EvA: An Evidence-First Audio Understanding Paradigm for LALMs

Benyou Wang; Liyin Liang; Shunian Chen; Xinyuan Xie; Yuhao Zhang; Zhiheng Liu; Zhiqiang Lv

REVIEW 2 major objections 5 minor 15 references

Stronger audio understanding depends on preserving acoustic evidence before reasoning begins, not on better reasoning alone.

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 ↗

T0 review · grok-4.5

2026-07-13 16:49 UTC pith:TAXJHO3F

load-bearing objection Solid dual-path systems paper with real gains and careful decontamination; the evidence-bottleneck story is useful framing but the Perception/Reasoning split is too soft to carry the causal claim alone. the 2 major comments →

arxiv 2603.27667 v2 pith:TAXJHO3F submitted 2026-03-29 cs.SD cs.AI

EvA: An Evidence-First Audio Understanding Paradigm for LALMs

Xinyuan Xie , Shunian Chen , Zhiheng Liu , Yuhao Zhang , Zhiqiang Lv , Liyin Liang , Benyou Wang This is my paper

classification cs.SD cs.AI

keywords large audio language modelsevidence bottleneckdual-path fusionacoustic perceptionhierarchical aggregationtime-aligned fusionaudio understandingEvA-Perception

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.

Large audio language models still fail in complex sound scenes mainly because they lose task-relevant acoustic evidence before any reasoning starts. The authors call this the evidence bottleneck: current systems lag humans far more on perception-style tasks (finding events, entities, and order) than on reasoning over those findings. EvA answers with a dual-path design that keeps a speech encoder and a non-speech event encoder, aggregates multi-band and multi-layer non-speech features, then fuses them onto a shared time line without compressing tokens. Trained on a new set of event-ordered captions and evidence-grounded questions, EvA leads open-source perception scores on major benchmarks, with the biggest lifts on perception-heavy splits, and human judges rate its captions as covering more fine-grained acoustic detail. The claim is simple: better listening infrastructure is the prerequisite for better audio understanding.

Core claim

Current large audio language models are limited more by how well they preserve task-relevant acoustic evidence upstream than by how well they reason downstream. Model–human gaps are substantially larger on perception-oriented tasks than on reasoning tasks, pointing to an evidence bottleneck created by speech-centric encoders, lossy compression, and weak temporal alignment. EvA removes that bottleneck by hierarchical multi-frequency, multi-layer aggregation of a non-speech encoder and non-compressive, time-aligned gated fusion into the language model’s audio tokens.

What carries the argument

EvA dual-path fusion: hierarchical evidence aggregation (frequency-gated pooling plus cascaded cross-layer attention over multiple CED layers) followed by time-aware alignment and gated inject-and-add fusion onto the Whisper token timeline, preserving full sequence length and temporal resolution.

Load-bearing premise

That splitting benchmark tasks into “perception” versus “reasoning” cleanly measures upstream evidence loss rather than leftover reasoning demands or label noise inside those subtasks.

What would settle it

Train a matched model that drops the non-speech path or replaces the non-compressive fusion with heavy temporal compression, then check whether it still matches EvA on the same perception-labeled splits after equal reasoning-side training; if it does, the claim that upstream evidence preservation is the primary bottleneck fails.

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

If this is right

Front-end encoder and fusion design must be fixed before further backend fine-tuning or reinforcement learning can recover lost acoustic cues.
Dual-path, time-aligned, non-compressive fusion improves perception accuracy across different language-model backbones.
Evidence-grounded data with event order and fine-grained QA raises fine-grained acoustic coverage in open-ended captioning.
Largest gains concentrate on perception-heavy benchmark splits, matching the evidence-first hypothesis.
The same fusion pattern transfers when ported to another audio-language architecture.

