File size: 24,598 Bytes
cd08cb0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
---
license: cc-by-4.0
pretty_name: PolyBench/PolyBench (mirror)
source_datasets:
- PolyBench/PolyBench
tags:
- laya
- laya-finetune
- mirror
- domain-general
- use-noncommercial
- genre-other
- label-choice
- annot-author
- annot-llm
- lang-en
language:
- en
task_categories:
- audio-classification
- question-answering
- multiple-choice
task_ids:
- multiple-choice-qa
size_categories:
- 1K<n<10K
configs:
- config_name: classification
  data_files:
  - split: test
    path: classification/test-*
- config_name: concurrency
  data_files:
  - split: test
    path: concurrency/test-*
- config_name: counting
  data_files:
  - split: test
    path: counting/test-*
- config_name: detection
  data_files:
  - split: test
    path: detection/test-*
- config_name: duration
  data_files:
  - split: test
    path: duration/test-*
dataset_info:
- config_name: classification
  features:
  - name: audio
    dtype: audio
  - name: sound_name
    dtype: string
  - name: question
    dtype: string
  - name: choices
    list: string
  - name: answer
    dtype: string
  - name: category
    dtype: string
  splits:
  - name: test
    num_bytes: 1149223260
    num_examples: 728
  download_size: 1092287859
  dataset_size: 1149223260
- config_name: concurrency
  features:
  - name: audio
    dtype: audio
  - name: sound_name
    dtype: string
  - name: question
    dtype: string
  - name: choices
    list: string
  - name: answer
    dtype: string
  - name: category
    dtype: string
  splits:
  - name: test
    num_bytes: 1149194877
    num_examples: 728
  download_size: 1092280880
  dataset_size: 1149194877
- config_name: counting
  features:
  - name: audio
    dtype: audio
  - name: sound_name
    dtype: string
  - name: question
    dtype: string
  - name: choices
    list: string
  - name: answer
    dtype: string
  - name: category
    dtype: string
  splits:
  - name: test
    num_bytes: 1149153249
    num_examples: 728
  download_size: 1092269268
  dataset_size: 1149153249
- config_name: detection
  features:
  - name: audio
    dtype: audio
  - name: sound_name
    dtype: string
  - name: question
    dtype: string
  - name: choices
    list: string
  - name: answer
    dtype: string
  - name: category
    dtype: string
  splits:
  - name: test
    num_bytes: 1149203869
    num_examples: 728
  download_size: 1092286935
  dataset_size: 1149203869
- config_name: duration
  features:
  - name: audio
    dtype: audio
  - name: sound_name
    dtype: string
  - name: question
    dtype: string
  - name: choices
    list: string
  - name: answer
    dtype: string
  - name: category
    dtype: string
  splits:
  - name: test
    num_bytes: 1149167602
    num_examples: 728
  download_size: 1092274924
  dataset_size: 1149167602
---

# PolyBench/PolyBench(複製)

## データの内容

多声の音声(複数の音が重なるクリップ)について推論を問う、選択式の質問応答のデータ。1 行が 1 つの音声と質問の組で、`audio`、`question`、`choices`、正解の `answer` を含む。タスクは音の種類の数え上げ、重なる音の分類、重なりの検出、同時性の判定、長さの比較の 5 つ。正解は音のクラスとタイムスタンプから作り、質問は LLM で言い換えたうえで人手で確かめた。

## データの概要

| 項目 | 値 |
|---|---|
| 行数 | 3,640(5 構成・5 分割の合計) |
| 列数 | 6 |
| 容量 | 6.62 GB(ファイル 759 本) |
| ファイルの形式 | wav 184、parquet 9、csv 5 |
| 言語 | en |
| 公開 | public |

### 構成と分割

| 構成(config) | 分割(split)ごとの行数 | 列数 |
|---|---|---:|
| `classification` | `test` 728 | 6 |
| `concurrency` | `test` 728 | 6 |
| `counting` | `test` 728 | 6 |
| `detection` | `test` 728 | 6 |
| `duration` | `test` 728 | 6 |

