Pith. sign in

REVIEW 2 major objections 4 minor 35 references

A filtered 3,000-question audio benchmark shows the best model answers only 58.3% correctly, meaning the task measures genuine audio understanding, not textual shortcuts.

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 ↗

On a new 3000-item audio-only benchmark stripped of text-solvable questions, the best AI system scores 58.33%, random guessing scores 25.46%, and all 36 competing systems fail the same 233 items.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Honest challenge summary whose benchmark's audio-dependence guarantee is weaker than advertised—useful new numbers, but filter completeness needs validation. the 2 major comments →

arxiv 2607.18718 v1 pith:D4CQMV56 submitted 2026-07-21 eess.AS

Summary of DCASE 2026 Task 5: Audio-Dependent Question Answering

Haolin He , Renhe Sun , Zheqi Dai , Xingjian Du , Chunyat Wu , Zining Liang , Zhengxi Liu , Jiahe Lei
show 11 more authors
Runbang Wang Jiayi Zhou Mingru Yang Xiquan Li Yun Chen Xie Chen Zhiyao Duan Weiqiang Wang Mark D. Plumbley Jian Liu Qiuqiang Kong
This is my paper
classification eess.AS
keywords audio-dependent question answeringbenchmark filteringlarge audio-language modelstextual priorsaudio understandingmultiple-choice evaluationDCASE challengeaudio reasoning
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 introduces ADQA-Bench, a 3,000-item multiple-choice audio benchmark whose items are explicitly filtered to remove questions answerable from text alone. The authors' claim is that the filtering pipeline (silent-audio probing, perplexity checks, LLM commonsense review, human verification) leaves only audio-dependent questions, so the measured accuracy of current large audio-language models reflects true audio understanding. The top system reaches 58.33% accuracy, well above the 25.46% random-guess baseline, and all 36 submissions miss the same 233 items, suggesting shared capability gaps. The paper also documents a consistent drop of about 12 percentage points from development to evaluation sets, indicating the hidden evaluation split is harder by design.

Core claim

The central discovery is that a carefully filtered audio question-answering benchmark can separate audio-driven performance from linguistic priors. The Audio-Dependency Filtering (ADF) cascade removes items that silent-audio LALMs, perplexity-based text preferences, or a text-only LLM can solve; the surviving 3,000 items are claimed to be audio-dependent. On this benchmark, the best system—an ensemble of MOSS-Audio-8B-Thinking and Qwen3-Omni-30B—achieves 58.33% accuracy, while the sub-10B lightweight winner scores 57.30%, only 1.03 percentage points behind. Every submission misses a common set of 233 items (7.8% of the evaluation set), and an oracle that picks the best answer across all syst

What carries the argument

The key mechanism is the Audio-Dependency Filtering (ADF) pipeline, a four-stage cascade that keeps only questions whose correct answer depends on the audio signal. First, Hard-ADF replaces audio with silence and discards items if more than one blind LALM selects the correct option. Second, Soft-ADF drops items where the correct option has the lowest language-model perplexity under silent audio. Third, a text-only LLM commonsense filter prunes questions answerable from world knowledge. Fourth, a two-pass human review verifies that the labeled answer is correct given the audio and that distractors are plausible but acoustically wrong. This cascade is the load-bearing component that defines wh

Load-bearing premise

The benchmark's validity rests on the assumption that the ADF cascade catches every text-solvable item—that is, any item that defeats all three automated filters and passes human review is genuinely audio-dependent, so if a text-solvable item slips through all filters, it is incorrectly counted as audio-dependent.

What would settle it

Take a random sample of ADQA-Bench evaluation items and run a strong text-only LLM (with no audio input) on the question and options, measuring its accuracy against the 25.46% random-guess baseline. If the text-only model scores significantly above that baseline on the same items, it demonstrates that some benchmark items are still solvable from text alone, falsifying the claim that all 3,000 items are audio-dependent. Similarly, removing option-length and distractor artifacts should bring the random-guess baseline to exactly 25.00%; if it does not, textual priors remain in the benchmark.

