PaperlayerView exact revision on arXiv

Source-linked AI summary

Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model

Senqiao Yang, Kaichen Zhang, Zhaoyang Jia, Jinghao Guo, Yifei Shen, Xinjie Zhang, Xiaoyi Zhang, Haoqing Wang, Xiao Li, Peng Zhang, Xiang An, Yin Xie, Zhening Liu, Xun Guo, Jiahao Li, Shicheng Zheng, Jinglu Wang, Zongyu Guo, Wenxuan Xie, Zihan Zheng, Yuxuan Luo, Bin Li, Yan Lu

arXiv:2607.24904v1cs.CVcs.CL

TL;DR

Existing VLMs handle sparse-frame reasoning well but remain slow and computationally heavy for continuous dynamic video. Mage-VL introduces codec-native, event-driven streaming with motion-salient visual updates and reports improvements across video and spatial reasoning, with up to 3.5× faster wall-clock inference.

  • Problem

    Existing VLMs perform well on sparsely sampled keyframes but remain slow and computationally heavy for continuous visual events because uniform sampling overlooks spatiotemporal redundancy.

  • Method

    Mage-VL combines codec-native motion-salient patch selection with a lightweight event gate that triggers a causal language decoder for proactive streaming interaction.

  • Results

    Mage-VL-4B matches flagship baselines on static image understanding, improves video and 2D/3D spatial reasoning, and achieves up to 3.5× wall-clock speedups over dense frame sampling.

  • Takeaways & Limitations

    Codec-native streaming establishes an accuracy-efficiency approach for multimodal modeling while supporting empirical findings on data efficiency, resolution scaling, and post-training.

  • Takeaways & Limitations

    Mage-VL remains limited in complex agentic workflows and mathematical reasoning, attributed to limited high-quality text data and omitted RL post-training.

Abstract

from arXiv · show

Standard vision-language models (VLMs) suffer from Moravec's paradox: they excel at complex offline visual reasoning but struggle with simple streaming perception tasks and process them inefficiently. We present Mage-VL, an efficient codec-native streaming foundation model for real-time multimodal understanding and interaction. At its core, our custom tokenizer, Mage-ViT, replaces uniform frame sampling by selectively encoding dynamic, entropy-rich regions using motion vectors and residual energy across sparse anchor (I) and predicted (P) frames. Operating at a 16 x 16 patch level, this reduces visual token consumption by over 75% while preserving spatiotemporal context. Trained from scratch on approximately 560M unlabeled images and 100M unlabeled video frames, Mage-ViT matches or outperforms flagship encoders trained on billions of image-text pairs. We establish AI4AI data pipelines encompassing prompt-code joint optimization for multimodal captioning and AI-driven performance diagnosis to guide training recipes. Furthermore, through a bio-inspired dual-system architecture - a lightweight System 1 event gate and a causal System 2 decoder - Mage-VL enables proactive streaming perception. Extensive evaluations show that Mage-VL-4B matches Qwen3-VL-4B on static tasks while achieving strong gains in video understanding and 2D/3D spatial reasoning, with up to a 3.5x wall-clock inference speedup, and comprehensively surpasses the 15B Phi-4-reasoning-vision baseline. Beyond model artifacts, we deliver seven key empirical findings covering pre-training data efficiency, variable-resolution scaling, codec system acceleration, VideoQA SFT redundancy, motion-spatial synergy, AI4AI data pipelines, and Zero-Vision SFT for multimodal RL.

1 Introduction

Mage-VL addresses the inefficiency of conventional VLMs on continuous video by using codec-native, motion-aware visual encoding and a dual-system streaming architecture. Trained with substantially less data than web-scale vision encoders, it achieves competitive static performance while improving video and spatial reasoning efficiency.