### 列(構成 `classification`・分割 `test`。どの構成も同じ列)

| 列 | 型 | 先頭の行の値 |
|---|---|---|
| `audio` | `audio` | `(音声・画像のファイル)` |
| `sound_name` | `string` | `S-0001.wav` |
| `question` | `string` | `During the unresolved period of the Birds sound …` |
| `choices` | `list<string>` | `[4 件]` |
| `answer` | `string` | `Bells` |
| `category` | `string` | `sound` |

行数と列は HF のデータセットビューアの集計(元のデータセットの現在の版で数えたもの)。全データセットの一覧は [HYUKI5/dataset-catalog](https://huggingface.co/datasets/HYUKI5/dataset-catalog) にある。

## 複製について

[`PolyBench/PolyBench`](https://huggingface.co/datasets/PolyBench/PolyBench) をリビジョン `94cde63abaa6b2ec1117f26f222bd6a45c3fe019` の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya(`convaiinnovations/laya`)の微調整に使う候補として、調査メモ(`Laya 学習用データセット調査`)に載せたデータを集めている。

## 出典

- 元のデータセット: https://huggingface.co/datasets/PolyBench/PolyBench
- 複製したリビジョン: `94cde63abaa6b2ec1117f26f222bd6a45c3fe019`(2026-05-20 03:57:21+00:00)
- 元のライセンス表記: `cc-by-4.0`

## 論文

- https://arxiv.org/abs/2603.05128

## アノテーションの方法

**🛠 構築のパイプライン**

このベンチマークは 3 つの段階で構築された(図 1 を参照):

1. **問題の観察。** 単声と多声の音声を比べることで、LALM が同時に存在する音源を幻覚したり混同したりすることが確かめられた。これが、重なりを対象とするベンチマークの動機となった。
2. **データと質問の準備。** 多声のクリップは実世界の録音から選別された。5 つのタスクの種類について、反復的な品質管理を伴う人と LLM の協働によって MCQA が生成された。Classification の誤答の選択肢(distractor)は、クリップに存在しない、意味的に無関係なイベントのカテゴリから選ばれた。意味的な関連性による近道を防ぐためである。
3. **指標と評価。** モデルは、選択肢の完全一致(NV-Embed-v2 による類似度を併用)で採点される。LLM の判定器は、任意で知覚の誤りと推論の誤りを診断する。

---

## 微調整に使うときの注意

- 調査メモの台帳での区分: C(NC・ND・研究専用・評価専用など、商用の利用を妨げうる記載がある)。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
- Laya の入力は 2,048 token まで(ModernBERT-large / mmBERT)。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
- 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。

## 調査メモの記述

| メモ | 節 | ライセンス(記載) | 確認 | 注記 |
|---|---|---|---|---|
| `macro_forecasting.md` | Q3. 予測質問のデータで、質問文 + 解決結果 + 群衆確率の時系列が揃うものは何か (L96) | 論文ページは CC BY-NC-SA 4.0。データ本体の表示は未確認 | 確認済 | 原文注記: 取得は GitHub 経由(OneDrive)と検索結果にあるが、未取得。HF の同名 ID と混同しないこと |

## 元のカード(日本語訳)

以下は元のデータセットのカードの本文を日本語に訳したもの(機械翻訳)。訳に疑問があるときは、末尾の原文を確かめる。

---

# PolyBench: A Benchmark for Compositional Reasoning in Polyphonic Audio

<p align="center">
  <img src="./polybench_1.png" alt="PolyBench pipeline overview" width="100%"/>
</p>

<p align="center">
  <em>図 1. ベンチマークの構築と評価のための PolyBench のパイプライン。</em>
</p>

<p align="center">
  📄 <a href="https://arxiv.org/abs/2603.05128">論文</a> &nbsp;|&nbsp;
  💻 <a href="https://github.com/visionchan/PolyBench">GitHub</a>
</p>

---

## 📖 概要

**PolyBench** は、Large Audio Language Models(LALM)について、**多声の音声における合成的推論**を評価するために特別に設計された最初のベンチマークである。現代の LALM は単声の音声に対する推論はできるが、複数の音イベントが同時に起きると一貫して性能が落ちる。PolyBench は、同時に起きるイベントとその関係についての推論を要する、慎重に設計された 5 つの多肢選択式の質問応答(MCQA)のサブセットを通じて、この根本的な限界を明らかにする。