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

If this is right

  • If the ADF filtering is sound, ADQA-Bench provides a reliable measure of audio-grounded reasoning, and the 58.33% top accuracy indicates substantial headroom for improving large audio-language models.
  • The consistent 11.91 percentage-point average drop from development to evaluation confirms that the hidden split is harder, making it a more reliable test of generalization than the development set.
  • The 233 universally-failed items suggest a shared capability gap among current backbones—likely stemming from common audio encoders or pre-training strategies—rather than random errors.
  • The small 1.03 percentage-point gap between an 8-billion-parameter model and a 96-billion-parameter ensemble implies that architectural choices, data curation, and inference-time ensembling can be as important as raw scale.
  • The finding that CoT-targeted fine-tuning helps on some backbones but not others, and that many reasoning-time interventions fail, indicates that audio-answer alignment is not yet a solved problem.

Where Pith is reading between the lines

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

  • A direct test of the filter's completeness would be to run a strong text-only language model on ADQA-Bench items; if it scores significantly above the 25.46% random-guess baseline, then some text-solvable items still pass the ADF cascade, and the benchmark's audio-dependence claim would need qualification.
  • The residual 'option-length and distractor artifacts' that push random-guess accuracy to 25.46% are in-paper evidence that the filter does not fully eliminate text-only signals; a stronger filter could target these artifacts directly.
  • Extending ADQA-Bench to open-ended generation, as the paper suggests, would require an audio-blind judge; if such a judge could be built, it would likely reveal even larger performance gaps than multiple-choice accuracy shows.
  • The shared failure on 233 items may point toward a specific audio property (e.g., fine-grained temporal or event-level cues) that current encoders under-represent; probing those items with targeted acoustic augmentations could isolate the missing capability.
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

2 major / 4 minor

Summary. This paper is the official summary of DCASE 2026 Task 5 on Audio-Dependent Question Answering. It introduces ADQA-Bench, a 3,000-item evaluation set produced by an Audio-Dependency Filtering (ADF) cascade (silent-audio probing, per-option perplexity, an LLM commonsense check, and human verification), and reports results from 14 teams and 36 submissions across two parameter tracks. The main findings are: a Chung-Ang University ensemble (Lim_CAU_4) reaches 58.33% top accuracy; the lightweight champion scores 57.30%; the average dev-to-eval gap across 30 comparable submissions is -11.91 pp; MOSS-Audio-8B-Thinking is the most common backbone; and 233 evaluation items are missed by all systems. The paper also surveys methodological trends (LoRA, SFT, GRPO/GDPO, prompt/voting ensembles) and explicitly flags several limitations, including dev-only baselines, sub-percentage-point tie levels, confounded paradigm/backbone comparisons, and the absence of an ablation for the winning system's acoustic tagger.

Significance. If the ADF filtering is sound, the paper is a valuable community resource: a public benchmark whose scoring is designed to penalize textual-prior guessing, plus a first-year snapshot of system capabilities. The paper deserves credit for repeatedly flagging its own inferential limits (dev-only baselines, sub-%-point ties, confounded comparisons, missing ablation) and for making the challenge results transparent. The 233 universally-failed items and the dev-to-eval gap are concrete, falsifiable measurements that should inform future benchmark design. The main caveat — the filter-completeness assumption — is discussed below; if not addressed, the benchmark's headline interpretation as measuring 'audio understanding' is not established.

major comments (2)
  1. [§2.1–§2.2] §2.1 defines an audio-dependent item as one whose correct answer 'cannot be recovered from the text alone.' The ADF cascade (§2.2) does not enforce this: Hard-ADF drops an item only if *more than one* silent-audio LALM picks it, so an item that exactly one blind model can solve from text is retained. Soft-ADF drops only when the correct option has lowest perplexity; a second-place textual cue survives. The LLM filter prunes only what one text-only LLM answers correctly; human review checks answer correctness and distractor plausibility, not text-solvability. The 25.46% random-guess floor in §2.4 (option-length/distractor artifacts) is in-paper evidence that non-audio signals remain. Since the abstract and Section 5 interpret the top accuracy, dev-to-eval drop, and 233 universal failures as measuring audio understanding, the central claim rests on an unvalidated filter-completeness assump
  2. [§5.2] §5.2 reports a mean dev-to-eval drop of -11.91 pp and attributes it to the evaluation set being 'designed to be harder than the development split.' The two splits are different item sets, so the gap conflates item-set difficulty with development-set overfitting and sampling variability. The paper acknowledges overfitting for some teams and n=30, but the causal attribution to a harder eval set is not supported by a controlled comparison. Provide construction details or a matched-difficulty analysis (e.g., same items in both conditions, or a difficulty-calibration table); otherwise limit the claim to the observed drop.
minor comments (4)
  1. [Table 3] The 'Mean (%)' for Qwen3-Omni-30B is computed over five submissions from only two teams; the text acknowledges this, but the table itself is unannotated. Add an asterisk or a note so the mean is not over-read.
  2. [§2.4 / Abstract] The 25.46% random-guess level is computed on the development set. The statement 'Every system beats the random-guess baseline' should clarify that eval-set scores are being compared with a dev-set estimate, not a matched eval-set baseline.
  3. [§3] The rule that 'gaps of about a percentage point should be read as ties' would benefit from a binomial confidence-interval justification for n=3000; as written it is a reasonable rule of thumb but presented as a fixed threshold.
  4. [Fig. 2] The caption mentions a dashed line marking the 49.65% mean, but the line is not labeled in the plot itself. Add a legend entry.

