Source-linked AI summary
LongVPO: From Anchored Cues to Self-Reasoning for Long-Form Video Preference Optimization
Zhenpeng Huang, Jiaqi Li, Zihan Jia, Xinhao Li, Desen Meng, Lingxue Song, Xi Chen, Liang Li, Limin Wang
TL;DR
Long-video reasoning remains difficult for short-context VLMs, while long-video annotations and suitable reference models are costly or unavailable. LongVPO addresses this with two-stage DPO using anchored short-clip preferences and self-training from recursively captioned long videos. Using only about 16K synthetic examples and no costly human labels, it achieves state-of-the-art results on long- and short-video benchmarks, while leaving inference efficiency for future work.
Problem
Short-context VLMs degrade on long-range video reasoning, and extending them is constrained by scarce long-video annotations and reference-model limitations.
Method
LongVPO uses two-stage DPO: anchored short-clip preferences with distractors in Stage 1, followed by self-training on scene-caption-based long-video reasoning in Stage 2.
Results
LongVPO achieves state-of-the-art performance on long- and short-video understanding benchmarks using about 16K synthetic samples without annotated long-video data.
Takeaways & Limitations
The results support LongVPO as a general-purpose approach for extending short-context VLMs to long-video understanding without costly long-video supervision.
Takeaways & Limitations
The work prioritizes performance improvement over inference computational efficiency and leaves integration with context-compression methods for future research.
Abstract
from arXiv · showhide
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.
1 Introduction
LongVPO addresses the difficulty of extending short-context VLMs to ultra-long videos without costly long-video annotations. Its two-stage preference-optimization framework uses anchored short-clip supervision and self-training to improve long-video understanding while preserving short-video performance.
- Motivation: Long-context visual reasoning remains challenging because VLM performance degrades on videos spanning over an hour.The paper identifies this degradation as a significant obstacle to scaling VLMs for long-form video understanding.
- Motivation: Short-context VLMs may be extended to long videos without expensive retraining or long-video labels, despite limited long-range visual inputs.This direction is motivated by their surprisingly competitive long-video benchmark results and strong foundational vision-language alignment.
- Method: LongVPO uses two training stages: anchored, interleaved short clips with distractor responses, followed by self-training for longer multi-segment reasoning.Stage 1 randomizes target positions and filters anchor questions for clip specificity; Stage 2 builds on retrieval skills using synthetic long-video supervision.
- Method: The framework constructs approximately 16K synthetic DPO instances from short visual contexts, eliminating the need for long-video annotations.The training set is substantially smaller than existing instruction-tuning datasets while transferring short-context supervision to long-video settings.
- Results: LongVPO outperforms existing long-video models across challenging benchmarks while maintaining competitive short-video performance.The reported comparison includes models trained with large-scale supervised or preference-optimized data, whereas LongVPO uses synthetic data without human labels.
2 Related work
Prior video-VLM research improves long-video understanding through context compression, temporal-window extension, and preference optimization. LongVPO differs by targeting ultra-long videos with a two-stage framework that transfers short-clip supervision and uses synthetic scene-level reasoning data.
- VLMs for Long Video Understanding: Long-video VLMs commonly compress or select visual context, extend temporal context windows, or construct long-form instruction datasets.These approaches combine architectural changes with data construction to address long-video understanding.
- DPO for Video-VLMs: DPO trains video-VLMs with queries, preferred responses, and rejected responses to improve alignment and spatial-temporal perception.Existing video-DPO methods generate preference data through perturbations such as frame cutout, clip dropping, rearrangement, and spatial misalignment.
- DPO for Video-VLMs: Many prior video-DPO methods focus on minute-level or short-form videos rather than long-form video understanding.The related-work discussion distinguishes these methods from efforts that align language models with long text contexts.
- LongVPO: LongVPO forms pseudo-long sequences from short clips and uses anchor-conditioned preferred responses alongside distractor-conditioned dispreferred responses.This construction simulates temporal misalignment while training retrieval from many candidate clips.
3 Method
LongVPO adapts short-context VLMs to long-video preference optimization through anchor-based synthetic supervision and self-training on coherent scene metadata. Its two stages address context mismatch, ambiguous distractors, and the need for multi-segment temporal reasoning.