  • Motivation: Conventional VLMs struggle with continuous video because uniform frame sampling repeatedly encodes unchanged content, causing computation to scale rapidly with video duration.This reflects a structural mismatch between continuous, event-driven perception and architectures built around static-image encoders.
  • Architecture: Mage-VL selectively encodes motion-salient patches using codec motion vectors and residual energy across fully processed I-frames and predicted frames.Its codec-native workflow reduces visual token consumption by approximately 75% while supporting real-time online video question answering.
  • Architecture: A lightweight System 1 gate monitors streaming inputs and triggers a causal System 2 decoder, combining low-latency event detection with deliberate reasoning.The design is inspired by biological vision’s filtering of redundant background signals and prioritization of dynamic motion.
  • Empirical findings: 560M unlabeled images and 100M unlabeled video frames suffice for Mage-ViT to perform on par with, and sometimes better than, encoders trained on billions of image–text pairs.The comparison includes SigLIP2 and MoonViT.
  • Results and contributions: Up to 3.5× wall-clock inference speedups accompany Mage-VL-4B’s parity with Qwen3-VL-4B on static tasks and stronger performance on video and spatial reasoning.The work also reports seven empirical findings spanning data efficiency, resolution scaling, codec acceleration, VideoQA SFT redundancy, motion-spatial synergy, AI4AI pipelines, and Zero-Vision SFT.

2 Related Work

Prior VLMs reduce visual redundancy mainly through patch pruning, temporal compression, or frame selection after dense encoding. Mage-VL instead makes codec-aligned sparse representation and streaming native to the visual front end, using video-native training and dual-system response triggering.

  • Visual tokenization: Most VLM encoders use dense patch grids, while DynamicViT, AdaViT, ToMe, FastV, and LLaVA-PruMerge explore adaptive token pruning or merging.These approaches target redundant visual tokens within image-centric encoder pipelines.
  • Video representation: Video models extend visual encoding temporally, but many reduce cost only after dense frame encoding or by selecting entire frames, retaining spatio-temporal redundancy.Representative methods include TimeSformer, ViViT, Video Swin Transformer, VideoMAE, LLaMA-VID, Chat-UniVi, SlowFast-LLaVA, MovieChat, LongVU, and VideoChat-Flash.
  • Codec-native modeling: Codec-based work exploits motion vectors and residuals, progressing from CoViAR and Video-LaVIT to codec-aligned patch sparsity and video-native VLM training.The cited systems include OneVision Encoder and LLaVA-OneVision-2.
  • Mage-VL: Mage-ViT allocates tokens to temporally informative regions across anchor and sparse predicted frames before visual encoding, enabling efficient long-video processing and incremental streaming.Unlike image-centric encoders, it is trained from scratch on both image and video data.
  • Mage-VL: Unlike frame-based video VLMs, Mage-VL directly consumes sparse codec-aligned streams, reducing temporal redundancy before language modeling through a video-native encoder trained from scratch.This contrasts with sampled-frame approaches that use dense per-frame tokenization.
  • Streaming VLMs: Unlike prior methods that add streaming mechanisms to conventional encoders, Mage-VL processes rolling codec-token windows and couples proactive response triggering with language generation through a dual-system architecture.Prior streaming work includes VideoLLM-online, MMDuet, Dispider, StreamMind, JoyAI, Flash-VStream, StreamingVLM, and InternLM-XComposer2.5-OmniLive.

3 Mage-ViT

Mage-ViT is a codec-native visual encoder that sparsely selects information-rich 16×16 patches while preserving spatiotemporal structure. It combines a Codec-ViT architecture with a two-stage image-and-video pre-training pipeline.

  • Architecture: Mage-ViT combines a codec-driven patchifier, a Vision Transformer trunk, and shared 3D rotary positional encoding.The trunk is a 24-layer pre-norm Vision Transformer with hidden dimension 1024 and 16 attention heads.
  • Codec-driven patchifier: The patchifier uses 16×16 patches and codec-derived importance from motion-vector magnitude and P-frame residual energy to select sparse visual tokens.It keeps all I-frame patches and selects top-k P-frame patches within token budget B.
  • Codec-driven patchifier: 75% token reduction is achieved for a 64-frame clip using B = 4096 tokens, while shared 3D rotary encoding retains relations across the un-pruned spatiotemporal grid.Mage-ViT also supports chunk-wise and collage patchification modes.
  • Pre-training: Mage-ViT is trained from scratch in two stages: variable-resolution image pre-training followed by joint image-and-video pre-training.Stage 1 uses image resolutions from 224 to 448 and aspect ratios from 1:2 to 2:1; Stage 2 trains 64-frame videos at resolution 256 with token budget 4096.
  • Pre-training: Training uses cluster discrimination over MetaCLIP-derived K-means visual prototypes to align semantically related image and short-video samples.The objective applies negative sampling against visual concept prototypes.

