Yixiao He, Menghao Zhang, Haifeng Sun, Jing Wang, Kangheng Lin, Jinghan Wang, Chenye Xu, Pengfei Ren, Qi Qi, Jingyu Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Online video understanding is pivotal for emerging video streaming services. However, existing Multimodal Large Language Models (MLLMs) encounter significant challenges in this domain, specifically in maintaining holistic visual perception under strict token budgets and attaining grounded semantic reasoning amidst dynamic context changes. To mitigate these issues, we propose OV-DPO, a systematic multimodal preference optimization approach. To ensure holistic visual perception, we introduce a visual preference objective that compels the model to strictly ground its reasoning in clear visual evidence. This is supported by the QGDFR strategy, which dynamically reallocates resolution budgets to preserve critical visual details for chosen samples while intentionally degrading them for rejected ones to construct valid visual contrasts. Concurrently, to promote grounded semantic reasoning, we employ a textual preference objective. We design the SROVA framework to construct high-quality preference pairs, utilizing self-refinement to generate factual chosen responses and visual degradation to induce language-prior-driven hallucinations as rejected samples. By jointly optimizing these objectives alongside anchored and supervised fine-tuning terms, OV-DPO ensures training stability and robust alignment. Experimental results demonstrate that our approach significantly outperforms baselines such as SFT and DPO. Notably, utilizing only 29K high-quality samples, our 3B-scale model achieves results competitive with significantly larger-scale models on both online and offline benchmarks, offering an efficient solution for online video understanding.

Large Vision-Language Models (LVLMs) suffer from severe object hallucinations, leading them to frequently generate outputs that do not correspond to the image content, significantly reducing the credibility and reliability of their responses. Recent research has attempted to enhance LVLMs by employing Direct Preference Optimization (DPO) to reduce hallucinations and improve response quality. However, these approaches typically utilize text-only preference response pairs, neglecting the influence of visual input in optimizing LVLMs. In this paper, we propose VA-DPO, a multimodal optimization objective. VA-DPO leverages the LVLMs' attention to select and corrupt critical parts of images, thereby constructing visual preference image pairs. This approach integrates both text and visual preference optimization objectives to achieve effective alignment optimization of LVLMs. We conduct extensive experiments on LVLMs of different sizes, and the results demonstrate that VA-DPO effectively reduces hallucinations across various tasks. Compared to other hallucination mitigation approaches, VA-DPO achieves more competitive results.

Yixiao He, Haifeng Sun, Pengfei Ren, Jingyu Wang, Huazheng Wang, Qi Qi, Zirui Zhuang, Jing Wang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Current Vision-Language Models (VLMs) struggle with fine-grained spatial reasoning, particularly when multi-step logic and precise spatial alignment are required. In this work, we introduce SpatialReasoner-R1, a vision-language reasoning model designed to address these limitations. To construct high-quality supervision for spatial reasoning, we design a Multi-Model Monte Carlo Tree Search (M3CTS) method that generates diverse, logically consistent Long Chain-of-Thought (LongCoT) reasoning trajectories. In addition, we propose a fine-grained Direct Preference Optimization (fDPO) method that introduces segment-specific preference granularity for descriptive grounding and logical reasoning, guided by a spatial reward mechanism that evaluates candidate responses based on visual consistency, spatial grounding, and logical coherence. Experimental results demonstrate that fDPO achieves relative performance gains of 4.1% and 9.0% over standard DPO on spatial qualitative and quantitative tasks, respectively. SpatialReasoner-R1, trained with fDPO, sets a new SoTA on SpatialRGPT-Bench, outperforming the strongest baseline by 9.4% in average accuracy, while maintaining competitive performance on general vision-language tasks.