Where Pith is reading between the lines

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

If the bottleneck is real, scaling chain-of-thought or RL reasoning alone will hit diminishing returns until non-speech encoders and fusion interfaces improve.
Future general-audio models may need native multi-band, multi-layer access rather than final-layer-only speech encoders.
The perception-versus-reasoning diagnostic split could become a reusable lens for other multimodal systems that undervalue upstream sensing.
Multilingual captions and expert music concepts remain open tests of whether evidence preservation generalizes beyond the current English-centric training mix.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Solid dual-path systems paper with real gains and careful decontamination; the evidence-bottleneck story is useful framing but the Perception/Reasoning split is too soft to carry the causal claim alone.

read the letter

The useful takeaway is practical: EvA is a clean dual-path fix on Kimi-Audio that keeps non-speech evidence at full temporal resolution, ships a large evidence-grounded dataset, and posts the best open-source Perception numbers on MMAU-Clean, MMAR, and MMSU under a unified zero-shot protocol. That is enough to make the paper worth reading if you build or evaluate LALMs.

What is actually new is the concrete fusion stack, not the dual-path idea itself. SALMONN and GAMA already had two streams; EvA’s combination—multi-layer CED aggregation with frequency-gated pooling, coverage-weighted time-aware interpolation onto the Whisper timeline, and gated inject-and-add that preserves token length—is a genuine engineering step past Q-Former compression and unaligned concatenation. The ablations isolate the CED path and the hierarchical aggregator; human captioning shows better label recall and object coverage without collapsing accuracy; and the MMAU decontamination (CLAP ≥0.80 plus manual audit, 149 samples removed) is more careful than most papers in this area. Gains transfer, at least partially, when they bolt the same design onto Qwen2.5-Omni. The DPI/chain-rule motivation is standard and qualitative; it does not force the result, which is fine.

The soft spot is real but not fatal. The central “evidence bottleneck” diagnosis leans hard on an author-defined Perception/Reasoning taxonomy (Appendix A.13). Many Perception items still involve order, concurrency, or light scene inference, so larger gaps and gains on that half are correlational support rather than a clean causal isolation of upstream evidence loss. Ablations and the human arena still show the dual path helps; they just do not fully rescue the split-dependent attribution. Other limitations are ordinary: English-only captions, soft AudioSet-Strong boundaries, pending full code/checkpoints, and a few free parameters (gate α, LoRA, layer indices, CLAP threshold). None of that sinks the systems contribution.

This is for people who train or evaluate audio-language models and care about non-speech evidence and temporal fidelity. It deserves a serious referee. I would bring it to reading group, cite the architecture and dataset when discussing dual-path fusion, and accept it for peer review.

Referee Report

2 major / 5 minor

Summary. The paper diagnoses an “evidence bottleneck” in Large Audio Language Models: larger model–human gaps on perception-oriented than reasoning-oriented splits of MMAU, MMAR and MMSU (Fig. 1) are taken to show that upstream acoustic-evidence preservation, not downstream reasoning, is often the limiting factor. It proposes EvA, a dual-path architecture on Kimi-Audio-7B that freezes Whisper and CED-Base, hierarchically aggregates multi-layer multi-band CED features (Eqs. 2–4), time-aligns them to the Whisper timeline without compression (Alg. 1), and injects them via gated additive fusion (Eqs. 5–6). Training uses a new EvA-Perception corpus (~54K event-ordered captions, ~500K evidence-grounded QA pairs) in a two-stage alignment + LoRA SFT recipe. Under a unified zero-shot protocol with a CLAP-decontaminated MMAU-Clean split, EvA reports the strongest open-source Perception numbers on MMAU-Clean, MMAR and MMSU, largest gains on perception-heavy splits, competitive CochlScene performance, supportive ablations (Table 3), transfer to Qwen2.5-Omni (Table 5), and improved human captioning metrics (Table 4).

Significance. If the evidence-first diagnosis and the architectural remedy hold, the work supplies a concrete, transferable design principle for LALMs: preserve complementary speech/non-speech evidence at full temporal resolution before the LLM, rather than relying solely on reasoning-backend SFT/RL. Strengths that raise the contribution above a pure engineering report include (i) a carefully audited MMAU-Clean protocol with released match files, (ii) ablations that isolate the CED path, frequency-gated pooling and non-compressive cross-layer fusion against a Q-Former baseline, (iii) a human open-ended captioning arena that measures fine-grained acoustic coverage beyond multiple-choice accuracy, and (iv) a successful transfer experiment on a second backbone. The information-flow motivation (DPI under deterministic maps, chain-rule complementarity) is standard and does not over-claim; the empirical package is the main contribution.

