VideoPASTA

VideoPASTA is a DPO-based framework that aligns video-language models using structured preference optimization. It leverages a dataset D = {(V, q, r⁺, r⁻)}, where V is the video, q is the query, r⁺ is the preferred aligned response, and r⁻ is a targeted adversarial response designed to introduce misalignment.

The core process involves generating preference pairs targeting three key failure modes:

1. Spatial Misalignment Generation: Uses InternVL2.5-38B to generate spatial queries (e.g., occlusion, depth, relative position). Preferred responses are generated from the baseline model using dense (32fps) sampling. Adversarial responses are generated using sparse (1fps) sampling and prompts that induce spatial errors (e.g., claiming occluded objects are visible, or all objects are equidistant).
2. Temporal Incoherence Generation: Generates temporal queries (e.g., event order, transitions, causality). Preferred responses use native frame rate sampling. Adversarial responses use sparse (1fps) sampling and prompts that distort time (e.g., describing sequential actions as simultaneous, ignoring transitions).
3. Cross-frame Disconnection Generation: Generates queries about long-range continuity (e.g., object/character persistence, setting changes). Preferred responses use uniform sampling across the video. Adversarial responses use sparse (1fps) sampling and prompts that break continuity (e.g., treating the same object/character in different scenes as unrelated).
4. Preference Data Filtering: Generated adversarial examples (3 per preferred response) are filtered using Qwen2.5-32B to ensure they introduce genuine, targeted misalignments and are distinct from the preferred responses. Preferred responses are also sanity-checked. This yields ~7k high-quality pairs.
5. Optimization: The model is trained using Direct Preference Optimization (DPO) on the filtered pairs. The DPO loss is computed separately for spatial, temporal, and cross-frame subsets and combined with weights.

This targeted approach enables robust alignment across multiple dimensions of video understanding.

VideoPASTA demonstrates significant improvements over its foundation model (Qwen2.5-VL) and outperforms state-of-the-art methods on several benchmarks. Key relative gains include +3.05% on VideoMME, +1.97% on NeXTQA, +1.69% on MVBench, and +1.31% on Long VideoBench. It achieves these results with only 7k preference pairs, contrasting with methods like Hound-DPO (17k pairs) and TPO (10k pairs). VideoPASTA outperforms LLaVA-Hound-DPO and i-SRT on all eight benchmarks and LLaVA-Video-TPO on seven.

Further analysis validates the effectiveness of VideoPASTA's components including training dynamics, the necessity of targeted failure modes, and robustness against adversarial attacks.

Training Dynamics

DPO training effectively learns preference boundaries with stabilizing reward accuracy.

Ablation of Modes

Combining all three targeted modes yields the best overall performance improvement.

Adversarial QA Performance

VideoPASTA significantly outperforms baselines in rejecting adversarial options across all categories.

Qualitative examples demonstrate VideoPASTA's improved understanding compared to the baseline Qwen2.5-VL. It correctly identifies spatial relationships (e.g., person clinging to aircraft vs. leaving), accurately captures temporal sequences in actions (e.g., cake decoration steps), and successfully connects narrative elements across frames (e.g., relating a disclaimer to illustrations of heat stress).