arXiv:2605.26513v2 Announce Type: replace
Abstract: Even balanced multimodal learning methods do not consistently translate additional modalities into better regression performance. To understand thi...
By Haojie Yin, Chengcheng Feng, Tianyi Liu, Tianqi Zhang, Kaizhu Huang
The paper introduces CAT‑GS, a training controller that stabilizes multimodal neural networks by addressing three failure modes: modality imbalance, unstable gating, and fusion interference. CAT‑GS calibrates teacher-derived reliability, applies a margin‑thresholded gating policy, caps gradient budgets, and uses fusion‑only PCGrad, all without altering model architectures or losses. Experiments on audio‑visual, tri‑modal, synthetic, and cross‑domain benchmarks show that CAT‑GS matches or surpasses strong imbalance‑aware baselines while producing smoother gating and fewer conflicting fusion gradients.
By Mahir Shahriar Tamim, Sharjil Khan, Md. Samiul Alim, Tanvir Ahmed Khan, Shafin Rahman, Nabeel Mohammed
VCMM: Variance-Calibrated Momentum for Multimodal Learning proposes a new optimizer that adapts momentum based on modality-specific gradient dynamics. It estimates minibatch noise and temporal drift online, using a Kalman-inspired controller to set modality-specific momentum and applies bias correction for the first moment. Experiments on four multimodal benchmarks show consistent improvements with modest training overhead.
By Zhongjing Gu, Chenyang Huang, Yufa Feng, Chong He, Qinxu Ding, Yiming Cui
The paper identifies that in multimodal learning, optimization often produces asymmetric certainty gains, with the stronger modality becoming more confident than the weaker one, which leads to imbalanced contributions and suboptimal performance. The authors attribute this issue to unimodal characteristics and propose a Max Confidence Regularization (MaxCR) method that tracks each modality’s semantic confidence via a nonlinear sparsity measure and applies max suppression and excitation to balance confidence levels. Experiments on standard datasets demonstrate that MaxCR improves overall performance compared to state‑of‑the‑art multimodal baselines.
By Longfei Huang, Xiangyu Wu, Yang Yang
arXiv:2609.16059v1 Announce Type: cross
Abstract: Multimodal instruction following (MMIF) is crucial for building generalist agents. However, current training paradigms rely heavily on Supervised Fin...
By Yirong Zeng, Zhang Sai, Yuxian Wang, Yutai Hou, Yufei Liu, Xiao Ding, Bibo Cai
Cross-modal alignment (CA) and cross-modal prediction (CP) are the dominant paradigms for multimodal representation learning, yet there is no systematic understanding of when each succeeds, when each fails, and when cross-modal training helps at all -- a gap that leaves practitioners, especially in scientific domains like biomedicine or astrophysics, with heterogeneous instruments and multiple levels of organization and measurement, unable to diagnose why standard methods underperform the best single modality. We develop a unified linear framework that addresses both questions.