Mahi Luthra, Jiayi Shen, Maxime Poli, Angelo Ortiz Tandazo, Yosuke Higuchi, Youssef Benchekroun, Martin Gleize, Charles-Éric Saint-James, Dongyan Lin, Phillip Rust, Angel Villar-Corrales, Surya, Vanessa Stark, Rashel Moritz, Juan Pino, Yann LeCun, Emmanuel Dupoux. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Text-to-audio generation (TTA) produces audio from a text description, learning from pairs of audio samples and hand-annotated text. However, commercializing audio generation is challenging as user-input prompts are often under-specified when compared to text descriptions used to train TTA models. In this work, we treat TTA models as a "blackbox" and address the user prompt challenge with two key insights: (1) User prompts are generally under-specified, leading to a large alignment gap between user prompts and training prompts. (2) There is a distribution of audio descriptions for which TTA models are better at generating higher quality audio, which we refer to as "audionese". To this end, we rewrite prompts with instruction-tuned models and propose utilizing text-audio alignment as feedback signals via margin ranking learning for audio improvements. On both objective and subjective human evaluations, we observed marked improvements in both text-audio alignment and music audio quality.

We introduce MusicFlow, a cascaded text-to-music generation model based on flow matching. Based on self-supervised representations to bridge between text descriptions and music audios, we construct two flow matching networks to model the conditional distribution of semantic and acoustic features. Additionally, we leverage masked prediction as the training objective, enabling the model to generalize to other tasks such as music infilling and continuation in a zero-shot manner. Experiments on MusicCaps reveal that the music generated by MusicFlow exhibits superior quality and text coherence despite being over $2\sim5$ times smaller and requiring $5$ times fewer iterative steps. Simultaneously, the model can perform other music generation tasks and achieves competitive performance in music infilling and continuation.

Social search has stably evolved across various species and is often used by humans to search for resources (such as food, information, social partners). In turn, these resources frequently come distributed in patches or clusters. In the current work, we use an ecologically inspired agent-based model to investigate whether social search and clustering are stable outcomes of the dynamical mutual interactions between the two. While previous research has studied unidirectional influences of social search on resource clustering and vice versa, the current work investigates the consequential patterns emerging from their two-way interactions over time. In our model, consumers evolved search strategies (ranging from competitive to social) as adaptations to their environmental resource structures, and resources varied in distributions (ranging from random to clustered) that were shaped by agents' consumption patterns. Across four experiments, we systematically analyzed the patterns of influence that search strategies and environment structure have on each other to identify stable attractor states of both. In Experiment 1, we fixed resource clustering at various levels and observed its influence on social search, and in Experiment 2, we observed the influence of social search on resource distribution. In both these experiments we found that increasing levels of one variable produced increases in the other; however, at very high levels of the manipulated variable, the dependent variable tended to fall. Finally in Experiments 3 and 4, we studied the dynamics that arose when resource clustering and social search could both change and mutually influence each other, finding that low levels of social search and clustering were stable attractor states. Our simple 2D model yielded results that qualitatively resemble those across a wide range of search domains (from physical search for food to abstract search for information), highlighting some stable outcomes of mutually interacting consumer/resource systems.