Circularity Check

2 steps flagged

ADQA-Bench's 'audio-dependent' property is defined by the ADF filter itself, whose retention rule admits text-solvable items and whose validity is inherited from overlapping-author prior work.

specific steps
  1. self definitional [§2.1–2.2 (Task Definition; Hard-ADF)]
    "Every question must be audio-dependent: the correct answer cannot be recovered from the text alone. ... An item is discarded if more than one blind model still picks the correct option, since the answer is then largely recoverable without listening."

    The definition of audio-dependent is 'the correct answer cannot be recovered from the text alone.' The only enforcement mechanism, Hard-ADF, discards an item only when more than one silent-audio model picks the correct option. An item that exactly one silent model can solve from text is therefore deliberately retained, even though the answer is text-recoverable by that model. The operational property is thus not the stated definition but 'survives the ADF cascade,' and items that survive are reported as audio-dependent. This is a definitional loop: ADQA-Bench membership is the filter output, and the filter output is treated as proof of audio dependence.

  2. self citation load bearing [§2.2, §2.3, §5.2 (ADF; Datasets; Dev-to-Eval Drop)]
    "The ADF cascade [1] is a four-stage pipeline (Fig. 1) that keeps only items whose correct answer depends on the audio signal. ... The hidden ADQA-Bench holds 3000 MCQ items filtered under the same protocol."

    The completeness of ADF is the load-bearing premise for calling the 3000 items audio-dependent and for interpreting the 11.91 pp dev-to-eval drop as evidence about a 'harder' audio split. That premise is imported from [1], whose author list overlaps with the present paper's, and is not independently validated or machine-checked here. The paper's own random-guess floor (25.46%, 'likely due to option-length and distractor artifacts') shows residual text-only signals survive the filter. Thus the central benchmark-validity claim reduces to an unverified self-citation rather than to external evidence.

full rationale

The challenge results themselves (leaderboard accuracies, dev-to-eval gaps, per-question failure distributions) are new measurements of submitted systems on a fixed evaluation set and are not fitted from the benchmark's construction, so the empirical claims have independent content. The circularity is concentrated in the benchmark's defining property: ADQA-Bench items are called 'audio-dependent,' but this property is operationalized as 'survives the ADF cascade.' Hard-ADF's more-than-one rule retains items that exactly one silent-audio model can answer from text, contradicting the §2.1 definition, and the filter's completeness is asserted via the authors' own prior work [1] rather than demonstrated. The paper even reports a 25.46% random-guess floor attributed to option-length and distractor artifacts, i.e., residual non-audio signals. Because all headline interpretations—58.33% top accuracy, the 11.91 pp dev-to-eval drop, and the 233 universal failures—are framed as evidence about audio understanding, the validity of the benchmark is load-bearing and rests on the self-cited ADF. This is a partial, definitional circularity (score 4), not a full reduction of the results to the inputs (score 6+), because the numerical accuracy measurements are not derived by fitting the benchmark's filter parameters.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The central claims rest on filter design choices inherited from the authors' prior work [1] plus domain assumptions about how text solvability is detectable. The winning-system hyperparameters (λ, option oversampling, binning quotas) are reported recipe details, not derivations. The 25.46% random-guess floor is both an honest standard-math adjustment and evidence that the filter's guarantee is incomplete.