4 Mage-VL

Mage-VL unifies image understanding, offline video reasoning, and proactive interaction over continuous streams using a shared codec-native visual encoder and causal language decoder. Its event gate enables response-triggered streaming, while codec-stream tokenization allocates visual tokens according to spatio-temporal variation for denser temporal coverage at comparable budgets.

  • Unified multimodal architecture: Mage-VL supports static images, variable-duration videos, event-triggered commentary, and conventional user-initiated question answering within one unified multimodal model.The architecture covers image, offline-video, and streaming modes.
  • Codec-native visual encoder: Mage-ViT shares one visual encoder across modalities, densely encoding anchor-frame patches while sparsely selecting predicted-frame patches associated with motion or visual changes.Anchor frames preserve complete scene context, whereas predicted frames retain only changed patches.
  • Proactive streaming mechanism: The cognition gate predicts pspeak = g(ht) and triggers generation when pspeak ≥τ, keeping background segments silent while activating responses for relevant events.During streaming, perception incrementally processes incoming codec windows, while generation uses a local sliding window of recent codec-token segments.
  • Training data and AI4AI optimization: The training corpus contains approximately 350M image–caption pairs, 54M image-instruction samples, 7.95M unique video–caption samples, and 3.35M streaming samples.These data supervise visual-language grounding, task-oriented reasoning, response formatting, and timing over causal video prefixes.
  • Training data and AI4AI optimization: An iterative AI4AI recaptioning pipeline uses GPT-5 rubric scoring, GitHub Copilot prompt refinement, a 200-image validation batch, and human approval before large-scale captioning.The optimized prompt consistently improves OCR, document, chart, perception, and general visual-reasoning benchmark performance.
  • Codec-native visual encoder: Codec-stream tokenization allocates tokens according to spatio-temporal variation rather than dense patch grids, enabling denser temporal coverage under a comparable visual-token budget.For video, codec-token windows are temporally ordered and combine complete anchor-frame context with sparse predicted-frame patches.

5 Experiments · 5.1 Experimental Setup

The experiments evaluate Mage-VL-4B with codec-native visual inputs against same-scale and larger multimodal baselines across representation, image, video, and proactive streaming tasks. Evaluation uses standardized lmms-eval protocols without model-specific prompt tuning or manual output correction.

  • 5.1 Experimental Setup: Mage-VL-4B combines the from-scratch MageViT visual tokenizer with a Qwen3-4B-Instruct-2507 language backbone.At inference, videos use codec-token streams on a shared 16×16 patch grid, with dense anchor frames and selectively encoded predicted-frame patches.
  • 5.1 Experimental Setup: The codec-canvas evaluation includes three operating points spanning different accuracy–efficiency trade-offs.Anchor frames are encoded densely, while predicted frames contribute only patches selected by codec-derived importance signals.
  • 5.1 Experimental Setup: Mage-ViT benefits consistently from larger visual-token budgets and performs best at the highest evaluated resolution, unlike fixed-resolution baselines that saturate or degrade.This comparison evaluates frozen visual features on Food-101 and ImageNet.
  • 5.1 Experimental Setup: Primary comparisons use Qwen3-VL-4B-Instruct at the same language-model scale, alongside Phi-4-MM (5.6B) and Phi-4-R-V (15B).The visual-representation study additionally compares Mage-ViT with public contrastive and self-supervised encoders, including SigLIP, SigLIP2, MetaCLIP2, AIMv2, DINOv3, and Moon…
  • 5.1 Experimental Setup: The evaluation covers representation quality, image understanding, video understanding, and proactive streaming perception.Tasks include image classification, video recognition, document and chart understanding, OCR, VQA, spatial intelligence, video QA, temporal grounding, tracking, and related video reasoning.
  • 5.1 Experimental Setup: Representation quality is measured with linear probes for image classification and attentive probes for video recognition, while downstream understanding spans image and video benchmarks.Image tasks include document and chart understanding, OCR, general VQA, and 2D/3D spatial intelligence; video tasks include video QA, temporal grounding, spatial reasoning, and referring-video tracking.
  • 5.1 Experimental Setup: All supported downstream benchmarks use lmms-eval with official prompt templates and scoring rules.The protocol excludes model-specific prompt tuning, semantic rewriting, and manual output correction; temporal-grounding and tracking outputs follow official evaluation procedures.
  • 5.1 Experimental Setup: Matched-budget video comparisons control the visual-input budget as specified in Section 5.5.Wall-clock measurements in Table 5 are collected on a single node, while temporal-grounding and tracking outputs are parsed using official procedures.

