AWhat changed
Developed hierarchical task-tag conditioning for heterogeneous audio-language pretraining; the later Qwen2 stage replaced tags with natural-language task prompts and added audio-instruction preference adaptation.
Scientific scope
Developed hierarchical task-tag conditioning for heterogeneous audio-language pretraining; the later Qwen2 stage replaced tags with natural-language task prompts and added audio-instruction preference adaptation. Inherits audio encoders, language models, multitask learning and DPO. One family package with explicit two-stage deltas, not two model-size entries. Qwen2-Audio outputs text responses; it is not the later Thinker/Talker speech-generation architecture. No universal audio understanding guarantee.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits audio encoders, language models, multitask learning and DPO. One family package with explicit two-stage deltas, not two model-size entries. Qwen2-Audio outputs text responses; it is not the later Thinker/Talker speech-generation architecture. No universal audio understanding guarantee.
Contributors
- Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.
- Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models
- Qwen2-Audio Technical Report
Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction