoning.
  • The same agent harness without training does not reliably help: some models actually score lower with tool access, so training is necessary for effective tool use.

Where Pith is reading between the lines

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

  • Editorial inference: because the OOD split is generated by the same LLM-based pipeline as the training corpus, the reported 70.94 OOD score may partly reflect imitation of the generator's style; a cleaner test would use human-designed OOD workflows.
  • Editorial inference: the hear-invoke-understand recipe could transfer to other modalities or to user-defined skill libraries at inference time, since the skill-retrieval step is what lets the agent handle tools not seen in training.
  • Editorial inference: the paper's own stated limitation in the conclusion — that the benchmark needs more complex and realistic scenarios — reinforces that real deployment would require field trials with novel, human-authored tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper formulates tool-interactive audio reasoning as a Hear-Invoke-Understand (HIU) process, constructs a large corpus (HIU-Corpus) of 65,492 interaction trajectories with 507.6 hours of audio, trains a multimodal agent (SpeechAgent-R) via trajectory-based SFT followed by multi-turn GRPO, and introduces HIU-Bench with ID and OOD splits. The main empirical claim is that SpeechAgent-R outperforms the base model under the same agent harness by 15.40 points on ID and 14.23 points on OOD tasks, and that this demonstrates generalization of skill and tool coordination to diverse unseen task settings.

Significance. If the generalization claim were convincingly established, this would be a useful contribution: the corpus is large, the two-stage training recipe is sensible, and the internal ablations in Table 2 and the oracle analysis in Table 4 provide credible evidence that trajectory-level SFT and RL improve interaction behavior on this benchmark. The paper also clearly separates direct-answer evaluation from agent-harness evaluation and reports per-tool performance, which is informative. However, the central OOD generalization claim is weakened because the OOD benchmark is generated by the same LLM pipeline as the training data and is scored with an author-designed rubric whose weights match the training reward weights. The internal ablations support the training recipe but do not, by themselves, establish transfer to genuinely unseen tool configurations.

major comments (3)
  1. [§5.1, Table 1] The OOD split is 'constructed through the same pipeline' as HIU-Corpus, using the same LLMs (Claude-Opus-4.7, Qwen3.5-122B) and the same task-template grammar. Task-level disjointness and MD5 file checks do not prevent leakage of stylistic patterns, skill-document phrasing, action formatting, or answer conventions. The model's OOD score of 70.94 could therefore reflect fitting the generator's conventions rather than acquiring transferable tool-coordination skills. I would like to see evaluation on independently constructed or human-annotated OOD tasks, or at least a cross-generator robustness check where OOD samples are produced by a different pipeline and scored by independent annotators.
  2. [§4.2 Eq. (4), §5.2, §6.3] The evaluation metric weights in HIU-Bench are 0.05/0.25/0.70, identical to the training reward weights λ_f/λ_t/λ_a, and Section 6.3 reports that these weights were selected by ablations that optimize the overall HIU-Bench score. Thus the headline numbers in Table 1 are partly a product of tuning the evaluation criterion to the training objective. The statement that evaluation metrics are 'independently computed on held-out annotations' addresses annotation leakage but not weight co-tuning. Please report results across a range of evaluation weights, or justify that the 0.05/0.25/0.70 choice is principled rather than selected to maximize the reported score.
  3. [§5.2, Table 4] The oracle-tool analysis that attributes remaining errors to tool-output quality and answer reasoning is performed only on the ID split. Since the central claim concerns OOD generalization, an OOD oracle analysis or an OOD error decomposition is needed. Without it, the paper does not show whether OOD gains come from skill/tool coordination or from the model learning the generator's interaction style and answer format.
minor comments (5)
  1. [§5.1] The text says ID and OOD share 'only two tools' while Figure 4(b) reports 'no overlap' in gold tool chains; please clarify the distinction between tool-level and chain-level overlap to avoid apparent inconsistency.
  2. [Table 1] The table headers are difficult to parse, especially the 'Tool Answer Total Time' spans. Please add explicit subheadings and define what 'Time' measures (wall-clock inference time? tool execution time?).
  3. [§6.3] The reward-weight ablation is described only qualitatively. Please provide a table with all tested weight settings and the resulting ID/OOD/overall scores, and state the number of random seeds used.
  4. [Table 3] The Base model achieves 100.0 on the sound-event-detector metric while SFT drops to 86.8; the text says tool execution performance remains 'comparable,' but this is a notable regression. Please explain whether this reflects different tool versions, metric sensitivity, or evaluation noise.
  5. [§7] The limitation paragraph is only one sentence. Given that the OOD benchmark is synthetic and generated in-house, a more detailed discussion of benchmark limitations and the need for external validation would strengthen the paper.