5.2 Evaluation of Mage-ViT

Mage-ViT delivers strong image and video representations without massive web-scale pre-training, while supporting variable-resolution scaling and codec-agnostic sparse token selection. Its video performance improves under extended temporal context at the same token budget, and its quality scales monotonically with resolution.

  • Evaluation setup: Mage-ViT is evaluated with frozen encoders using linear probes for images and attentive probes for videos, covering representation quality, resolution scaling, and codec robustness.The evaluation spans limited pre-training data, visual-token budgets from 196 to 676 tokens per image, and different codec families.
  • Image representation performance: Mage-ViT achieves 99.33% on CIFAR-10, 82.01% on SUN397, 95.60% on Food-101, and 85.69% on ImageNet-1K, rivaling top-tier encoders.It closely matches SigLIP2 on ImageNet-1K, which reaches 85.92%.
  • Video representation performance: 85.13% on HMDB-51 and 84.83% on Kinetics-400 outperform SigLIP and DINOv3 under 16-frame uniform sampling, while Diving-48 reaches 60.45%.With 64-frame codec-based sparse sampling at the same 4096-token budget, Diving-48 rises to 64.14% (+3.69%) and HMDB-51 reaches 85.17%.
  • Variable-resolution scaling: Mage-ViT’s accuracy improves monotonically through 676 tokens, exceeding 96.1% on Food-101 and 86.3% on ImageNet-1K, unlike fixed-resolution baselines that degrade.SigLIP2 peaks at 256 tokens on ImageNet (~85.9%) and declines to ~84.7% at 676 tokens, while OV-Encoder drops beyond 484 tokens.
  • Cross-codec robustness: Codec-guided patch selection generalizes from HEVC to DCVC-RT without codec-specific retraining, achieving nearly identical average performance while using a smaller average token canvas.Motion magnitude, residual energy, and neural-codec negative log-likelihood estimate local temporal predictability, which Mage-ViT uses independently of codec syntax or latent representations.

5.3 Evaluation of Mage-VL

Mage-VL is competitive with Qwen3-VL-4B on static multimodal tasks and delivers especially strong gains in video understanding and spatial intelligence. These capabilities arise without long VideoQA SFT, while dynamic video training further enhances static 2D/3D spatial reasoning.

  • Static multimodal evaluation: Mage-VL achieves top results on MMStar (67.32), MME-Perception (1709.54), and CV-Bench (87.79), while comprehensively surpassing similarly-sized Phi-4-MM across general VQA.Against Qwen3-VL-4B, the models trade wins within small margins on general VQA.
  • Static multimodal evaluation: Mage-VL outperforms Qwen3-VL on CV-Bench-3D (94.75 vs. 92.30), EmbSpatial (82.67 vs. 77.50), and CrossPoint (80.00 vs. 26.90).Despite using roughly a quarter of its parameters, Mage-VL also surpasses Phi-4-R-V (15B) on most spatial benchmarks.
  • Video understanding: Mage-VL improves over Qwen3-VL-4B on VideoMME (+4.3), MLVU (+7.2), LongVideoBench (+3.5), LVBench (+2.6), NextQA (+3.3), and VideoEval-Pro (+24.5).Qwen3-VL remains stronger on MV-Bench, TempCompass, and MMVU.
  • Video understanding: Mage-VL gains +7.6, +17.1, and +22.5 on Timelens Charades/ActivityNet/QVHighlight, +11.0 on VSI-Bench, and roughly +18 to +20 J&F points on referring-video tracking.It also gains +8.1 on ReasonVOS, with improvements concentrated in localization-heavy regimes.
  • Video understanding: Strong long VideoQA capabilities require no long VideoQA SFT: short video SFT plus detailed video captions enables long VideoQA generalization zero-shot.Dense temporal-visual alignment from high-quality captions allows the base LLM to bypass heavy VideoQA instruction tuning.
  • Video–spatial synergy: Dynamic video sequences with viewpoint transitions and spatial interactions enhance Mage-VL’s 2D/3D spatial reasoning alongside temporal video capabilities.Continuous video streams provide inductive bias for fine-grained geometric perception through perspective shifts, camera trajectories, and object-environment interactions.

