MuseReason: From Musical Evidence to Counterfactual Editing via Bidirectional Affective Reasoning

Music emotion understanding and emotion conditioned music generation have made significant progress in recent years, but they are usually modelled as relatively independent tasks. Existing music emotion understanding methods mainly focus on emotion state prediction and lack an explicit characterisation of the musical evidence supporting emotional judgements. Emotion conditioned music generation methods typically take the target emotion directly as a condition, with less reasoning about which musical attributes should be adjusted based on the source musical state and which constraints should be maintained. This paper proposes MuseReason, a bidirectional emotion reasoning framework for music emotion understanding and counterfactual editing. MuseReason uses temporal music representations to extract verifiable musical evidence. It then combines the musical evidence with the target affect to predict structured music adjustments and preservation constraints for counterfactual editing. The experimental results show that MuseReason demonstrates capabilities in emotional understanding, structured output, and temporal emotional modelling. Structured music adjustment can support counterfactual editing by considering emotional changes while maintaining the source music. MuseReason extends music emotion modelling from a single emotional prediction to a unified framework that connects musical evidence, emotional reasoning, structured music adjustment, and generated result verification.