Abstract:Current multimodal models often suffer from shallow reasoning, leading to errors caused by incomplete or inconsistent thought processes. To address this limitation, we propose Self-Verification and Self-Rectification (SVSR), a unified framework that explicitly integrates self-verification and self-rectification into the model's reasoning pipeline, substantially improving robustness and reliability in complex visual understanding and multimodal reasoning tasks. SVSR is built on a novel three-stage training paradigm. First, we construct a high-quality unified preference dataset by refining reasoning traces from pre-trained vision-language models, incorporating both forward and backward reasoning to embed self-reflective signals. Second, we perform cold-start supervised fine-tuning on this dataset to learn structured, multi-step reasoning behaviors. Third, we apply a Semi-online Direct Preference Optimization (Semi-online DPO) process, continuously augmenting the training corpus with high-quality, model-generated reasoning traces filtered by a powerful teacher VLM. This pipeline enables the model to learn, elicit, and refine its ability to self-verify and self-rectify. Extensive experiments across diverse benchmarks demonstrate that SVSR improves reasoning accuracy and enables stronger generalization to unseen tasks and question types. Notably, once trained with explicit self-reflective reasoning, the model also exhibits improved implicit reasoning ability, outperforming strong baselines even when no explicit reasoning traces are provided. These results highlight the potential of SVSR for building more dependable, introspective, and cognitively aligned multimodal systems.
Abstract:As an emerging wireless communication technology, movable antennas (MAs) offer the ability to adjust the spatial correlation of steering vectors, enabling more flexible beamforming compared to fixed-position antennas (FPAs). In this paper, we investigate the use of MAs for two typical near-field beamforming scenarios: beam nulling and multi-beam forming. In the first scenario, we aim to jointly optimize the positions of multiple MAs and the beamforming vector to maximize the beam gain toward a desired direction while nulling interference toward multiple undesired directions. In the second scenario, the objective is to maximize the minimum beam gain among all the above directions. However, both problems are non-convex and challenging to solve optimally. To gain insights, we first analyze several special cases and show that, with proper positioning of the MAs, directing the beam toward a specific direction can lead to nulls or full gains in other directions in the two scenarios, respectively. For the general cases, we propose a discrete sampling method and an alternating optimization algorithm to obtain high-quality suboptimal solutions to the two formulated problems. Furthermore, considering the practical limitations in antenna positioning accuracy, we analyze the impact of position errors on the performance of the optimized beamforming and MA positions, by introducing a Taylor series approximation for the near-field beam gain at each target. Numerical results validate our theoretical findings and demonstrate the effectiveness of our proposed algorithms.