cs.MMMar 20, 2026

Leum-VL Technical Report

Authors: Yuxuan He, Chaiming Huang, Yifan Wu, Hongjun Wang, Chenkui Shen, Jifan Zhang, Long Li

Organizations: Hainan Sihe Data Technology Co., Ltd.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

A short video succeeds not simply because of what it shows, but because of how it schedules attention -- yet current multimodal models lack the structural grammar to parse or produce this organization. Existing models can describe scenes, answer event-centric questions, and read on-screen text, but they are far less reliable at identifying timeline-grounded units such as hooks, cut rationales, shot-induced tension, and platform-facing packaging cues. We propose SV6D (Structured Video in Six Dimensions), inspired by professional storyboard practice in film and television production, a representation framework that decomposes internet-native video into six complementary structural dimensions -- subject, aesthetics, camera language, editing, narrative, and dissemination -- with each label tied to physically observable evidence on the timeline. We formalize a unified optimization objective over SV6D that combines Hungarian-matched temporal alignment, dimension-wise semantic label distance, and quality regularization. Building on this framework, we present Leum-VL-8B, an 8B video-language model that realizes the SV6D objective through an expert-driven post-training pipeline, further refined through verifiable reinforcement learning on perception-oriented tasks. Leum-VL-8B achieves 70.8 on VideoMME (w/o subtitles), 70.0 on MVBench, and 61.6 on MotionBench, while remaining competitive on general multimodal evaluations such as MMBench-EN. We also construct FeedBench, a benchmark for structure-sensitive short-video understanding. Our results indicate that the missing layer in video AI is not pixel generation but structural representation: grounded on the timeline, linked to visible evidence, and directly consumable by downstream workflows such as editing, retrieval, recommendation, and generation control, including text-heavy internet video formats with overlays and image-text layouts.

Figures & tables

Figure 1: Video understanding benchmark comparison of Leum-VL-8B against 8B-scale models.
Figure 2: Overview of the data synthesis pipeline. Starting from raw internet video, the pipeline proceeds through shot segmentation, LoRA-based structural annotation, scorer-based quality filtering, LLM-assisted structural extension, and dense report generation, producing the SV6D SFT training corpus.
DimensionSub-dimensions
Camera languageShot size
Camera position
Shooting angle
Lens focal length
Camera movement
Depth of field
Table 1: Per-dimension label taxonomy of FeedBench. Sub-dimensions are listed for each of the six structural dimensions.
DimensionTemporal anchorTask Types# Labels# Items
SubjectIntervalSingle, Ordered888 †
AestheticsIntervalSingle, Multi, Ordered416,645
Camera languageIntervalSingle, Multi, Ordered385,927
EditingInterval / TimestampSingle, Multi, Ordered65949
NarrativeIntervalSingle, Multi, Ordered[WIP][WIP]
DisseminationIntervalSingle, Multi40194
Table 2: Overview of FeedBench, containing 60 short-form videos from 27 content categories. All core items are grounded in either a temporal interval or a timestamp, instantiated as closed-vocabulary multiple-choice or IoU-scored temporal localization tasks. † Subject items are currently IoU-only (temporal localization); narrative items are under construction ([WIP]). Single, Multi, and Ordered denote single-label, multi-label, and ordered-transition items, respectively.
ParameterValue
Group size GG8
Prompts per step390 (13 task types ×\times 30)
Sequences per step (GBS)3,120 ( 390×8390\times 8 )
PPO mini-batch size390 (single gradient step)
Learning rate1×10−61\times 10^{-6} (cosine →1×10−7\to 1\times 10^{-7} )
KL penaltydisabled
Table 3: GRPO training hyperparameters.
Task TypeReward
Temporal grounding0.2 Format+0.4 IoU+0.4 Label0.2\,\mathrm{Format}+0.4\,\mathrm{IoU}+0.4\,\mathrm{Label}
Temporal action localization0.2 Format+0.8 IoU0.2\,\mathrm{Format}+0.8\,\mathrm{IoU}
OCR (regular & handwritten)EditDist(y^, y∗)\mathrm{EditDist}(\hat{y},\,y^{*})
Chain-of-thought0.3 Format+0.7 Judge0.3\,\mathrm{Format}+0.7\,\mathrm{Judge}
Table 4: Reward function by task type.
Figure 3: GRPO training dynamics over 150 steps. Left : actor entropy decreases steadily, indicating the policy becomes more confident. Center : mean reward score rises from ∼\sim 0.66 to ∼\sim 0.80. Right : mean response length remains stable, showing no reward hacking through verbosity.
ModelSubjectAestheticsCamera LanguageEditingNarrativeDissemination
Qwen3-VL-8B-Instruct †62.344.042.049.2–44.9
Qwen3.5-9B ( Team, 2025b ) †69.652.342.056.4–56.4
Leum-VL-8B78.364.853.074.8–71.1
Table 5: Performance of Leum-VL-8B on FeedBench.
CategoryBenchmarkLeum-VL-8BQwen3-VL-8B 1Keye-VL-8B Thinking 2GLM-4.1V-9B Thinking 3MiniCPM-V-4.5-8B 4
General VQAMMBench-EN test{}_{\text{test}}84.884.592.085.884.2
MMBench-CN test{}_{\text{test}}83.984.784.7
HallusionBench56.561.162.763.261.2
RealWorldQA73.271.573.5—72.1
MMStar67.570.980.572.972.1
BLINK65.269.154.9 ‡65.142.0 ‡
Table 6: Performance of Leum-VL-8B and other 8B-scale models on open benchmarks.
More in this topic

Video Understanding

Video understanding develops methods for recognizing actions, tracking events, grounding evidence in time, and reasoning across visual and audio streams.