free parameters (5)
  • Hard-ADF disqualification threshold = >1 blind-model correct picks
    An item is dropped only if more than one silent-audio LALM picks the correct option; items with 0 or 1 blind-model hits pass. Permissive, hand-set criterion from [1] (§2.2).
  • Soft-ADF perplexity criterion = correct option has lowest silent-audio perplexity
    Items are dropped when the correct option has the lowest text perplexity under silent audio; criterion inherited from [1] and summarized in §2.2.
  • LLM commonsense filter criterion = text-only LLM answers correctly
    Items the LLM can answer from commonsense are pruned; items it fails on are kept even if another text-only reasoner could solve them (§2.2).
  • Lim_CAU ensemble weight λ = ≈0.46
    Weighted log-prob fusion of MOSS-Audio-8B-Thinking and Qwen3-Omni-30B views in the winning system (§4.4). Reported recipe detail, not load-bearing for the benchmark's validity claim.
  • Lim_CAU option-layout oversampling = A/B/C/D = 1.0/1.0/1.0/1.3
    Oversamples option-D targets to counter reported positional bias in MOSS outputs (§4.4). Winning-system training detail, hand-set.
axioms (6)
  • domain assumption A LALM answering correctly with silent audio identifies text-solvable items
    Hard-ADF's validity; assumes the blind model pool exploits the same textual priors as evaluated systems (§2.2).
  • domain assumption Perplexity of the correct option under silent audio exposes linguistic shortcuts
    Soft-ADF's validity; assumes LM perplexity ranks options by text-solvability (§2.2).
  • domain assumption A text-only LLM's commonsense success approximates human text-only solvability
    LLM commonsense filter; items the LLM fails on can still be text-solvable by other means (§2.2).
  • domain assumption Two-pass human review guarantees correct labels and plausible-but-wrong distractors
    No inter-annotator agreement or label audit is reported; the 233 universal failures are meaningful as capability gaps only if labels are correct (§2.2, §5.1).
  • domain assumption Dev and eval splits are filtered under equivalent protocols, making dev-to-eval deltas interpretable
    The 11.91 pp mean drop and the 'eval designed to be harder' claim (§5.2) assume filter parity across splits built from different source pools at different times.
  • standard math Random-guess floor computed from observed option artifacts (25.46%), standard binomial reasoning
    Four-way MCQ floor; the paper correctly uses the empirical floor rather than the ideal 25% (§2.4).
invented entities (2)
  • ADQA-Bench independent evidence
    purpose: Hidden 3000-item evaluation set whose items allegedly cannot be answered from text alone; defines the challenge's headline numbers.
    Falsifiable handles exist: future systems can be scored on it, a text-only probe could test the audio-dependence claim, and the paper reports oracle headroom (92.2%) and 233 universal failures. A benchmark, not a graviton-style entity.
  • Audio-Dependency Filtering (ADF) cascade independent evidence
    purpose: Four-stage procedure (silent probe, perplexity, LLM commonsense, human review) claimed to guarantee audio dependence of surviving items.
    The development split is public, so the filter's outputs are auditable, and its completeness is a falsifiable claim (a text-only system at chance on survivors would confirm; the 25.46% floor hints at leakage). However, it originates in the authors' own prior work [1] and is not independently benchmarked here.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Summary of DCASE 2026 Task 5: Audio-Dependent Question Answering." pith.science (2026). https://pith.science/paper/D4CQMV56

@misc{pith2026260718718,
  author       = {Pith},
  title        = {Pith review of: Summary of DCASE 2026 Task 5: Audio-Dependent Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4CQMV56}},
  note         = {Machine review of arXiv:2607.18718}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