5.4 Streaming Video Understanding

Mage-VL is evaluated for both response timing and response quality in continuous video streams. It achieves strong streaming perception accuracy while using codec-native sparse perception and event-driven response generation in live scenarios.

  • Response timing: Mage-VL achieves superior event-triggering performance under strict, zero-tolerance canvas-position matching without domain-specific over-tuning.StreamMind was trained in-distribution on SoccerNet-Caption, while JoyAI is evaluated at 1 Hz with a relaxed ±1-second window.
  • Response quality: Mage-VL achieves 79.84% average accuracy on Real-Time Visual Perception and 48.15% on Backward Tracing on OVO-Bench.Queries use the four most recent frames sampled at 1 fps under SimpleStream’s recent-window protocol.
  • Response quality: 64.00% overall OVO-Bench accuracy establishes Mage-VL as the state-of-the-art streaming architecture score.The score is reported across Real-Time Visual Perception and Backward Tracing.
  • Qualitative illustration: In a live 2026 World Cup broadcast case study, Mage-VL dynamically retains 16×16 patches containing salient motion and novel visual details while suppressing predictable backgrounds.During routine gameplay from 02.16s–24.50s, the event gate stays silent, illustrating codec-native sparse perception and event-driven response generation.

5.5 Efficiency Analysis

Codec-native tokenization dynamically packs motion-salient patches across extended temporal horizons, producing a superior accuracy-efficiency frontier over uniform sampling. It also delivers up to 3.5× wall-clock inference speedups alongside strong accuracy gains.

  • Accuracy-budget scaling: The performance advantage is most pronounced in low-to-medium budget regimes (N ≤16), especially on dense temporal localization benchmarks.The affected benchmarks include TimeLens-ActivityNet, TimeLens-Charades, and TimeLens-QVHighlights, where sparse uniform sampling frequently misses crucial content.
  • Wall-clock efficiency: 57.9% accuracy at 270 s under Mage-VL tc8 on VideoMME outperforms Qwen3-VL’s 59.7% at 463 s in compute time.Mage-VL matches Qwen3-VL’s accuracy at tc16 with 61.8% accuracy at 439 s.
  • Key finding: Finding 5 identifies codec-native inputs as significantly boosting video representation efficiency.The finding summarizes the combined accuracy-budget and wall-clock efficiency results.
  • Accuracy-budget scaling: Codec-native inputs establish a superior accuracy-efficiency frontier over uniform frame sampling under matched visual token budgets.Mage-VL uses adaptive temporal grouping and salient patch packing over an 8× denser source-frame sequence, while the baseline uniformly samples RGB frames.
  • Wall-clock efficiency: Up to 3.5× wall-clock inference speedups result from dynamically packing motion-salient patches across extended temporal horizons.The codec-native visual representation compresses redundant video content before vision encoding, yielding end-to-end speedups alongside higher performance under matched token budgets.

6 Discussions

AI-guided prompt and code co-design improves caption quality across diverse multimodal benchmarks, while replacing visual SFT with pure-text SFT before RL substantially strengthens multimodal performance and reduces RL compute.

  • AI-Driven Data Pipeline Optimization: AI-driven caption optimization consistently improves all nine evaluated multimodal benchmarks, with the largest gains on text- and document-heavy tasks.Improvements include InfoVQA (+5.62), OCRBench (+3.80), ChartQA (+1.52), and DocVQA (+1.08).
  • AI-Driven Data Pipeline Optimization: The optimized pipeline also improves general visual reasoning and perception, including RealWorldQA (+3.79) and MMBench-CN (+1.37), without narrow task-specific overfitting.Multi-dimensional feedback improves caption quality globally, and code-prompt co-design can extend prompt tuning through harness modifications such as video timestamp overlays.
  • Zero-Vision SFT for Multimodal RL: Zero-Vision RL achieves 54.28% versus 48.96% accuracy (+5.33% absolute margin) across 24 image benchmarks, winning 19 tasks and converging in 2,175 versus 4,415 steps.The proposed pipeline replaces visual SFT with pure-text math and code instruction data before applying OpenMMReasoner RL.
  • Zero-Vision SFT for Multimodal RL: Zero-Vision RL gains 7.92% on reasoning tasks and 9.52% on OCR and chart understanding, including WeMath-Loose (+12.29%) and CharXiv-Desc (+27.30%).It also improves General VQA by +0.64% and Others by +6.38%, with PixmoCount reaching +14.61%.
  • Zero-Vision SFT for Multimodal RL: Bypassing visual SFT challenges its presumed necessity and enables compute-efficient multimodal RL for addressing Mage-VL’s agentic capability gap.The motivation identifies omitted RL post-training, limited compute, and the lack of joint text-multimodal mixed pre-training as constraints.