See all 116 papers

Explore similar work

CardsList
May 19, 2026

HAVEN: Hierarchically Aligned Multimodal Benchmark for Unified Video Understanding

While Multimodal Large Language Models (MLLMs) exhibit strong performance on standard video tasks, their ability to faithfully summarize and reason over complex narratives remains poorly evaluated. Existing summarization benchmarks fragment supervision across isolated granularities, such as keyframes, key shots, or disjointed text summaries, failing to capture the inherently hierarchical structure of cross-modal alignment. To address this critical gap, we introduce HAVEN, a hierarchically aligned multimodal benchmark for unified video understanding. HAVEN pioneers a fully granular (frame, shot, and video levels) and fully multimodal (video and text) dataset architecture, complete with explicit, continuous alignment between modalities. Built upon this unified annotation paradigm, we propose a comprehensive evaluation suite spanning summarization, temporal reasoning, multimodal grounding, and saliency ranking. Extensive benchmarking of state-of-the-art MLLMs exposes a persistent gap between surface-level textual fluency and grounded multimodal understanding. Ultimately, HAVEN advances the evaluation of multimodal systems beyond traditional QA formats, offering a rigorous, standardized testbed to drive future research in interpretable, hierarchical video understanding. We publicly release the dataset, benchmark suite, and evaluation protocols.

Mengqi Shi, Haopeng Zhang

Department of Information and Computer Sciences · University of Hawaii at Manoa

Multimodal GroundingVideo SummarizationMultimodal Model EvaluationTemporal Video UnderstandingMultimodal Understanding
AskPDF
Aug 23, 2026

Improving Spatial-Temporal Reasoning in Video-Language Models with Structured Video Prompting

Video-language models (VLMs) remain brittle on tasks that require tracking events over time and grounding answers in specific spatial regions. We propose that part of this limitation can be addressed through better organization of visual evidence at inference time. We introduce structured video prompting, a training-free inference-time method that augments the input video with lightweight spatial structure and temporal structure, providing explicit anchors for organizing evidence across space and time without changing model weights or decoding and without altering the question prompt in the main comparison. We evaluate this approach on two complementary video benchmarks and two open video-language models. Across these settings, structured inputs improve performance in several cases, with gains varying by model and task. Our findings suggest that some failures of VLMs arise not only from reasoning capacity, but also from how video evidence is presented at inference time. These results highlight structured video prompting as a simple and practical direction for improving video understanding.

Sadegh Mohammadian

Sharif University of Technology

Video-Language ModelsStructured PromptingTemporal Video Understanding
AskPDF
Jun 1, 2026

Moment-Video: Diagnosing Temporal Fidelity of Video MLLMs on Momentary Visual Events

Video multimodal large language models (MLLMs) have made rapid progress in general and long-form video understanding, yet their ability to capture and use brief, answer-critical visual evidence remains underexplored. Momentary actions or state transitions may determine an answer while occupying only a small fraction of the video. Performance on broad video understanding tasks alone does not establish whether models can reliably identify and use this temporally localized evidence. We introduce Moment-Video, a benchmark for evaluating temporal fidelity through momentary visual event understanding. Questions are constructed around visually observable events or event sequences, requiring models to notice, count, describe, or reason about transient evidence. The benchmark contains 1,000 human-verified video-QA pairs across seven domains and 25 subcategories, covering four tasks: Temporal Occurrence, Temporal Counting, Action Description, and Temporal Reasoning. We evaluate 34 proprietary and open-source MLLMs. Under the main evaluation settings, the best-performing model, Seed-2.0-Pro, achieves 39.6% overall accuracy, while most open-source models remain below 25%. Diagnostic analyses show that denser frame sampling improves some models, but substantial errors persist even when sampled frames cover all annotated necessary event intervals in an audited subset. Performance also declines on longer videos, suggesting additional temporal-localization challenges. By exposing these limitations, Moment-Video provides a diagnostic benchmark to guide the evaluation and development of video MLLMs that more reliably understand brief, answer-critical visual events.

Xiaolin Liu, Yilun Zhu, Xiangyu Zhao +9

Shandong University · Southeast University · Shanghai Jiao Tong University +1

VLM EvaluationVideo QAVideo-Language ModelsTemporal Video Understanding
AskPDF