主な特徴:

- **実世界の多声音声**。3 つの確立された音イベントのデータセット(DataSED、DESED、MAESTRO-Real)からサンプリングしたものに加え、Concurrency のサブセット用に AudioTime の単声のクリップを含む。
- **5 つの階層的な MCQA タスク**。counting、classification、detection、concurrency、duration の各タスクを扱う。
- **人と LLM の協働による質問の生成**。反復的な品質管理により、論理的に一貫した `<question, choices, answer>` の三つ組を確保する。
- **標準化された評価プロトコル**。正解率(Accuracy)と F1 を指標とし、知覚の誤りと推論の誤りを細かく分析するため、LLM を用いた判定器(judge)を備える。

---

## 🧩 タスク

PolyBench は 5 つの MCQA のサブセットを含み、多声の推論の依存関係の同心円状の階層によって構成されている:

| タスク | 目的 | 質問の例 |
|---|---|---|
| **Counting** | 音声に含まれる異なる音イベントのカテゴリの数を特定する。 | *「音声ファイルにはいくつの異なる音源が含まれるか。」* |
| **Duration** | 音声の中で最も長い音イベントを特定する。 | *「どの音イベントが最も長いか。」* |
| **Concurrency** | 音声のクリップに多声の現象が起きているかを判定する。 | *「揚げる音が止まらない間に、同時に別の音イベントが存在するか。」* |
| **Classification** | 指定された基準の音イベントと時間的に重なる、別のイベントのカテゴリを特定する。 | *「終わる前に犬の音イベントと時間的に重なるのは、どの種類の音イベントか。」* |
| **Detection** | 重なりが最初に起きる時間区間と、重なりの起きた回数の合計を推論する。 | *「この音声クリップで、複数の音が最初に重なるのはどの時点か。」* |

各タスクでは、基本となる質問のテンプレートを Qwen3-Max により **意味的に等価な 20 個のバリアント**に書き換え、その後、音声サンプルに無作為に割り当てて、プロンプトのテンプレートによる偏りを減らす。すべての三つ組は、論理的な一貫性について人手による品質確認を経る。

---

## 📊 データセットの統計

<p align="center">
  <img src="polybench_2.png" alt="PolyBench dataset statistics" width="61.8%"/>
</p>

<p align="center">
  <em>図 2. 左: 最上位の AudioSet のオントロジーのカテゴリごとの、イベントタグの分布(CE = Channel/environment/background、MU = Music、NS = Natural sounds、AS = Animal sounds、HS = Human sounds、ST = Sounds of things)。右: 重なっている音源の数(同時発生する 2、3、4 個のイベント)ごとのクリップの割合。</em>
</p>

### 音声の出典

| 出典 | # Clips | 備考 |
|---|---|---|
| **DataSED** | 169 | 音イベントのクラス 22 個。クリップの長さは 3〜89 秒。多声のセグメントだけ。 |
| **DESED** | 259 | 公式の評価用セット(home シナリオ)から、多声のクリップをサンプリングしたもの。 |
| **MAESTRO-Real** | 300 | 屋内と屋外の環境。長い録音は一律に 25 秒に切り出した。 |
| **AudioTime**(Concurrency のみ) | — | Concurrency のサブセットの均衡を取り、「常に Yes」という自明な偏りを避けるため、負例として加えた単声のクリップ。 |

すべてのクリップには **正確な正解の音イベントのクラスとタイムスタンプ**が付いている。これらは、人手で検証する前に、正解をプログラムで導き出すために使われた。

---

## 🚀 使い方

```python
from datasets import load_dataset

# Load a specific task config (counting | duration | concurrency | classification | detection)
ds = load_dataset("PolyBench/PolyBench", "counting")
print(ds["test"][0])
# {'audio': {'array': ..., 'sampling_rate': ...},
#  'sound_name': 'S-0001.wav',
#  'question': 'How many different sources can be found in the audio file?',
#  'choices': ['2', '4', '5', '3'],
#  'answer': '3',
#  'category': 'sound'}

# Load all configs
for config in ["counting", "duration", "concurrency", "classification", "detection"]:
    ds = load_dataset("PolyBench/PolyBench", config)
    print(f"{config}: {len(ds['test'])} examples")
```