DCASE~2026 Task~5 introduces Audio-Dependent Question Answering (ADQA), which tests whether large audio-language models answer from the audio rather than from textual priors. An Audio-Dependency Filtering (ADF) pipeline combines silent-audio probing, per-option perplexity, a large language model (LLM) commonsense check, and human review to remove items solvable from text alone. The 3000 items that pass form the ADQA-Bench evaluation set, spanning music, speech, and environmental audio. The inaugural edition draws 14 teams and 36 submissions across two tracks defined by total parameter count (up to 100B and under 10B). A Chung-Ang University ensemble of MOSS-Audio-8B-Thinking and Qwen3-Omni-30B reaches the top overall accuracy at \pct{58.33}, and a MOSS-only configuration from the same team leads the sub-10B track at \pct{57.30}. Across the 30 submissions with a comparable development score, evaluation accuracy falls by 11.91 percentage points (pp) on average (median 10.91\,pp) on the hidden evaluation split, which is designed to be harder than the development split. The most common building blocks are: the MOSS-Audio-8B-Thinking backbone (13 of 36 submissions), Low-Rank Adaptation (LoRA) fine-tuning on AudioMCQ-StrongAC, and preference or reinforcement-learning objectives -- Group Relative Policy Optimization (GRPO) in five teams, Group reward-Decoupled Normalization Policy Optimization (GDPO) in two. At test time, prompt engineering is near-universal, and majority or choice-permutation voting is common. Every system misses the same set of 233 evaluation items.

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

35 extracted references · 16 linked inside Pith

  1. [1]

    Measuring audio’s impact on correctness: Audio-contribution-aware post-training of large audio language models,

    H. He, X. Du, R. Sun, Z. Dai, Y . Xiao, M. Yang, J. Zhou, X. Li, Z. Liu, Z. Liang, C. Wu, Q. He, T. Lee, X. Chen, W.-L. Zheng, W. Wang, M. D. Plumbley, J. Liu, and Q. Kong, “Measuring audio’s impact on correctness: Audio-contribution-aware post-training of large audio language models,” arXiv preprint arXiv:2509.21060, 2025. [Online]. Available: https://ar...

    arXiv 2025

  2. [2]

    MMAU: A massive multi-task audio understanding and reasoning benchmark,

    S. Sakshi, U. Tyagi, S. Kumar, A. Seth, R. Selvakumar, O. Nieto, R. Duraiswami, S. Ghosh, and D. Manocha, “MMAU: A massive multi-task audio understanding and reasoning benchmark,” inProc. Int. Conf. Learn. Represent. (ICLR), 2025. [Online]. Available: https://arxiv.org/abs/2410.19168

    Pith/arXiv arXiv 2025

  3. [3]

    MMAR: A challenging benchmark for deep reasoning in speech, audio, music, and their mix,

    Z. Ma, Y . Ma, Y . Zhu, C. Yang, Y .-W. Chao, R. Xu, W. Chen, Y . Chen, Z. Chen, J. Cong, K. Li, K. Li, S. Li, X. Li, X. Li, Z. Lian, Y . Liang, M. Liu, Z. Niu, T. Wang, Y . Wang, Y . Wang, Y . Wu, G. Yang, J. Yu, R. Yuan, Z. Zheng, Z. Zhou, H. Zhu, W. Xue, E. Benetos, K. Yu, E.-S. Chng, and X. Chen, “MMAR: A challenging benchmark for deep reasoning in sp...

    Pith/arXiv arXiv 2025

  4. [4]

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

    D. Wang, J. Li, J. Wu, D. Yang, X. Chen, T. Zhang, and H. Meng, “MMSU: A massive multi-task spoken language understanding and reasoning benchmark,” arXiv preprint arXiv:2506.04779, 2025. [Online]. Available: https://arxiv.org/abs/2506.04779

    Pith/arXiv arXiv 2025

  5. [5]

    Audio-grounded hard-example training with acoustic tagging for audio-dependent question answering,

    H. J. Kim, B. Kim, J. C. Ryu, Y . R. Kim, Y . Oh, B. E. Choi, C. Lim, and I.-Y . Kwak, “Audio-grounded hard-example training with acoustic tagging for audio-dependent question answering,” Chung-Ang Univ., DCASE2026 Challenge, Jun. 2026

    2026

  6. [6]

    Qwen3-Omni technical report,

    J. Xuet al., “Qwen3-Omni technical report,” arXiv preprint arXiv:2509.17765, 2025. [Online]. Available: https://arxiv.org/abs/2509. 17765

    Pith/arXiv arXiv 2025

  7. [7]

    Fun-Audio-Chat technical report,

    Tongyi Fun Teamet al., “Fun-Audio-Chat technical report,” arXiv preprint arXiv:2512.20156, 2025. [Online]. Available: https://arxiv.org/ abs/2512.20156

    arXiv 2025

  8. [8]

