Temporal Gains, Spatial Costs: Revisiting Video Fine-Tuning in Multimodal Large Language Models

• URL: http://arxiv.org/abs/2603.17541v1
• Date: Wed, 18 Mar 2026 09:46:44 GMT
• Title: Temporal Gains, Spatial Costs: Revisiting Video Fine-Tuning in Multimodal Large Language Models
• Abstract summary: We systematically study how Video-SFT reshapes visual capabilities in MLLMs.<n>Video-SFT reliably improves video performance, but often yields limited gains or even degradation on static image benchmarks.<n>We study an instruction-aware Hybrid-Frame strategy that adaptively allocates frame counts and partially mitigates the image-video trade-off.
• Score: 36.34630132055548
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Multimodal large language models (MLLMs) are typically trained in multiple stages, with video-based supervised fine-tuning (Video-SFT) serving as a key step for improving visual understanding. Yet its effect on the fine-grained evolution of visual capabilities, particularly the balance between spatial and temporal understanding, remains poorly understood. In this paper, we systematically study how Video-SFT reshapes visual capabilities in MLLMs. Across architectures, parameter scales, and frame sampling settings, we observe a consistent pattern: Video-SFT reliably improves video performance, but often yields limited gains or even degradation on static image benchmarks. We further show that this trade-off is closely tied to temporal budget: increasing the number of sampled frames generally improves video performance, but does not reliably improve static image performance. Motivated by this finding, we study an instruction-aware Hybrid-Frame strategy that adaptively allocates frame counts and partially mitigates the image-video trade-off. Our results indicate that Video-SFT is not a free lunch for MLLMs, and preserving spatial understanding remains a central challenge in joint image-video training.

Related papers

• Think-Clip-Sample: Slow-Fast Frame Selection for Video Understanding [18.742640666790827]
  Think-Clip-Sample (TCS) is a training-free framework that enhances long video understanding through two key components.<n>Extensive experiments on MLVU, LongVideoBench, and VideoMME demonstrate that TCS consistently improves performance across different MLLMs.
  arXiv  Detail & Related papers  (2026-01-16T15:14:04Z)
• Zero-Shot Video Translation and Editing with Frame Spatial-Temporal Correspondence [81.82643953694485]
  We present FRESCO, which integrates intra-frame correspondence with inter-frame correspondence to formulate a more robust spatial-temporal constraint.<n>Our method goes beyond attention guidance to explicitly optimize features, achieving high spatial-temporal consistency with the input video.<n>We verify FRESCO adaptations on two zero-shot tasks of video-to-video translation and text-guided video editing.
  arXiv  Detail & Related papers  (2025-12-03T15:51:11Z)
• FrameMind: Frame-Interleaved Video Reasoning via Reinforcement Learning [65.42201665046505]
  Current video understanding models rely on fixed frame sampling strategies, processing predetermined visual inputs regardless of the specific reasoning requirements of each question.<n>This static approach limits their ability to adaptively gather visual evidence, leading to suboptimal performance on tasks that require broad temporal coverage or fine-grained spatial detail.<n>We introduce FrameMind, an end-to-end framework trained with reinforcement learning that enables models to dynamically request visual information during reasoning through Frame-Interleaved Chain-of-Thought (FiCOT)<n>Unlike traditional approaches, FrameMind operates in multiple turns where the model alternates between textual reasoning and active visual perception, using tools to extract
  arXiv  Detail & Related papers  (2025-09-28T17:59:43Z)
• Episodic Memory Representation for Long-form Video Understanding [52.33907540905242]
  Large Video Language Models excel at general video understanding but struggle with long-form context window limits.<n>We introduce Video-EM, a training free framework inspired by the principles of human memory.<n>Video-EM achieves performance gains of 4-9 percent over respective baselines while utilizing fewer frames.
  arXiv  Detail & Related papers  (2025-08-13T04:33:07Z)
• KFFocus: Highlighting Keyframes for Enhanced Video Understanding [33.69757683688046]
  We propose KFFocus, a method designed to efficiently compress video tokens and emphasize the informative context present within video frames.<n>By assigning varying condensation ratios to frames based on their contextual relevance, KFFocus efficiently reduces token redundancy while preserving informative content details.<n>We also introduce a multimodal modeling module that encodes both the temporal relationships between video frames and the spatial structure within each frame.
  arXiv  Detail & Related papers  (2025-08-12T14:57:03Z)
• STORM: Token-Efficient Long Video Understanding for Multimodal LLMs [116.4479155699528]
  STORM is a novel architecture incorporating a dedicated temporal encoder between the image encoder and the Video-LLMs.<n>We show that STORM achieves state-of-the-art results across various long video understanding benchmarks.
  arXiv  Detail & Related papers  (2025-03-06T06:17:38Z)
• Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization [52.63845811751936]
  Video pre-training is challenging due to the modeling of its dynamics video. In this paper, we address such limitations in video pre-training with an efficient video decomposition. Our framework is both capable of comprehending and generating image and video content, as demonstrated by its performance across 13 multimodal benchmarks.
  arXiv  Detail & Related papers  (2024-02-05T16:30:49Z)
• A Simple Recipe for Contrastively Pre-training Video-First Encoders Beyond 16 Frames [57.758863967770594]
  We build on the common paradigm of transferring large-scale, image--text models to video via shallow temporal fusion.<n>We expose two limitations to the approach: (1) decreased spatial capabilities, likely due to poor video--language alignment in standard video datasets, and (2) higher memory consumption, bottlenecking the number of frames that can be processed.
  arXiv  Detail & Related papers  (2023-12-12T16:10:19Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.