## 📏 評価のプロトコル

- **形式:** MCQA。モデルは音声と文章の指示を受け取り、選択肢から 1 つを選ぶ。
- **推論モデル:** 最終的な選択の前に Chain-of-Thought を出力するよう指示される。
- **指標:** MMAU-Pro に従う正解率(Accuracy、ACC)と F1。
- **回答の照合:** モデルの応答と正解の選択肢の意味的な類似度の計算に **NV-Embed-v2** を使う。文字列の完全一致や正規表現による照合の脆さを避けるためである。
- **誤りの分析:** LLM を用いた判定器が各誤りを調べ、*知覚の誤解*か*推論の失敗*かのラベルを付ける。

---

## 🏆 ベースラインの結果

多声の MCQA の設定における、PolyBench 上の代表的な LALM の性能(%)(論文の Table 1 より):

| Model | Counting F1 | Counting ACC | Duration F1 | Duration ACC | Concurrency F1 | Concurrency ACC | Classification F1 | Classification ACC | Detection F1 | Detection ACC |
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-Omni-30B-A3B | 55.9 | **57.5** | **70.2** | **68.5** | 85.2 | 76.0 | **77.2** | **77.9** | **57.3** | **63.4** |
| R1-AQA | 34.5 | 30.1 | 42.1 | 43.2 | **93.7** | **90.4** | 29.5 | 33.1 | 22.4 | 26.2 |
| Audio Flamingo 3 | 56.3 | 53.4 | 67.4 | 65.8 | 53.1 | 37.7 | 47.8 | 50.3 | 36.2 | 37.2 |
| TimeAudio + Qwen3-8B | 47.7 | 43.8 | 63.7 | 63.7 | 86.9 | 78.1 | 53.8 | 50.3 | 53.5 | 51.7 |
| AUDSEMTHINKER-QA GRPO | **57.3** | 51.4 | 67.1 | 64.4 | 59.1 | 43.2 | 70.6 | 70.3 | 39.6 | 37.2 |


## 🛠 構築のパイプライン

このベンチマークは 3 つの段階で構築された(図 1 を参照):

1. **問題の観察。** 単声と多声の音声を比べることで、LALM が同時に存在する音源を幻覚したり混同したりすることが確かめられた。これが、重なりを対象とするベンチマークの動機となった。
2. **データと質問の準備。** 多声のクリップは実世界の録音から選別された。5 つのタスクの種類について、反復的な品質管理を伴う人と LLM の協働によって MCQA が生成された。Classification の誤答の選択肢(distractor)は、クリップに存在しない、意味的に無関係なイベントのカテゴリから選ばれた。意味的な関連性による近道を防ぐためである。
3. **指標と評価。** モデルは、選択肢の完全一致(NV-Embed-v2 による類似度を併用)で採点される。LLM の判定器は、任意で知覚の誤りと推論の誤りを診断する。

---

## 📚 引用

PolyBench を使う場合は、次を引用すること:

```bibtex
@article{chen2026polybench,
  title   = {PolyBench: A Benchmark for Compositional Reasoning in Polyphonic Audio},
  author  = {Chen, Yuanjian and Xiao, Yang and Yin, Han and Liu, Xubo and Huang, Jinjie and Dang, Ting},
  journal = {arXiv preprint arXiv:2603.05128},
  year    = {2026}
}
```

---

## 👥 著者

Yuanjian Chen¹, Yang Xiao², Han Yin³, Xubo Liu⁴, Jinjie Huang¹, Ting Dang²

¹ Harbin University of Science and Technology · ² The University of Melbourne · ³ KAIST · ⁴ University of Surrey

---

## 📬 連絡先

- 2010400002@stu.hrbust.edu.cn
- yxiao9550@student.unimelb.edu.au
- hanyin@kaist.ac.kr