Circularity Check

2 steps flagged

The headline HIU-Bench score is the same weighted sum as the GRPO training reward, with the reward weights selected on that benchmark, and the OOD split is generated by the same LLM pipeline as the training corpus; the central generalization claim is therefore partially circular.

specific steps
  1. fitted input called prediction [Section 5.2 (Evaluation Protocol) vs. Section 4.2 / Eqs. (3)-(5) and Section 6.3]
    "We define S total on HIU-Bench from three evaluation dimensions: format correctness (Sf), tool interaction (St) and answer quality (Sa), weighted by 0.05/0.25/0.70, with their effects analyzed in Section 6.3."

    The GRPO reward in Eq. (4) is R = 0.05 R_f + 0.25 R_t + 0.70 R_a, and Section 5.2 defines the benchmark total with the identical weights 0.05/0.25/0.70; it also states that S_f and S_a follow the same criteria as R_f and R_a, and that for ID tasks S_t follows the training-time scorer. Thus the reported HIU-Bench total is the very objective that GRPO maximizes. Section 6.3 further says 'Reward weight ablations show that 0.05/0.25/0.70 weighting achieves the best overall score', so the evaluation weights were selected by maximizing this same benchmark. The headline improvement of SpeechAgent-R over baselines is therefore partly the training reward, with weights tuned on the test benchmark, restated as an evaluation result rather than an independent measure of skill/tool coordination.

  2. other [Section 5.1 (Benchmark Construction) vs. Section 4.1 (Corpus Construction)]
    "HIU-Bench comprises an ID split that follows the same task, skill and tool distributions as HIU-Corpus while using independent audio instances, and an OOD split constructed through the same pipeline with new task scenarios and skill compositions."

    HIU-Corpus is the training data, built by a three-stage pipeline in Section 4.1 that uses Claude-Opus-4.7 to check QA pairs and trajectories and Qwen3.5-122B to generate trajectories. The OOD split of HIU-Bench is 'constructed through the same pipeline', so the supposedly out-of-distribution test tasks are generated by the same LLMs, template grammar, skill-document style, and action-format conventions as the training trajectories. Task-level disjointness does not remove this shared generative source. Consequently, the OOD score largely measures how well the model imitates the generator's conventions, not generalization to an independent distribution of tool configurations. The central claim of generalization to 'diverse task settings and adaptive tool interactions' is thus supported by a

full rationale

The two circular steps above affect the main quantitative support for the paper's central claim. First, the evaluation metric on HIU-Bench is definitionally aligned with the GRPO training reward: the same weights (0.05/0.25/0.70) appear in Eq. (4) and in Section 5.2, S_f/S_a are said to follow R_f/R_a, and S_t follows the training-time scorer for ID tasks. Moreover, Section 6.3 states that this weighting was chosen as the one that gives the best overall score on HIU-Bench, meaning the benchmark result is at least partially fitted. Second, the OOD split is generated by the same LLM pipeline as HIU-Corpus, so the claimed out-of-distribution generalization is not measured against a genuinely independent distribution; the model can succeed by matching the generator's style. That said, the paper does contain some independent content: the trajectory-SFT vs. answer-only-SFT ablation and the oracle-tool analysis show behavioral differences under the same harness, and the answer-quality metrics are computed on held-out annotations. No load-bearing self-citation chain or imported uniqueness theorem is present. These independent components prevent the paper from being fully circular, but the headline ID/OOD scores used to demonstrate skill/tool-coordination generalization are substantially compromised by the two reductions above.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim rests on a large, AI-generated corpus and an in-house benchmark. The only fitted numbers are the reward/evaluation weights. No new physical or model entities are introduced beyond the datasets and skill documents.

