Video-language models (VLMs) achieve strong multimodal understanding but remain prone to hallucinations, especially when reasoning about actions and temporal order. Existing mitigation strategies, such as textual filtering or random video perturbations, often fail to address the root cause: over-reliance on language priors rather than fine-grained visual dynamics. We propose a scalable framework for counterfactual video generation that synthesizes videos differing only in actions or temporal structure while preserving scene context. Our pipeline combines multimodal LLMs for action proposal and editing guidance with diffusion-based image and video models to generate semantic hard negatives at scale. Using this framework, we build CounterVid, a synthetic dataset of ~26k preference pairs constructed from short counterfactual action clips and targeting both action recognition and controlled action-sequence ordering. We further introduce MixDPO, a unified Direct Preference Optimization approach that jointly leverages textual and visual preferences. Across Qwen2.5-VL and InternVL3 backbones, MixDPO substantially improves action recognition and temporal ordering, yields gains on most standard video hallucination benchmarks, and largely preserves general video understanding. Our source code, trained models, and dataset are available at https://aimagelab.github.io/CounterVid.

Large language models (LLMs) have been deployed in a wide spectrum of domains and applications due to their strong language understanding capabilities obtained through pretraining. However, their performance on specific domain is usually suboptimal due to limited exposure to domain-specific tasks. Adapting LLM to the cinematic domain post unique challenges as it consists of complicated stories with limited textual information accessible from the subtitle or script alone. In this paper, we decompose the movie understanding capability into a suite of narrative understanding tasks based on narrative theory. We construct a dataset for these tasks based on resources in the movie domain, and use it to examine the effect of different domain adaptation strategies. Both the dataset and the models are made publicly available.Our experiment results show the effectiveness of our approach in improving the narrative understanding of LLMs and highlight the trade-offs between domain-specific and general instruction capabilities.