- 3.1 Background: DPO aligns a policy model with preferences by comparing preferred and dispreferred response likelihoods against a frozen reference model.The preference margin is controlled by β, which determines how sharply the policy prefers the preferred response.
- 3.3 Stage 1: Efficient Short-to-Long Learning from Anchored Cues: Stage 1 synthesizes anchor-centric preference triples by making each question answerable only from one short clip, concatenating distractors, and generating a plausible incorrect response from non-anchor clips.The resulting composite sequence exposes the model to long-range contextual variation while preserving an identifiable answer source.
- 3.3 Stage 1: Efficient Short-to-Long Learning from Anchored Cues: Scene-similarity and question-specificity filtering reduce visual overlap among clips and verify that questions depend on multiple distinct elements of the anchor.Samples with overly similar distractors can be replaced or discarded, while insufficiently specific questions are filtered.
- 3.3 Stage 1: Efficient Short-to-Long Learning from Anchored Cues: Stage 1 evaluates the reference model only on the anchor clip, avoiding context-length mismatch while reducing computational and memory costs.This approximation assumes that non-anchor segments provide no relevant signal for answering the anchor question.
- 3.4 Stage 2: Self-Training for Long Video Preference Alignment: Stage 2 uses recursive captioning to create context-aware scene descriptions, then generates scene-grounded reasoning queries and preferred or degraded responses for long-video self-training.Reasoning traces cite relevant scene IDs; degraded responses use partial evidence or irrelevant scenes to induce incomplete reasoning or hallucination.
- 3.5 Total Objective: Both stages combine their DPO loss with a weighted supervised fine-tuning loss on the preferred response.The total objective adds α times the negative log-likelihood of the preferred response to the stage-specific DPO loss.
4 Experiment
LongVPO is evaluated across long- and short-video benchmarks, ablations, context-length tests, and qualitative tasks. Results show competitive long-video performance with synthetic data, preserved short-video ability, improved multi-segment reasoning, and stronger scaling than the baseline.
- Evaluation Setup: LongVPO uses 16k publicly sourced samples, split into 10k Stage 1 and 6k Stage 2 examples, without costly human annotations or closed-source tools.Stage 1 uses LLaVA-Video-178K captions, while Stage 2 uses scene-segmented unlabeled videos from Vript.
- Main Results: LongVPO achieves competitive performance on long-video benchmarks while maintaining competitive short-video performance on MVBench, including a +1.1 improvement.The evaluation covers long-video benchmarks, VideoMME, and MVBench across durations from seconds to two hours.
- Stage Effectiveness: Stage 2 consistently improves most settings by aggregating question-relevant information across multiple real-video segments.This contrasts with Stage 1’s focus on localizing a single segment and improves alignment with realistic scenarios.
- Generality: LongVPO remains effective beyond short-context models, consistently surpassing counterparts trained on InternVL2.5 when applied to InternVideo2.5.The result indicates additional long-context gains remain available even for a model pretrained on long videos.
- Scaling: As input frames increase, LongVPO gains progressively while InternVL-2.5-8B saturates, demonstrating stronger use of extended temporal information.LongVPO improves from 256 to 512 frames and maintains superior performance across benchmarks.
- Ablations: DPO on synthesized Stage 1 preference pairs produces the largest improvements and requires approximately 70% of the training time of full-long-video reference scoring.Removing scene filtering degrades performance, while semantic filtering supports robust long-video understanding.
- Qualitative Comparison: LongVPO accurately counts all 5 pumpkin-carving instances, whereas Qwen2.5-VL, Qwen2-VL, and LLaVA-Video fail to provide correct counts.The task tests action recognition and temporal instance tracking across extended durations.
5 Conclusion
LongVPO extends short-context VLMs to long-video comprehension through synthetic DPO instances from short visual contexts, without annotated long-video data. It achieves state-of-the-art performance on long- and short-video benchmarks, while inference computational efficiency remains an open limitation.
- LongVPO is a two-stage DPO framework for long-video understanding.It incrementally extends short-context VLM capabilities using synthetic DPO instances constructed from short visual contexts.
- LongVPO requires no annotated long-video data.
- LongVPO achieves state-of-the-art performance on both long- and short-video understanding benchmarks.
- Inference computational efficiency is not prioritized and is left for future integration with context-compression approaches.