free parameters (2)
  • Reward weights (lambda_f, lambda_t, lambda_a) = 0.05, 0.25, 0.70
    Chosen after ablation on HIU-Bench (Section 6.3) to achieve the best overall score. These same weights are used as evaluation metric weights (Section 5.2), so the model is scored on the same axis it was tuned on.
  • Evaluation metric weights (Sf, St, Sa) = 0.05, 0.25, 0.70
    Set equal to the training reward weights after reward-weight ablations. The choice of these weights affects all benchmark numbers, including the headline 84.17, 70.94, and 80.05.
axioms (4)
  • domain assumption Skill documents describing tools, interfaces, and calling rules are sufficient for the agent to use both seen and unseen tools correctly at inference.
    Problem formulation (Section 3) and the OOD split (Section 5.1) rely on the agent interpreting natural-language skill documents to call tools it never trained on; no training examples for OOD tools are provided.
  • domain assumption AI-generated trajectories from Qwen3.5 models, verified by Claude-Opus-4.7 and manual correction, are high-quality supervision for learning interaction behavior.
    Section 4.1 'Trajectory Generation and Verification'; the quality of the SFT signal depends on these generated trajectories being correct and representative.
  • domain assumption The agent harness executes tools with deterministic, correct or partially correct outputs during evaluation.
    Section 5.2 and Table 3 assume tool outputs can be compared against gold metrics; in reality tool outputs are imperfect, as shown by the Oracle analysis in Table 4.
  • standard math GRPO is a valid policy-gradient objective for multi-turn RL.
    Section 4.2 Eq. 3 adopts GRPO from prior work; no proof is given in this paper.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Hear, Invoke, and Understand: A Skill-Calling Multimodal Agent for Large Audio Language Models." pith.science (2026). https://pith.science/paper/BCXIHRX3

@misc{pith2026260801881,
  author       = {Pith},
  title        = {Pith review of: Hear, Invoke, and Understand: A Skill-Calling Multimodal Agent for Large Audio Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BCXIHRX3}},
  note         = {Machine review of arXiv:2608.01881}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Complex acoustic problems may require models to perform acoustic operations, interact with external tools and reason over the resulting textual or processed-audio observations rather than answer directly from a fixed audio input. We study such problems as tool-interactive audio reasoning and develop SpeechAgent-R, an audio agent that coordinates its intrinsic multimodal understanding with external skills and tools. To support this capability, we construct HIU-Corpus, comprising 65,492 interaction trajectories and 507.6 hours of audio across 24 tasks, 8 skills and 9 tools. SpeechAgent-R first learns structured interaction behaviors through trajectory-based supervised fine-tuning and then improves its decisions through multi-turn reinforcement learning. We further introduce HIU-Bench to jointly evaluate task performance, interaction quality and generalization to diverse task settings. It contains 1,395 samples across 56 tasks, including in-distribution (ID) and out-of-distribution (OOD) splits with substantial shifts in tool usage and workflow composition. SpeechAgent-R achieves 84.17 on ID tasks and 70.94 on OOD tasks, improving over the base model under the same agent harness by 15.40 and 14.23 points. These results demonstrate that learning skill and tool coordination improves audio agents' ability to handle diverse task settings and adaptive tool interactions.

Figures

Figures reproduced from arXiv: 2608.01881 by Kun Zhan, Minghao Cai, Pan Zhou, Tian-Hao Zhang, Xin Wang, Xinyuan Qian, Yilin Ren, Yuwen Wang, Zhichao Wang, Ziyang Jiang.

Figure 1
Figure 1. Figure 1: Overview of the Hear-Invoke-Understand frame￾work. SpeechAgent-R hears the acoustic scene, invokes re￾quired skills through task reasoning, and understands the problem by integrating the returned evidence. actively acquired and integrated (Sussman 2017). For exam￾ple, recovering the speech of a target speaker from a spatial recording may require localization, speaker identification, speech extraction and s… view at source ↗
Figure 2
Figure 2. Figure 2: Construction pipeline of HIU-Corpus and HIU-Bench, covering task collection, skill construction, trajectory generation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training and interaction workflow of SpeechAgent-R. (a) Two-stage training with trajectory-based SFT and multi-turn [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of tool coverage and workflow compositions between ID and OOD splits in HIU-Bench. (a) Tool-family [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative cases comparing the trajectories generated by the SFT and RL models, illustrating differences in skill [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

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

25 extracted references · 15 linked inside Pith

  1. [4]

    Fu, Y.; Cheng, L.; and Lv, S

    FSD50K: an Open Dataset of Human-Labeled Sound Events.CoRR, abs/2010.00475. Fu, Y.; Cheng, L.; and Lv, S

    Pith/arXiv arXiv 2010

  2. [8]

