# The Visual Bottleneck: Sparse-Frame Adaptation of MLLMs for Joint Spatial-Temporal Video Grounding

arXiv: https://arxiv.org/abs/2607.24570 · HTML: https://mlsanity.com/paper/2607.24570/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 16 of its 23 references that are in the map.
- First posted on arXiv 27 Jul 2026.

## Builds on (most cited first, 16 of 16)

- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection](https://mlsanity.com/paper/2303.05499/) (2023)
- [InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks](https://mlsanity.com/paper/2312.14238/) (2023)
- [Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding](https://mlsanity.com/paper/2306.02858/) (2023)
- [LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://mlsanity.com/paper/2403.13372/) (2024)
- [HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips](https://mlsanity.com/paper/1906.03327/) (2019)
- [Localizing Moments in Video with Natural Language](https://mlsanity.com/paper/1708.01641/) (2017)
- [TALL: Temporal Activity Localization via Language Query](https://mlsanity.com/paper/1705.02101/) (2017)
- [Hollywood in Homes: Crowdsourcing Data Collection for Activity Understanding](https://mlsanity.com/paper/1604.01753/) (2016)
- [VTimeLLM: Empower LLM to Grasp Video Moments](https://mlsanity.com/paper/2311.18445/) (2023)
- [LLaVA Steering: Visual Instruction Tuning with 500x Fewer Parameters through Modality Linear Representation-Steering](https://mlsanity.com/paper/2412.12359/) (2024)
- [Context-Guided Spatio-Temporal Video Grounding](https://mlsanity.com/paper/2401.01578/) (2024)
- [Video-GroundingDINO: Towards Open-Vocabulary Spatio-Temporal Video Grounding](https://mlsanity.com/paper/2401.00901/) (2023)
- [Rethinking the Video Sampling and Reasoning Strategies for Temporal Sentence Grounding](https://mlsanity.com/paper/2301.00514/) (2023)
- [Structured Video-Language Modeling with Temporal Grouping and Spatial Grounding](https://mlsanity.com/paper/2303.16341/) (2023)

## Abstract (arXiv)

Large-scale video platforms process millions of uploads hourly, requiring moderation systems that can localize when and where policy violations occur within each video. Processing every frame is infeasible at scale, so systems are constrained to sparse inputs of 8 to 16 frames per video. Yet state-of-the-art multimodal large language models (MLLMs) are pretrained on dense sequences of hundreds of frames, creating a fundamental mismatch between training and deployment conditions. This mismatch causes severe performance collapse: the Qwen3-VL 8B model drops from 56.0% to 22.3% temporal mIoU when frames are reduced to 16, a 60.2% relative degradation. We present a systematic empirical study of training strategies to close this gap for spatial-temporal video grounding. Our results suggest that visual feature extraction is the dominant bottleneck under sparse-frame inputs. Adapting only the final three ViT layers, 4% of total parameters, achieves 68.8% temporal mIoU and surpasses a zero-shot 8B model using dense inputs by 12.8 points. Language model fine-tuning, by contrast, offers negligible or negative returns. A boundary-aware sampling strategy, Hybrid16, further improves temporal mIoU by 26 points over uniform sampling when temporal boundaries are available. We conclude that for sparse-frame video grounding, training strategy dominates model scale: a fine-tuned 2B model consistently outperforms a zero-shot 8B model, with or without dense frame access.