Audio Visual Question Answering on Music-AVQA (Robustness Evaluation) benchmark leaderboard · SOTA2 Research

Audio Visual Question Answering on Music-AVQA (Robustness Evaluation)

85.15Music-AVQA Clean AccuracyQwen3-Omni-30B

Evaluation Results

Method
Qwen3-Omni-30B2026.04	85.15	83.23	—	9.58	15.16
Qwen2.5-Omni-7B2026.04	82.99	80.91	—	18.83	18.6
Gemini-3.1-Flash-Lite-preview2026.04	81.32	80.01	—	10.54	16.12
Negative Language Modeling Loss2026.01	80.7	74.3	6.4	—	—
Encoder-Based Cosine Similarity Loss2026.01	80.7	74.7	6	—	—
Vision Attention Suppression Loss2026.01	80.7	77.7	3	—	—
Audio Attention Amplification Loss2026.01	80.7	76.9	3.8	—	—
Attention Randomization Loss2026.01	80.7	75.8	4.9	—	—
Hidden-State Similarity Loss2026.01	80.7	76.6	4.1	—	—
Combined Loss (SOUNDBREAK)2026.01	80.7	75	5.7	—	—
Gemini-2.5-Flash-Lite2026.04	80.68	75.4	—	11.75	19.68
Gemini-3.1-Flash-Lite-preview2026.04	71.97	70.62	—	2.14	6.84
Gemini-2.5-Flash-Lite2026.04	68.99	67.24	—	2.01	4.52
Qwen2.5-Omni-7B2026.04	66.94	56.18	—	4.34	15.51
PandaGPT2026.04	64.41	64.46	—	26.73	26.96
Gemini-3.1-Flash-Lite-preview2026.04	62.63	49.96	—	16.83	34.21
Qwen3-Omni-30B2026.04	61.54	55.09	—	2.15	8.11
Gemini-2.5-Flash-Lite2026.04	60.15	47.49	—	17.99	33.26
Qwen2.5-Omni-7B2026.04	57.01	43.76	—	22.2	38.62
Qwen3-Omni-30B2026.04	56.57	53.96	—	18.48	33.33
ChatBridge2026.04	51.38	50	—	27.47	30.09
ChatBridge2026.04	39.99	34.38	—	15.83	25.55
ChatBridge2026.04	37.64	35.21	—	23.68	24.53
PandaGPT2026.04	35.98	35.93	—	10.04	10.98
PandaGPT2026.04	34.93	34.93	—	29.02	29.12