LongVPO: From Anchored Cues to Self-Reasoning for Long-Form Video Preference Optimization

TL;DR

LongVPO tackles ultra-long video understanding with no long-video annotations by a two-stage Direct Preference Optimization framework. Stage 1 creates anchor-centered triples from short clips, using an anchor-only likelihood approximation to train a model with mixed short-long context and disambiguated supervision, while Stage 2 applies recursive captioning and LLM-guided multi-segment reasoning to align preferences over real long videos through self-generated queries and dispreferred responses, all within the DPO objective. Trained on only about 16K synthetic examples, LongVPO achieves state-of-the-art results on multiple long-video benchmarks and maintains competitive short-video performance, demonstrating a scalable path for efficient long-form video understanding. The approach leverages anchored cues, scene-level metadata, and staged optimization to bridge short-to-long contexts without expensive annotations, indicating strong practical potential for scalable multimodal video analysis.

Abstract

We present LongVPO, a novel two-stage Direct Preference Optimization framework that enables short-context vision-language models to robustly understand ultra-long videos without any long-video annotations. In Stage 1, we synthesize preference triples by anchoring questions to individual short clips, interleaving them with distractors, and applying visual-similarity and question-specificity filtering to mitigate positional bias and ensure unambiguous supervision. We also approximate the reference model's scoring over long contexts by evaluating only the anchor clip, reducing computational overhead. In Stage 2, we employ a recursive captioning pipeline on long videos to generate scene-level metadata, then use a large language model to craft multi-segment reasoning queries and dispreferred responses, aligning the model's preferences through multi-segment reasoning tasks. With only 16K synthetic examples and no costly human labels, LongVPO outperforms the state-of-the-art open-source models on multiple long-video benchmarks, while maintaining strong short-video performance (e.g., on MVBench), offering a scalable paradigm for efficient long-form video understanding.

Figures (13)

• Figure 1: Context Position Bias Probing.Left: A short video segment (visualized as a $4 \times 4$ grid) is embedded within a much longer padded sequence and processed chronologically. Right: Performance is plotted against each frame’s L1 distance from the question token. The middle-position drop indicates a strong positional bias (“lost-in-the-middle”). The Upper Bound shows performance without padding, revealing degradation under long-context settings.
• Figure 2: Comparison of prior methods with our proposed two-stage method.
• Figure 3: Overview of our two-stage training framework. Stage 1: Short clips are concatenated to form a pseudo-long video. The target model generates the query ($q_i$) and preferred response ($y_i^+$) conditioned on the anchor clip and its caption, while dispreferred responses ($y_i^-$) are generated by prompting the model to answer based on non-anchor (distractor) clips, simulating temporal misalignment errors. Stage 2: For unlabeled long videos, an LLM generates the query ($q_i$) and reasoning trace ($r_i$) based on scene synthetic captions. The target MLLM then generates the preferred response ($y_i^+$) based on corresponding scene context, and generates dispreferred responses ($y_i^-$) under degraded context (e.g., partial or irrelevant scenes).
• Figure 4: The V-NIAH results of our baseline InternVL2.5-8B and LongVPO. "Frame Depth" indicates the position where the needle image is located, ranging from 0% to 100% (from the beginning to the end of the video).
• Figure 5: Comparison of Stage 1 training using SFT and DPO. Additional results are provided in the appendix.