    Kimi-Audio technical report,

    KimiTeamet al., “Kimi-Audio technical report,” arXiv preprint arXiv:2504.18425, 2025. [Online]. Available: https://arxiv.org/abs/2504. 18425

    Pith/arXiv arXiv 2025

  9. [9]

    MiMo-Audio: Audio language models are few-shot learners,

    Xiaomi LLM-Core Teamet al., “MiMo-Audio: Audio language models are few-shot learners,” arXiv preprint arXiv:2512.23808, 2025. [Online]. Available: https://arxiv.org/abs/2512.23808

    arXiv 2025

  10. [10]

    Step-Audio 2 technical report,

    B. Wuet al., “Step-Audio 2 technical report,” arXiv preprint arXiv:2507.16632, 2025. [Online]. Available: https://arxiv.org/abs/2507. 16632

    Pith/arXiv arXiv 2025

  11. [11]

    MOSS-Audio technical report,

    C. Yanget al., “MOSS-Audio technical report,” arXiv preprint arXiv:2606.01802, 2026. [Online]. Available: https://arxiv.org/abs/2606. 01802

    Pith/arXiv arXiv 2026

  12. [12]

    GDPO: Group reward-decoupled normalization policy optimization for multi-reward RL optimization,

    S.-Y . Liuet al., “GDPO: Group reward-decoupled normalization policy optimization for multi-reward RL optimization,” arXiv preprint arXiv:2601.05242, 2026. [Online]. Available: https://arxiv.org/abs/2601. 05242

    Pith/arXiv arXiv 2026

  13. [13]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inProc. Int. Conf. Learn. Represent. (ICLR), 2022. [Online]. Available: https://arxiv.org/abs/2106.09685

    Pith/arXiv arXiv 2022

  14. [14]

    Learning from audio-dependency errors: Data curation strategies based on model confusion patterns in audio question answering,

    H. Nam, “Learning from audio-dependency errors: Data curation strategies based on model confusion patterns in audio question answering,” Independent, DCASE2026 Challenge, Jun. 2026

    2026

  15. [15]

    QLoRA: Efficient finetuning of quantized LLMs,

    T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer, “QLoRA: Efficient finetuning of quantized LLMs,” inProc. NeurIPS, 2023. [Online]. Available: https://arxiv.org/abs/2305.14314

    Pith/arXiv arXiv 2023

  16. [16]

    Audio-dependent question answering at the DCASE 2026 challenge,

    W. Hu, Y . Cao, and J. Yang, “Audio-dependent question answering at the DCASE 2026 challenge,” Inst. of Acoustics, CAS, DCASE2026 Challenge, Jun. 2026

    2026

  17. [17]

    Training-free inference- time exploration for audio-dependent question answering,

    Z. Yin, Q. Cao, P. Deng, Y . Tan, and S. Li, “Training-free inference- time exploration for audio-dependent question answering,” XJTLU, DCASE2026 Challenge, Jun. 2026

    2026

  18. [18]

    Selective multi-modal RAG for DCASE 2026 task 5: Audio-dependent question answering,

    Y . Cheng, J. Zhao, R. Wan, P. Chang, Y . Chen, and W. Wang, “Selective multi-modal RAG for DCASE 2026 task 5: Audio-dependent question answering,” Univ. of Surrey, DCASE2026 Challenge, Jun. 2026

    2026

  19. [19]

    Structured audio reasoning and robust multi- sample inference for DCASE 2026 audio-dependent question answering challenge,

    Y . Zhang, J. Liu, and M. Li, “Structured audio reasoning and robust multi- sample inference for DCASE 2026 audio-dependent question answering challenge,” Wuhan Univ. and CUHK-Shenzhen, DCASE2026 Challenge, Jun. 2026

    2026

  20. [20]

    DCASE 2026 audio-dependent question answering task,

    A. Tathe, “DCASE 2026 audio-dependent question answering task,” UIUC and CMU, DCASE2026 Challenge, Jun. 2026

    2026

  21. [21]

    Curriculum learning for audio-dependent question answering: Technical report for DCASE 2026 task 5,