<details>
<summary>原文(英語)</summary>

# PolyBench: A Benchmark for Compositional Reasoning in Polyphonic Audio

<p align="center">
  <img src="./polybench_1.png" alt="PolyBench pipeline overview" width="100%"/>
</p>

<p align="center">
  <em>Figure 1. PolyBench pipeline for benchmark construction and evaluation.</em>
</p>

<p align="center">
  📄 <a href="https://arxiv.org/abs/2603.05128">Paper</a> &nbsp;|&nbsp;
  💻 <a href="https://github.com/visionchan/PolyBench">GitHub</a>
</p>

---

## 📖 Overview

**PolyBench** is the first benchmark specifically designed to evaluate **compositional reasoning in polyphonic audio** for Large Audio Language Models (LALMs). While modern LALMs are capable of reasoning over monophonic audio, they consistently degrade when multiple sound events co-occur. PolyBench exposes this fundamental bottleneck through five carefully designed Multiple-Choice Question Answering (MCQA) subsets that require reasoning over concurrent events and their relationships.

Key features:

- **Real-world polyphonic audio** sampled from three established sound event datasets (DataSED, DESED, MAESTRO-Real), plus monophonic clips from AudioTime for the Concurrency subset.
- **Five hierarchical MCQA tasks** covering counting, classification, detection, concurrency, and duration estimation.
- **Human–LLM collaborative question generation** with iterative quality control to ensure logically coherent `<question, choices, answer>` triplets.
- **Standardized evaluation protocol** with Accuracy and F1 metrics, plus an LLM-based judge for fine-grained perception-vs-reasoning error analysis.

---

## 🧩 Tasks

PolyBench contains five MCQA subsets, organized by a concentric hierarchy of polyphonic reasoning dependencies:

| Task | Goal | Example Question |
|---|---|---|
| **Counting** | Identify the number of distinct sound event categories in the audio. | *"How many distinct sources does the audio file contain?"* |
| **Duration** | Identify the sound event with the longest duration in the audio. | *"Which sound event has the longest duration?"* |
| **Concurrency** | Decide whether polyphonic phenomena occur in the clip. | *"While the frying sound has not ceased, is there a simultaneous presence of another sound event?"* |
| **Classification** | Identify the category of another event that overlaps with a specified anchor event. | *"What type of sound event overlaps temporally with the dog sound event before it ends?"* |
| **Detection** | Reason about the temporal interval(s) where overlap first occurs and the total number of overlap occurrences. | *"At what moment in this audio clip do multiple sounds first coincide?"* |

For each task, a fundamental question template is rewritten into **20 semantically equivalent variants** by Qwen3-Max, then randomly assigned across audio samples to reduce prompt-template bias. All triplets go through manual quality checks for logical coherence.

---

## 📊 Dataset Statistics

<p align="center">
  <img src="polybench_2.png" alt="PolyBench dataset statistics" width="61.8%"/>
</p>

<p align="center">
  <em>Figure 2. Left: distribution of event tags across top-level AudioSet ontology categories (CE = Channel/environment/background, MU = Music, NS = Natural sounds, AS = Animal sounds, HS = Human sounds, ST = Sounds of things). Right: proportion of clips by number of overlapping sound sources (2, 3, and 4 concurrent events).</em>
</p>

### Audio sources

| Source | # Clips | Notes |
|---|---|---|
| **DataSED** | 169 | 22 sound event classes; clip duration 3–89 s; polyphonic segments only. |
| **DESED** | 259 | Sampled polyphonic clips from the official evaluation set (home scenario). |
| **MAESTRO-Real** | 300 | Indoor/outdoor environments; long recordings uniformly clipped to 25 s. |
| **AudioTime** (Concurrency only) | — | Monophonic clips added as negative samples to balance the Concurrency subset and avoid trivial "always Yes" bias. |

All clips come with **exact ground-truth event classes and timestamps**, which were used to programmatically derive correct answers before manual verification.