    Liu, X.; Wang, X.; and Sahidullah, M

    Full-Duplex-Bench- v3: Benchmarking Tool Use for Full-Duplex Voice Agents Under Real-World Disfluency.CoRR, abs/2604.04847. Liu, X.; Wang, X.; and Sahidullah, M

    Pith/arXiv arXiv

  3. [10]

    Maben, L

    MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix.CoRR, abs/2505.13032. Maben, L. M.; Lakshmy, G. G.; and Radhakrishnan, S

    Pith/arXiv arXiv

  4. [11]

    InIEEE Au- tomatic Speech Recognition and Understanding Workshop, ASRU2025,Honolulu,HI,USA,December6-10,2025,1–4

    AURA: Agent for Understanding, Reasoning, and Automated Tool Use in Voice-Driven Tasks. InIEEE Au- tomatic Speech Recognition and Understanding Workshop, ASRU2025,Honolulu,HI,USA,December6-10,2025,1–4. IEEE. Nagrani, A.; Chung, J. S.; and Zisserman, A

    2025

  5. [12]

    In Lacerda, F., ed.,18th Annual Conference of the Inter- national Speech Communication Association, Interspeech 2017, Stockholm, Sweden, August 20-24, 2017, 2616–2620

    Vox- Celeb: A Large-Scale Speaker Identification Dataset. In Lacerda, F., ed.,18th Annual Conference of the Inter- national Speech Communication Association, Interspeech 2017, Stockholm, Sweden, August 20-24, 2017, 2616–2620. ISCA. Pahwa, R.; Beedu, A.; and Priye, P

    2017

  6. [13]

    Rong, Y.; Li, C.; and Yu, D

    Audio2Tool: Speak, Call, Act - A Dataset for Benchmarking Speech Tool Use.CoRR, abs/2604.22821. Rong, Y.; Li, C.; and Yu, D

    Pith/arXiv arXiv

  7. [14]

    Sakshi, S.; Tyagi, U.; and Kumar, S

    AudioGenie-Reasoner: ATraining-FreeMulti-AgentFrameworkforCoarse-to-Fine Audio Deep Reasoning.CoRR, abs/2509.16971. Sakshi, S.; Tyagi, U.; and Kumar, S

    arXiv

  8. [16]

    Qwen3-Omni Technical Report.CoRR, abs/2509.17765. Team, Q

    Pith/arXiv arXiv

  9. [17]

    Tian, F.; Zhang, X

    Qwen3.5-Omni Technical Report.CoRR, abs/2604.15804. Tian, F.; Zhang, X. T.; and Zhang, Y

    Pith/arXiv arXiv

  10. [18]

    Tong, S.; Li, X.; and Wang, Y

    Step-Audio-R1 Technical Report.CoRR, abs/2511.15848. Tong, S.; Li, X.; and Wang, Y

    arXiv

  11. [19]

    Wang, B.; Zou, X.; and Lin, G

    AuTAgent: A Rein- forcement Learning Framework for Tool-Augmented Audio Reasoning.CoRR, abs/2602.13685. Wang, B.; Zou, X.; and Lin, G

    arXiv

  12. [20]

    AudioBench: A Uni- versal Benchmark for Audio Large Language Models. In Proceedings of the 2025 Conference of the Nations of the AmericasChapteroftheAssociationforComputationalLin- guistics:HumanLanguageTechnologies,NAACL2025-Vol- ume 1: Long Papers, Albuquerque, New Mexico, USA, April 29-May4,2025,4297–4316.AssociationforComputational Linguistics. Wang, ...

    2025

  13. [21]

    Wang, Y.; Peng, J.; and Li, H

    MSU-Bench: To- wards Understanding the Conversational Multi-talker Sce- narios.CoRR, abs/2508.08155. Wang, Y.; Peng, J.; and Li, H

    Pith/arXiv arXiv

  14. [22]

    Wijngaard,G.;Formisano,E.;andDumontier,M.2025

    Audio-Mind: An Au- ditableAgenticFrameworkforAudioUnderstanding.CoRR, abs/2605.28480. Wijngaard,G.;Formisano,E.;andDumontier,M.2025. Au- dioToolAgent: An Agentic Framework for Audio-Language Models.CoRR, abs/2510.02995. Wu, D.; Zhang, X.; and Yang, D

    Pith/arXiv arXiv 2025

  15. [23]