major comments (2)

The central attribution of both the model–human gap (Fig. 1) and EvA’s largest gains (Table 2) to an “evidence bottleneck” rests on the author-defined Perception/Reasoning partition in Appendix A.13 (MMAU information-extraction → Perception; MMAR Signal+Perception → Perception; MMSU native tags retained). Many items still require event order, concurrency or scene inference, so residual reasoning demand and label noise are not cleanly isolated. Without an independent validation of the taxonomy (e.g., human rating of residual inference load, or a controlled probe set of pure entity/temporal extraction), the claim that upstream evidence preservation—not downstream reasoning—is the primary limiting factor remains correlational. The dual-path ablations and human captioning results still show that the architecture helps; they do not fully rescue the split-dependent causal story.
Table 2 and the Stage-2 ablations (Table 3) compare EvA to the Kimi-Audio-7B-Instruct backbone and to a same-data single-path baseline, but do not fully disentangle the contribution of the new architecture from the contribution of EvA-Perception itself. A control that fine-tunes the original Kimi-Audio backbone (or a strong peer) on the same EvA-Perception mixture without the CED Aggregator / inject-and-add path would clarify how much of the Perception lift is data versus fusion design. The existing S2(1) “w/o CED path” row is helpful but still uses the Whisper adapter and LoRA recipe introduced for EvA; a cleaner backbone-only control would strengthen the architectural claim.

minor comments (5)

Fig. 1 averages only two models; stating the per-model gaps (or adding error bars) would make the perception–reasoning disparity more transparent.
Eqs. (5)–(6) introduce a learnable gate α and a √2 scaling factor without reporting the learned α range or an ablation on the scaling; a short note would aid reproducibility.
The visual-leakage audit (Appendix A.11) flags 1.4% of captions; stating whether those items were retained or filtered in the released training set would close a small data-quality loop.
Table 1 caption length (67.99) is much longer than prior corpora; a short qualitative example of how length correlates with temporal ordering would help readers interpret the comparison.
Minor notation: H_l vs. H_agg vs. H_aligned are introduced densely in §4.1; a one-line symbol table would reduce cognitive load.

Circularity Check

0 steps flagged

No circular derivation: evidence-bottleneck framing and dual-path design are motivated by observation and standard DPI/chain-rule arguments, then tested on external benchmarks and human evaluation.

full rationale

The paper’s load-bearing chain is observational and empirical, not definitional. Sec. 3 motivates an “evidence bottleneck” from measured model–human gaps (Fig. 1) and applies the standard data-processing inequality under fixed deterministic maps (Eq. 1) plus the chain-rule complementarity of dual paths; neither identity is fitted to the target metrics nor defined in terms of EvA’s later scores. The architecture (hierarchical frequency/layer aggregation, time-aware non-compressive inject-and-add fusion) is a design choice, not a quantity recovered from the evaluation labels. EvA-Perception is constructed from AudioSet-Strong and expert models, then the model is evaluated zero-shot on external suites (MMAU-Clean, MMAR, MMSU, CochlScene) and a human captioning arena; ablations (Table 3) hold training data fixed when isolating the CED path. Self-citations (e.g., FusionAudio as a dataset comparison) are not uniqueness theorems and do not force the Perception gains. Author-defined Perception/Reasoning taxonomy (App. A.13) may confound causal attribution of those gains, but that is a validity concern, not a reduction of a claimed prediction to its inputs by construction. No fitted parameter is renamed as a prediction; no result is equivalent to its inputs by definition.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 3 invented entities

The central empirical claim rests on standard information-theoretic inequalities, the assumption that CED and Whisper supply complementary evidence, the authors’ Perception/Reasoning taxonomy, and a handful of architectural free parameters (gate init, LoRA rank, interpolation weights). No new physical entities are postulated; the ‘evidence bottleneck’ is a diagnostic label, not an ontological invention.