---

## 🚀 Usage

```python
from datasets import load_dataset

# Load a specific task config (counting | duration | concurrency | classification | detection)
ds = load_dataset("PolyBench/PolyBench", "counting")
print(ds["test"][0])
# {'audio': {'array': ..., 'sampling_rate': ...},
#  'sound_name': 'S-0001.wav',
#  'question': 'How many different sources can be found in the audio file?',
#  'choices': ['2', '4', '5', '3'],
#  'answer': '3',
#  'category': 'sound'}

# Load all configs
for config in ["counting", "duration", "concurrency", "classification", "detection"]:
    ds = load_dataset("PolyBench/PolyBench", config)
    print(f"{config}: {len(ds['test'])} examples")
```



## 📏 Evaluation Protocol

- **Format:** MCQA. The model receives the audio plus a textual instruction and selects one option.
- **Reasoning models:** prompted to emit Chain-of-Thought before the final choice.
- **Metrics:** Accuracy (ACC) and F1, following MMAU-Pro.
- **Answer matching:** **NV-Embed-v2** is used to compute semantic similarity between the model's response and the gold option, avoiding the brittleness of exact-string or regex matching.
- **Error analysis:** an LLM-based judge inspects each error and labels it as *perceptual misunderstanding* vs. *reasoning failure*.

---

## 🏆 Baseline Results

Performance (%) of representative LALMs on PolyBench under the polyphonic MCQA setting (from the paper, Table 1):

| Model | Counting F1 | Counting ACC | Duration F1 | Duration ACC | Concurrency F1 | Concurrency ACC | Classification F1 | Classification ACC | Detection F1 | Detection ACC |
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-Omni-30B-A3B | 55.9 | **57.5** | **70.2** | **68.5** | 85.2 | 76.0 | **77.2** | **77.9** | **57.3** | **63.4** |
| R1-AQA | 34.5 | 30.1 | 42.1 | 43.2 | **93.7** | **90.4** | 29.5 | 33.1 | 22.4 | 26.2 |
| Audio Flamingo 3 | 56.3 | 53.4 | 67.4 | 65.8 | 53.1 | 37.7 | 47.8 | 50.3 | 36.2 | 37.2 |
| TimeAudio + Qwen3-8B | 47.7 | 43.8 | 63.7 | 63.7 | 86.9 | 78.1 | 53.8 | 50.3 | 53.5 | 51.7 |
| AUDSEMTHINKER-QA GRPO | **57.3** | 51.4 | 67.1 | 64.4 | 59.1 | 43.2 | 70.6 | 70.3 | 39.6 | 37.2 |


## 🛠 Construction Pipeline

The benchmark was built in three stages (see Figure 1):

1. **Problem observation.** Contrasting monophonic and polyphonic audio confirmed that LALMs hallucinate or confuse concurrent sources, motivating a benchmark targeting overlap.
2. **Data and question preparation.** Polyphonic clips were curated from real-world recordings; MCQAs were generated for the five task types via human–LLM collaboration with iterative quality control. Distractor options for Classification were drawn from semantically unrelated event categories not present in the clip, to prevent semantic-relevance shortcuts.
3. **Metric and evaluation.** Models are scored by exact-option match (with NV-Embed-v2 similarity); an LLM judge optionally diagnoses perception vs. reasoning errors.

---

## 📚 Citation

If you use PolyBench, please cite:

```bibtex
@article{chen2026polybench,
  title   = {PolyBench: A Benchmark for Compositional Reasoning in Polyphonic Audio},
  author  = {Chen, Yuanjian and Xiao, Yang and Yin, Han and Liu, Xubo and Huang, Jinjie and Dang, Ting},
  journal = {arXiv preprint arXiv:2603.05128},
  year    = {2026}
}
```

---

## 👥 Authors

Yuanjian Chen¹, Yang Xiao², Han Yin³, Xubo Liu⁴, Jinjie Huang¹, Ting Dang²

¹ Harbin University of Science and Technology · ² The University of Melbourne · ³ KAIST · ⁴ University of Surrey

---

## 📬 Contact

- 2010400002@stu.hrbust.edu.cn
- yxiao9550@student.unimelb.edu.au
- hanyin@kaist.ac.kr

</details>