    Xie,Z.;Lin,M.;andLiu,Z.2025

    Echo: Towards Ad- vanced Audio Comprehension via Audio-Interleaved Rea- soning.CoRR, abs/2602.11909. Xie,Z.;Lin,M.;andLiu,Z.2025. Audio-Reasoner:Improv- ingReasoningCapabilityinLargeAudioLanguageModels. InProceedings of the 2025 Conference on Empirical Meth- odsinNaturalLanguageProcessing,EMNLP2025,Suzhou, China, November 4-9, 2025, 23829–23851. Associati...

    arXiv 2025

  16. [24]

    CoRR, abs/2606.15141

    EChO-Agent: Evidence Chain Orchestration Agent for Audio Reasoning. CoRR, abs/2606.15141. Zhao, Z.; Jiang, Y.; and Liu, H

    arXiv

  17. [2017]

    Chen, L.; Chen, H.; and Cai, Y

    AISHELL-1: An Open- SourceMandarinSpeechCorpusandASpeechRecognition Baseline.CoRR, abs/1709.05522. Chen, L.; Chen, H.; and Cai, Y

    Pith/arXiv arXiv

  18. [2018]

    Li, L.; Chen, H.; Li, Z.; Hu, Q.; Kang, J.; Li, J.; Xie, L.; and Li,Y.2026

    Spoken SQuAD: A Study of Mitigating the Impact of Speech Recognition Errors on Listening Comprehension.CoRR, abs/1804.00320. Li, L.; Chen, H.; Li, Z.; Hu, Q.; Kang, J.; Li, J.; Xie, L.; and Li,Y.2026. Audio-Cogito:TowardsDeepAudioReasoning in Large Audio Language Models. arXiv:2604.12527. Lin, G.; Chen, C.; and Chen, Z

    Pith/arXiv arXiv 2026

  19. [2020]

    arXiv:2005.11262

    LibriMix: An Open-Source Dataset for Generalizable Speech Separation. arXiv:2005.11262. Dinkel,H.;Li,G.;andLiu,J.2025.MiDashengLM:Efficient Audio Understanding with General Audio Captions.CoRR, abs/2508.03983. Fonseca, E.; Favory, X.; and Pons, J

    Pith/arXiv arXiv 2005

  20. [2021]

    Jain, D.; Shukla, H.; and Rajeev, G

    AISHELL-4: An Open Source Dataset for Speech Enhancement, Separation, Recognition and Speaker Diarization in Conference Sce- nario.CoRR, abs/2104.03603. Jain, D.; Shukla, H.; and Rajeev, G

    Pith/arXiv arXiv

  21. [2022]

    Ma, Z.; Ma, Y.; and Zhu, Y

    ASVspoof 2021: Towards Spoofed and Deepfake Speech Detection in the Wild.CoRR, abs/2210.02437. Ma, Z.; Ma, Y.; and Zhu, Y

    Pith/arXiv arXiv 2021

  22. [2023]

    LibriSQA: Ad- vancing Free-form and Open-ended Spoken Question An- swering with a Novel Dataset and Framework.CoRR, abs/2308.10390

    Pith/arXiv arXiv

  23. [2024]

    Sussman, E

    MMAU: A Mas- siveMulti-TaskAudioUnderstandingandReasoningBench- mark.CoRR, abs/2410.19168. Sussman, E. S

    Pith/arXiv arXiv

  24. [2025]

    KimiTeam;Ding,D.;andJu,Z.2025

    VoiceAgent- Bench:AreVoiceAssistantsreadyforagentictasks?CoRR, abs/2510.07978. KimiTeam;Ding,D.;andJu,Z.2025. Kimi-AudioTechnical Report.CoRR, abs/2504.18425. Kumar,S.;Sedlácek,S.;andLokegaonkar,V.2026.MMAU- Pro: A Challenging and Comprehensive Benchmark for Holistic Evaluation of Audio General Intelligence. InForti- ethAAAIConferenceonArtificialIntellige...

    arXiv 2025

  25. [2026]

    CoRR, abs/2602.10439

    AudioRouter: Data EfficientAudioUnderstandingviaRLbasedDualReasoning. CoRR, abs/2602.10439. Chu,Y.;Xu,J.;andYang,Q.2024. Qwen2-AudioTechnical Report.CoRR, abs/2407.10759. Cosentino, J.; Pariente, M.; Cornell, S.; Deleforge, A.; and Vincent, E

    arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.