free parameters (4)

learnable gate α = small positive init (exact value not stated)
Scalar initialized small and trained; controls how much CED evidence is added to each audio token (Eq. 5).
LoRA rank / alpha = 64 / 64
r=64, α=64 on q/k/v/o projections; chosen for Stage-2 SFT capacity.
CED layer indices for aggregation = 4, 8, L
Layers 4, 8, L selected by hand for multi-scale fusion.
CLAP decontamination threshold = 0.80
Cosine ≥0.80 chosen after manual bucket inspection to remove source-audio duplicates.

axioms (3)

standard math Data-processing inequality for deterministic maps: I(Z;Y) ≤ I(Z;O) ≤ I(Z;H) ≤ I(Z;X) under fixed parameters.
Invoked in §3.1 and Appendix A.6 to argue that post-training cannot restore evidence lost upstream.
domain assumption Whisper and CED-Base extract complementary acoustic evidence (speech vs non-speech).
Stated in §3.2 and used to justify the dual-path design; supported by ablations but not independently proven for every scene.
ad hoc to paper The authors’ Perception/Reasoning partition of MMAU/MMAR/MMSU subtasks correctly isolates evidence extraction from higher-order inference.
Defined in §5.1 and Appendix A.13; load-bearing for the ‘bottleneck’ diagnosis.

invented entities (3)

evidence bottleneck no independent evidence
purpose: Diagnostic label for the observed larger model–human gap on perception versus reasoning.
Introduced in the abstract and §1; useful framing but not an independently measurable physical quantity.
EvA dual-path architecture (hierarchical Aggregator + inject-and-add fusion) independent evidence
purpose: Concrete mechanism claimed to preserve multi-scale non-speech evidence without temporal compression.
Core technical contribution; falsifiable via the reported ablations and transfer to Qwen2.5-Omni.
EvA-Perception dataset independent evidence
purpose: Evidence-grounded captions and QA pairs used for alignment and instruction tuning.
New resource; construction pipeline is described and a visual-leakage audit is supplied.

pith-pipeline@v1.1.0-grok45 · 26261 in / 2956 out tokens · 34268 ms · 2026-07-13T16:49:01.286979+00:00 · methodology

0 comments

read the original abstract

Large Audio Language Models (LALMs) still struggle in complex acoustic scenes because they often fail to preserve task-relevant acoustic evidence before reasoning begins. We identify this error pattern as the evidence bottleneck: state-of-the-art systems show larger deficits in acoustic evidence extraction than in downstream reasoning, suggesting that upstream perception is often the limiting factor. To address this problem, we propose EvA (Evidence-First Audio), a dual-path architecture that enhances acoustic evidence preservation through hierarchical aggregation and non-compressive, time-aligned fusion. We also build EvA-Perception, a large-scale training set with about 54K event-ordered captions and 500K evidence-grounded QA pairs. Under a unified zero-shot protocol, EvA achieves the best open-source \emph{Perception} results on MMAU, MMAR, and MMSU, with the largest gains on perception-heavy splits. Human evaluation on open-ended captioning further shows improved fine-grained acoustic coverage and caption quality. These results support the evidence-first hypothesis: stronger audio understanding depends on preserving acoustic evidence before reasoning. Project can be found at https://satsuki2486441738.github.io/EvA/.

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 5 linked inside Pith

[1]

arXiv preprint arXiv:2506.01111

Fusionaudio-1.2 m: Towards fine-grained au- dio captioning with multimodal contextual fusion. arXiv preprint arXiv:2506.01111. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, and 1 others. 2024. Qwen2-audio technical report.arXiv preprint arXiv:2407.10759. Gheorghe Comanici, Eric Biebe...

Pith/arXiv arXiv 2024
[2]

Arushi Goel, Sreyan Ghosh, Jaehyeon Kim, Sonal Ku- mar, Zhifeng Kong, Sang gil Lee, Chao-Han Huck Yang, Ramani Duraiswami, Dinesh Manocha, Rafael Valle, and Bryan Catanzaro

Gama: A large audio-language model with ad- vanced audio understanding and complex reasoning abilities.arXiv preprint arXiv:2406.11768. Arushi Goel, Sreyan Ghosh, Jaehyeon Kim, Sonal Ku- mar, Zhifeng Kong, Sang gil Lee, Chao-Han Huck Yang, Ramani Duraiswami, Dinesh Manocha, Rafael Valle, and Bryan Catanzaro. 2025. Audio flamingo 3: Advancing audio intelli...