7 Conclusion and Limitations

Mage-VL is a lightweight 4B-parameter streaming vision–language model built around the codec-aligned, dynamically sparse Mage-ViT encoder, a causal decoder, and an event-driven System 1 gate. It matches flagship static-image baselines while improving video and spatial reasoning with up to 3.5× wall-clock speedups, but remains limited in complex agentic workflows and mathematical reasoning.

  • Conclusion: Mage-VL is a lightweight 4B-parameter streaming vision–language model centered on the custom-designed Mage-ViT encoder.Mage-ViT is trained from scratch as a codec-agnostic front-end that dynamically extracts motion- and entropy-rich patches across traditional and neural codecs.
  • Conclusion: Mage-VL combines Mage-ViT with a causal language decoder and event-driven System 1 gate for offline reasoning, spatial grounding, and proactive streaming interaction.
  • Conclusion: Up to 3.5× wall-clock inference speedups accompany uniform improvements across video QA, temporal grounding, and 2D/3D spatial reasoning benchmarks.Mage-VL-4B matches flagship baselines on static image understanding while delivering these gains over dense frame sampling.
  • Limitations and Future Work: Mage-VL remains limited in complex agentic workflows and mathematical reasoning because of insufficient high-quality text data and omitted RL post-training.Future iterations plan Zero-Vision SFT alongside RL and extension to native audio-visual capabilities.

Contributor List

The paper lists 23 contributors, including Senqiao Yang, Kaichen Zhang, Zhaoyang Jia, Jinghao Guo, Yifei Shen, and Xinjie Zhang.

  • Contributor List: The contributor list names 23 researchers, with Senqiao Yang, Kaichen Zhang, Zhaoyang Jia, Jinghao Guo, Yifei Shen, and Xinjie Zhang marked by symbols.The passage also lists Xiaoyi Zhang, Haoqing Wang, Xiao Li, Peng Zhang, Xiang An, Yin Xie, Zhening Liu, Xun Guo, Jiahao Li, Shicheng Zheng, Jinglu Wang, Zongyu Guo, Wenxuan Xie, Zihan Zheng, Yuxuan Luo, Bin Li, and Yan Lu.

A Recaptioning System Prompt

The recaptioning system prompt instructs Qwen3-VL-32B to produce a single, detailed, visually grounded caption that integrates applicable visual, textual, contextual, and reasoning dimensions. An optimized version further emphasizes objective, assertive, coherent prose and supports approximately 350M image–caption pairs for image-caption training.

  • Core caption requirements: The prompt requires one coherent caption integrating all applicable dimensions while ensuring every detail is directly verifiable from the image.It explicitly prohibits separate sections and omits dimensions that do not apply.
  • Coverage dimensions: It covers scene content, actions, environment, emotions, spatial relationships, character identity, OCR, documents, mathematics, coding, science, and other task-specific visual information.The dimensions include object attributes, interactions, composition, text transcription, structured-data interpretation, planning, and perception.
  • Prompt optimization: The optimized prompt was iteratively developed to improve visual coverage, reduce redundant descriptions, preserve coherent organization, and maintain high OCR fidelity.The resulting captions support the approximately 350M image–caption pairs used in image-caption training.
  • Optimized writing style: The optimized prompt directs the model to describe only directly visible facts in natural, flowing prose rather than mechanical attribute lists.Related details should be integrated through clauses, appositives, conjunctions, and smooth transitions.
  • Text and structured information: The prompt requires exact OCR transcription, immediate English translations for non-English text, contextual explanations, and accurate reproduction of data from tables, charts, and labels.It also applies to screenshots, documents, and structured interfaces by requiring descriptions of layout and interface elements.
Loading 2607.24904v1…