    Q. Huang, Y . Pan, X. Zhou, R. Li, and M. Unoki, “Curriculum learning for audio-dependent question answering: Technical report for DCASE 2026 task 5,” JAIST, DCASE2026 Challenge, Jun. 2026

    2026

  22. [22]

    Task-leaf routed MiMo-Audio for DCASE 2026 task 5,

    J. Kim, L. Song, and H.-M. Park, “Task-leaf routed MiMo-Audio for DCASE 2026 task 5,” Sogang Univ., DCASE2026 Challenge, Jun. 2026

    2026

  23. [23]

    Qwen2.5-Omni with all-audio audio-TAH for audio-dependent question answering,

    C. Wu, D. Wu, and Y . Huang, “Qwen2.5-Omni with all-audio audio-TAH for audio-dependent question answering,” Xiamen Univ. and Tsinghua Univ., DCASE2026 Challenge, Jun. 2026

    2026

  24. [24]

    Fun-Audio-Chat-8B with LoRA fine-tuning for audio- dependent question answering,

    C. Zheng, “Fun-Audio-Chat-8B with LoRA fine-tuning for audio- dependent question answering,” Hangzhou Dianzi Univ., DCASE2026 Challenge, Jun. 2026

    2026

  25. [25]

    GISP@HEU’s submission for DCASE 2026 task 5: A LoRA-SFT fine-tuned audio-dependent question answering system,

    F. Xiao, Q. Zhu, and J. Guan, “GISP@HEU’s submission for DCASE 2026 task 5: A LoRA-SFT fine-tuned audio-dependent question answering system,” HEU and UTS, DCASE2026 Challenge, Jun. 2026

    2026

  26. [26]

    Audio-dependent question answering with attention-anchored reinforcement learning on MiMo-Audio,

    H. Song, “Audio-dependent question answering with attention-anchored reinforcement learning on MiMo-Audio,” BIT, DCASE2026 Challenge, Jun. 2026

    2026

  27. [27]

    Audio question answering at the DCASE 2026 challenge,

    H. Xu and R. Zhang, “Audio question answering at the DCASE 2026 challenge,” HUST, DCASE2026 Challenge, Jun. 2026

    2026

  28. [28]

    Qwen2.5-Omni technical report,

    J. Xuet al., “Qwen2.5-Omni technical report,” arXiv preprint arXiv:2503.20215, 2025. [Online]. Available: https://arxiv.org/abs/2503. 20215

    Pith/arXiv arXiv 2025

  29. [29]

    DeepSeekMath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . K. Li, Y . Wu, and D. Guo, “DeepSeekMath: Pushing the limits of mathematical reasoning in open language models,” arXiv preprint arXiv:2402.03300, 2024. [Online]. Available: https://arxiv.org/abs/2402.03300

    Pith/arXiv arXiv 2024

  30. [30]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inProc. ICML, 2023. [Online]. Available: https://arxiv.org/abs/2212.04356

    Pith/arXiv arXiv 2023

  31. [31]

    pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe,

    H. Bredin, “pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe,” inProc. Interspeech, 2023

    2023

  32. [32]

    ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,” inProc. Interspeech, 2020. [Online]. Available: https://arxiv.org/abs/2005.07143

    Pith/arXiv arXiv 2020

  33. [33]

    Large-scale Contrastive Language-Audio Pretraining with feature fusion and keyword-to-caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, M. Nezhurina, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale Contrastive Language-Audio Pretraining with feature fusion and keyword-to-caption augmentation,” inProc. ICASSP, 2023. [Online]. Available: https://arxiv.org/abs/2211.06687

    Pith/arXiv arXiv 2023

  34. [34]

    librosa: Audio and music signal analysis in Python,

    B. McFee, C. Raffel, D. Liang, D. P. W. Ellis, M. McVicar, E. Battenberg, and O. Nieto, “librosa: Audio and music signal analysis in Python,” in Proc. SciPy, 2015

    2015

  35. [35]

    Prometheus 2: An open source language model specialized in evaluating other language models,

    S. Kim, J. Suk, S. Longpre, B. Y . Lin, J. Shin, S. Welleck, G. Neubig, M. Lee, K. Lee, and M. Seo, “Prometheus 2: An open source language model specialized in evaluating other language models,” arXiv preprint arXiv:2405.01535, 2024. [Online]. Available: https://arxiv.org/abs/2405.01535

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.