Pith/arXiv arXiv 2025
[3]

Il-Young Jeong and Jeongsoo Park

IEEE. Il-Young Jeong and Jeongsoo Park. 2022. Cochlscene: Acquisition of acoustic scene data using crowdsourc- ing.Preprint, arXiv:2211.02289. Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. 2019. Audiocaps: Generating cap- tions for audios in the wild. InProceedings of the 2019 Conference of the North American Chapter of the Association ...

Pith/arXiv arXiv 2022
[4]

Luoyi Sun, Xuenan Xu, Mengyue Wu, and Weidi Xie

Mmau: A massive multi-task audio under- standing and reasoning benchmark.arXiv preprint arXiv:2410.19168. Luoyi Sun, Xuenan Xu, Mengyue Wu, and Weidi Xie

Pith/arXiv arXiv
[5]

dog bark- ing

Auto-acd: A large-scale dataset for audio- language representation learning. InProceedings of the 32nd ACM International Conference on Multime- dia, pages 5025–5034. Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. 2023. Salmonn: Towards generic hearing abilities for large language models.arXiv prepri...

Pith/arXiv arXiv 2023
[6]

Give priority to audio tags

Multimodal Parsing:Extract key sound events, source characteristics, environment, and music elements from each source. Give priority to audio tags. From ASR, only detect voice presence and non-content features, possibly aiding environment/emo- tion inference, but never include speech text itself. 2.Auditory Fact Determination and Cross-modal Correction: •...
[7]

plane), multiple possible sources, or common perceptual misinterpretations.Never include visual-based ambiguities

Pure Auditory Ambiguity Inference:List ambiguities only from hearing, such as similarity of sounds (e.g., car vs. plane), multiple possible sources, or common perceptual misinterpretations.Never include visual-based ambiguities
[8]

sounds like

Reliability Assessment:If audio facts are extremely scarce, or sources are severely conflicting and cannot yield reliable auditory facts, directly output:UNCERTAIN_AUDIO_INFORMATION_DETECTED. 5.Final Audio Caption Generation:If reliable: • Generate a fluent, precise, audio-focused English description that preserves event order, number of occurrences, audi...
[9]

Do not invent facts beyond what the description supports

Grounding:Useonlythe given description as grounding. Do not invent facts beyond what the description supports. Do not produce ASR-style verbatim transcripts
[10]

Each line is one JSON object

Output format:Return JSON Lines (JSONL). Each line is one JSON object. No markdown, no backticks, no extra prose, no blank lines
[11]

No trailing commas

JSON validity:Use straight ASCII quotes. No trailing commas. All keys required. Types must match the schema
[12]

Brevity:Closed-ended answers must be short (yes/no, true/false, a label, a small set, or a number) or MCQ (A–D, answer is a single letter)
[13]

Provide exact words/phrases taken from the description that justify the answer

Evidence: support_span must not be empty. Provide exact words/phrases taken from the description that justify the answer. Do not add commentary
[14]

Set it to trueonlyfor the most difficult items in the batch; otherwise false

Difficulty tagging:Use the boolean field ishard. Set it to trueonlyfor the most difficult items in the batch; otherwise false
[15]

caption/description/text

No meta-reference to source:In both question and answer, never mention the existence of any “caption/description/text”. Write from anaudio-listenerperspective, not a reader perspective. Schema (each JSON object must match): { "question": string, "answer": string, "answer_style": "close" | "open", "type": one of [ "presence", "counting", "temporal_order", ...

[1] [1]

arXiv preprint arXiv:2506.01111

Fusionaudio-1.2 m: Towards fine-grained au- dio captioning with multimodal contextual fusion. arXiv preprint arXiv:2506.01111. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, and 1 others. 2024. Qwen2-audio technical report.arXiv preprint arXiv:2407.10759. Gheorghe Comanici, Eric Biebe...

