A Review of Video Large Language Models: From Spatiotemporal Representations to Reliable Reasoning

Video large language models connect video observations with user needs through natural language, allowing summarization, event localization, question answering, and ongoing interaction to draw on shared model capabilities. Yet accepting more frames or generating longer answers does not establish that a system has retained and correctly used the information required by a task. This review examines video understanding since the emergence of large language models, starting from application needs and proceeding through visual interfaces and representations, information management for long videos, intermediate reasoning, data and training, and capability validation. We first compare the pathways through which visual information enters a language model and examine how temporal computation and learning objectives jointly determine the distinctions a model can make. We then trace observation, compression, storage, and retrieval to explain how efficiency methods make decisions, where they save computation, and under what conditions omitted information can be recovered. On this basis, we distinguish relational organization, active observation, and judgment revision as functions of video reasoning, and analyze the roles of text, spatiotemporal anchors, latent features, and tool operations. The analysis indicates that fine-grained capabilities require training examples and objectives that explicitly distinguish the relevant changes; efficiency on long videos requires attention to both information retention and subsequent access; and the value of additional reasoning should be measured by the specific errors it corrects. We connect these findings to sample organization, parameter updates, process feedback, and diagnostic experiments, and discuss the conditions under which predictive and language learning, persistent memory, and transfer to generation and action can be effective. The resulting analysis supports technical choices under specific task requirements. Project repository: https://github.com/LeslieZhoa/Awesome_Video_Large_Language_Models

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24310883
Primary Topic
Multimodal Machine Learning Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Review of Video Large Language Models: From Spatiotemporal Representations to Reliable Reasoning

Xiang Zhao
Zenodo (CERN European Organization for Nuclear Research)
Multimodal Machine Learning Applications
preprint

A Review of Video Large Language Models: From Spatiotemporal Representations to Reliable Reasoning

Xiang Zhao
preprint en

Abstract

Video large language models connect video observations with user needs through natural language, allowing summarization, event localization, question answering, and ongoing interaction to draw on shared model capabilities. Yet accepting more frames or generating longer answers does not establish that a system has retained and correctly used the information required by a task. This review examines video understanding since the emergence of large language models, starting from application needs and proceeding through visual interfaces and representations, information management for long videos, intermediate reasoning, data and training, and capability validation. We first compare the pathways through which visual information enters a language model and examine how temporal computation and learning objectives jointly determine the distinctions a model can make. We then trace observation, compression, storage, and retrieval to explain how efficiency methods make decisions, where they save computation, and under what conditions omitted information can be recovered. On this basis, we distinguish relational organization, active observation, and judgment revision as functions of video reasoning, and analyze the roles of text, spatiotemporal anchors, latent features, and tool operations. The analysis indicates that fine-grained capabilities require training examples and objectives that explicitly distinguish the relevant changes; efficiency on long videos requires attention to both information retention and subsequent access; and the value of additional reasoning should be measured by the specific errors it corrects. We connect these findings to sample organization, parameter updates, process feedback, and diagnostic experiments, and discuss the conditions under which predictive and language learning, persistent memory, and transfer to generation and action can be effective. The resulting analysis supports technical choices under specific task requirements. Project repository: https://github.com/LeslieZhoa/Awesome_Video_Large_Language_Models

Zenodo (CERN European Organization for Nuclear Research)
Multimodal Machine Learning Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon