Towards Mitigating Modality Bias in Vision-Language Models for Temporal Action Localization

Temporal Action Localization (TAL) requires identifying both the boundaries and categories of actions in untrimmed videos. While vision-language models (VLMs) offer rich semantics to complement visual evidence, existing approaches tend to overemphasize linguistic priors at the expense of visual performance, leading to a pronounced modality bias. We propose ActionVLM, a vision-language aggregation framework that systematically mitigates modality bias in TAL. Our key insight is to preserve vision as the dominant signal while adaptively exploiting language only when beneficial. To this end, we introduce (i) a debiasing reweighting module that estimates the language advantage-the incremental benefit of language over vision-only predictions-and dynamically reweights language modality accordingly, and (ii) a residual aggregation strategy that treats language as a complementary refinement rather than the primary driver. This combination alleviates modality bias, reduces overconfidence from linguistic priors, and strengthens temporal reasoning. Experiments on THUMOS14 show that our model outperforms state-of-the-art by up to 3.2% mAP.

Hybrid Driven Learning for Channel Estimation in Intelligent Reflecting Surface Aided Millimeter Wave Communications

Intelligent reflecting surfaces (IRS) have been proposed in millimeter wave (mmWave) and terahertz (THz) systems to achieve both coverage and capacity enhancement, where the design of hybrid precoders, combiners, and the IRS typically relies on channel state information. In this paper, we address the problem of uplink wideband channel estimation for IRS aided multiuser multiple-input single-output (MISO) systems with hybrid architectures. Combining the structure of model driven and data driven deep learning approaches, a hybrid driven learning architecture is devised for joint estimation and learning the properties of the channels. For a passive IRS aided system, we propose a residual learned approximate message passing as a model driven network. A denoising and attention network in the data driven network is used to jointly learn spatial and frequency features. Furthermore, we design a flexible hybrid driven network in a hybrid passive and active IRS aided system. Specifically, the depthwise separable convolution is applied to the data driven network, leading to less network complexity and fewer parameters at the IRS side. Numerical results indicate that in both systems, the proposed hybrid driven channel estimation methods significantly outperform existing deep learning-based schemes and effectively reduce the pilot overhead by about 60% in IRS aided systems.