Pith/arXiv arXiv 2024

[2] [2]

Arushi Goel, Sreyan Ghosh, Jaehyeon Kim, Sonal Ku- mar, Zhifeng Kong, Sang gil Lee, Chao-Han Huck Yang, Ramani Duraiswami, Dinesh Manocha, Rafael Valle, and Bryan Catanzaro

Gama: A large audio-language model with ad- vanced audio understanding and complex reasoning abilities.arXiv preprint arXiv:2406.11768. Arushi Goel, Sreyan Ghosh, Jaehyeon Kim, Sonal Ku- mar, Zhifeng Kong, Sang gil Lee, Chao-Han Huck Yang, Ramani Duraiswami, Dinesh Manocha, Rafael Valle, and Bryan Catanzaro. 2025. Audio flamingo 3: Advancing audio intelli...

Pith/arXiv arXiv 2025

[3] [3]

Il-Young Jeong and Jeongsoo Park

IEEE. Il-Young Jeong and Jeongsoo Park. 2022. Cochlscene: Acquisition of acoustic scene data using crowdsourc- ing.Preprint, arXiv:2211.02289. Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. 2019. Audiocaps: Generating cap- tions for audios in the wild. InProceedings of the 2019 Conference of the North American Chapter of the Association ...

Pith/arXiv arXiv 2022

[4] [4]

Luoyi Sun, Xuenan Xu, Mengyue Wu, and Weidi Xie

Mmau: A massive multi-task audio under- standing and reasoning benchmark.arXiv preprint arXiv:2410.19168. Luoyi Sun, Xuenan Xu, Mengyue Wu, and Weidi Xie

Pith/arXiv arXiv

[5] [5]

dog bark- ing

Auto-acd: A large-scale dataset for audio- language representation learning. InProceedings of the 32nd ACM International Conference on Multime- dia, pages 5025–5034. Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. 2023. Salmonn: Towards generic hearing abilities for large language models.arXiv prepri...

Pith/arXiv arXiv 2023

[6] [6]

Give priority to audio tags

Multimodal Parsing:Extract key sound events, source characteristics, environment, and music elements from each source. Give priority to audio tags. From ASR, only detect voice presence and non-content features, possibly aiding environment/emo- tion inference, but never include speech text itself. 2.Auditory Fact Determination and Cross-modal Correction: •...

[7] [7]

plane), multiple possible sources, or common perceptual misinterpretations.Never include visual-based ambiguities

Pure Auditory Ambiguity Inference:List ambiguities only from hearing, such as similarity of sounds (e.g., car vs. plane), multiple possible sources, or common perceptual misinterpretations.Never include visual-based ambiguities

[8] [8]

sounds like

Reliability Assessment:If audio facts are extremely scarce, or sources are severely conflicting and cannot yield reliable auditory facts, directly output:UNCERTAIN_AUDIO_INFORMATION_DETECTED. 5.Final Audio Caption Generation:If reliable: • Generate a fluent, precise, audio-focused English description that preserves event order, number of occurrences, audi...

[9] [9]

Do not invent facts beyond what the description supports

Grounding:Useonlythe given description as grounding. Do not invent facts beyond what the description supports. Do not produce ASR-style verbatim transcripts

[10] [10]

Each line is one JSON object

Output format:Return JSON Lines (JSONL). Each line is one JSON object. No markdown, no backticks, no extra prose, no blank lines

[11] [11]

No trailing commas

JSON validity:Use straight ASCII quotes. No trailing commas. All keys required. Types must match the schema

[12] [12]

Brevity:Closed-ended answers must be short (yes/no, true/false, a label, a small set, or a number) or MCQ (A–D, answer is a single letter)

[13] [13]

Provide exact words/phrases taken from the description that justify the answer

Evidence: support_span must not be empty. Provide exact words/phrases taken from the description that justify the answer. Do not add commentary

[14] [14]

Set it to trueonlyfor the most difficult items in the batch; otherwise false

Difficulty tagging:Use the boolean field ishard. Set it to trueonlyfor the most difficult items in the batch; otherwise false

[15] [15]

caption/description/text

No meta-reference to source:In both question and answer, never mention the existence of any “caption/description/text”. Write from anaudio-listenerperspective, not a reader perspective. Schema (each JSON object must match): { "question": string, "answer": string, "answer_style": "close" | "open", "type": one of [ "presence", "counting", "temporal_order", ...