Taming Hallucinations: Boosting MLLMs' Video Understanding via Counterfactual Video Generation

TL;DR

This work tackles visual hallucinations in Multimodal Large Language Models by addressing the imbalance between text and video data. It introduces DualityForge, a diffusion-based controllable video editing framework with embedded counterfactual context to synthesize high-quality QA pairs and paired real/edited videos, culminating in the DualityVidQA dataset. To exploit the contrastive nature of the data, the authors propose DNA-Train, a two-stage SFT-RL regime with $\ell_1$-normalized advantages that stabilizes training and enforces visual grounding. Experiments show substantial reductions in hallucinations and improved general video understanding, with a reported 24.0% relative gain on DualityVidQA-Test and strong generalization across benchmarks. The work includes open-sourcing the dataset and code, demonstrating a practical path to scalable, grounded video understanding for MLLMs.

Abstract

Multimodal Large Language Models (MLLMs) have made remarkable progress in video understanding. However, they suffer from a critical vulnerability: an over-reliance on language priors, which can lead to visual ungrounded hallucinations, especially when processing counterfactual videos that defy common sense. This limitation, stemming from the intrinsic data imbalance between text and video, is challenging to address due to the substantial cost of collecting and annotating counterfactual data. To address this, we introduce DualityForge, a novel counterfactual data synthesis framework that employs controllable, diffusion-based video editing to transform real-world videos into counterfactual scenarios. By embedding structured contextual information into the video editing and QA generation processes, the framework automatically produces high-quality QA pairs together with original-edited video pairs for contrastive training. Based on this, we build DualityVidQA, a large-scale video dataset designed to reduce MLLM hallucinations. In addition, to fully exploit the contrastive nature of our paired data, we propose Duality-Normalized Advantage Training (DNA-Train), a two-stage SFT-RL training regime where the RL phase applies pair-wise $\ell_1$ advantage normalization, thereby enabling a more stable and efficient policy optimization. Experiments on DualityVidQA-Test demonstrate that our method substantially reduces model hallucinations on counterfactual videos, yielding a relative improvement of 24.0% over the Qwen2.5-VL-7B baseline. Moreover, our approach achieves significant gains across both hallucination and general-purpose benchmarks, indicating strong generalization capability. We will open-source our dataset and code.

Figures (11)

• Figure 1: Comparison of MLLMs on counterfactual video QA. Each example shows key frames, the multiple‑choice question, and the models’ reasoning and answers. General models (Qwen‑VL, GPT‑4o) rely on language priors, picking plausible but visually incorrect answer. Our DNA‑Train-7B model grounds reasoning in the video and selects the counterintuitive yet correct answer.
• Figure 2: Overview of video editing pipelines. There are three pipelines for different types of counterfactual context: Visual Anomaly: pixel-level video editing via OpenCV Semantic Anomaly: an MLLM selects an object for editing, followed by mask generation, VACE-based editing, and majority-vote verification using multiple SOTA MLLMs. Common Sense Anomaly: an MLLM propose commonsense violations, FLUX-Kontext edits frames, edits are re-verified by multiple MLLMs, and VACE interpolates the final video.
• Figure 3: Overview of the DualityForge framework and DualityVidQA dataset. Starting with real, web-sourced videos, the DualityForge framework first embeds the counterfactual (CF) context, including visual, semantic, and commonsense, into it with video editing pipeline. The embedded context is then provided alongside the video to an MLLM to produce detailed captions and QA pairs. The dataset comprises three splits: DualityVidQA-SFT with real and counterfactual video-QA pairs (54K + 50K) for SFT; DualityVidQA-RL with 20K shared-question contrastive video-answer pairs (one question, two real/CF videos) for RL; and DualityVidQA-Test (600 pairs), which shares the same contrastive structure as DualityVidQA-RL and covers diverse counterfactual categories.
• Figure 4: Overview of DNA-Train framework. We first perform SFT on our dual dataset to initialize the model. During RL, we sample a group of responses for both real and CF videos, compute their rewards based on task correctness, and calculate the $\ell_1$ norm of intra-group advantages. Finally, we normalize the advantages across the dual groups to ensure balanced gradients.
• Figure A.1: Dense Caption Prompt Template