Yuxin Ye

Yinghao Ma, Haiwen Xia, Hewei Gao et al.

While music generation models have evolved to handle complex multimodal inputs mixing text, lyrics, and reference audio, evaluation mechanisms have lagged behind. In this paper, we bridge this critical gap by establishing a comprehensive ecosystem for music reward modeling under Compositional Multimodal Instruction (CMI), where the generated music may be conditioned on text descriptions, lyrics, and audio prompts. We first introduce CMI-Pref-Pseudo, a large-scale preference dataset comprising 110k pseudo-labeled samples, and CMI-Pref, a high-quality, human-annotated corpus tailored for fine-grained alignment tasks. To unify the evaluation landscape, we propose CMI-RewardBench, a unified benchmark that evaluates music reward models on heterogeneous samples across musicality, text-music alignment, and compositional instruction alignment. Leveraging these resources, we develop CMI reward models (CMI-RMs), a parameter-efficient reward model family capable of processing heterogeneous inputs. We evaluate their correlation with human judgments scores on musicality and alignment on CMI-Pref along with previous datasets. Further experiments demonstrate that CMI-RM not only correlates strongly with human judgments, but also enables effective inference-time scaling via top-k filtering. The necessary training data, benchmarks, and reward models are publicly available.

Yuxin Ye, Yan Liu, Shujian Yu

Continual learning in medical image segmentation involves sequential data acquisition across diverse domains (e.g., clinical sites), where task interference between past and current domains often leads to catastrophic forgetting. Existing continual learning methods fail to capture the complex dependencies between tasks. We introduce a novel framework that mitigates forgetting by establishing and enhancing complex dependencies between historical data and the network in the present task. Our framework features a dual-alignment strategy, the cross-network alignment (CNA) module aligns the features extracted from the bottleneck layers of the current and previous networks, respectively, while the cross-representation alignment (CRA) module aligns the features learned by the current network from historical buffered data and current input data, respectively. Implementing both types of alignment is a non-trivial task. To address this, we further analyze the linear and nonlinear forms of the well-established Hilbert-Schmidt Independence Criterion (HSIC) and deliberately design feature mapping and feature pairing blocks within the CRA module. Experiments on medical image segmentation task demonstrate our framework's effectiveness in mitigating catastrophic forgetting under domain shifts.