Appendix Overview
The appendix organizes supplementary material covering context-bias experiments, qualitative comparisons, implementation details, and long-video benchmark performance.
- Section A describes the experimental design behind the context-bias analysis.
- Section B presents qualitative results across diverse long-video scenarios.
- Section C provides implementation details of the method.
- Table 3 compares performance on long-video benchmarks and marks improvements over base models with red upward arrows.
A Core Experimental Design for Context Position Bias Probing (Main Fig. 1)
The context-position-bias experiment simulates long contexts by padding and randomly relocating meaningful video frames. It tests consistency across context lengths and invariance to the location of valid frames.
- Evaluation Setup: The evaluation selects unambiguous MVBench tasks to ensure valid labels.
- Padding Strategy: Long-context scenarios embed original video frames in a larger grid surrounded by meaningless padding frames.
- Random Placement: Random placement tests whether performance depends on the spatial location of meaningful video content.
- Consistency across Context Lengths: An ideal model should preserve task performance across short and long contexts without changing task semantics.
- Position Invariance: An ideal model should be position-invariant because padding frames contain no meaningful information.
- Observed Results: Existing long-context models show position sensitivity and a short-to-long performance drop relative to LongVPO.
B Qualitative Results
The appendix reports qualitative comparisons across diverse long-video scenarios, where LongVPO remains competitive despite synthetic training. Additional figures address visual semantics, cross-temporal association, temporal order, and detail comprehension, alongside context-bias analyses.
- Qualitative Results: LongVPO demonstrates competitive open-ended QA performance across diverse long-video domains despite being trained on synthetic data.
- Context-Bias Analysis: Figures 8 and 9 provide context-bias analyses, with Figure 9 showing that the lost-in-the-middle phenomenon persists under shortened input context.
- Visual Semantic Understanding: Figure 10 presents long-video visual semantic understanding results.
- Cross-temporal Scene Association: Figure 11 presents long-video cross-temporal scene association results.
- Temporal Order Analysis: Figure 12 presents long-video temporal order analysis results.
- Detail Comprehension: Figure 13 presents long-video detail comprehension results.
C Implementation Details
Experiments use a 16k-sample, two-stage LongVPO training setup with full-model fine-tuning and extended video contexts. Training uses H100 hardware, sequence parallelism, and specified optimization and evaluation settings.
- Hardware Setup: Experiments run on 4 × 8 NVIDIA H100 GPUs with 80GB per GPU, using DeepSpeed Ulysses sequence parallelism for memory-efficient training.
- Training Strategy: 16k curated samples are split into 10k for Stage 1 and 6k for Stage 2, with each model variant trained for 1 epoch.The split balances training efficiency with robust adaptation.
- Full-Model Fine-Tuning: The entire model is fine-tuned end-to-end, including the vision encoder, vision-language connector, and LLM backbone.
- Optimization Settings: Optimization uses a composite KL-divergence and SFT loss with β=0.01 and α=1.0, a 5e-7 learning rate, batch size 8, cosine scheduling, and 0.01 warm-up.
- Training Duration: Each model variant requires approximately 10 hours with DeepSpeed Ulysses enabled, compared with about 1 hour without it under the stated configuration.
- Evaluation Settings: Evaluation follows InternVL2.5 and InternVideo2.5 protocols and uses up to 512 frames per input to test long-context scalability.
NeurIPS Paper Checklist
The checklist records that the paper reports its contributions, experimental details, limitations, and computational resources, while noting that it does not provide statistical significance analysis. It also states that data and code will be released upon acceptance and that the work has no additional societal impacts identified by the authors.
- Claims and Contributions: The authors state that the abstract and introduction outline the work’s contributions and summarize them at the introduction’s end.
- Theory Assumptions and Proofs: The paper reports no theoretical results, so theorem assumptions and complete proofs are marked not applicable.
- Experimental Reproducibility: The methods, experiments, and appendix provide information intended to reproduce the main experimental results.
- Open Access to Data and Code: Data and code are stated to be available upon acceptance.
- Experimental Setting and Resources: Training and test details are reported in the experiments section and appendix, while computational resources are detailed in the appendix.
- Experiment Statistical Significance: The paper does not include error bars or extensive statistical significance analysis because of limited computational resources.