ViMU is the first benchmark for evaluating video models on metaphorical and subtextual understanding using hint-free questions grounded in multimodal evidence.
super hub Mixed citations
Qwen3-VL Technical Report
Mixed citation behavior. Most common role is background (47%).
846
Pith papers citing it
Background
47% of classified citations
abstract
We introduce Qwen3-VL, the most capable vision-language model in the Qwen series to date, achieving superior performance across a broad range of multimodal benchmarks. It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B) and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency-quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger pure-text understanding, surpassing comparable text-only backbones in several cases; (ii) robust long-context comprehension with a native 256K-token window for both text and interleaved multimodal inputs, enabling faithful retention, retrieval, and cross-referencing across long documents and videos; and (iii) advanced multimodal reasoning across single-image, multi-image, and video tasks, demonstrating leading performance on comprehensive evaluations such as MMMU and visual-math benchmarks (e.g., MathVista and MathVision). Architecturally, we introduce three key upgrades: (i) an enhanced interleaved-MRoPE for stronger spatial-temporal modeling across images and video; (ii) DeepStack integration, which effectively leverages multi-level ViT features to tighten vision-language alignment; and (iii) text-based time alignment for video, evolving from T-RoPE to explicit textual timestamp alignment for more precise temporal grounding. Under comparable token budgets and latency constraints, Qwen3-VL achieves superior performance in both dense and Mixture-of-Experts (MoE) architectures. We envision Qwen3-VL serving as a foundational engine for image-grounded reasoning, agentic decision-making, and multimodal code intelligence in real-world workflows.
hub tools
citation-role summary
background 121
method 61
baseline 50
dataset 5
other 4
citation-polarity summary
claims ledger
- abstract We introduce Qwen3-VL, the most capable vision-language model in the Qwen series to date, achieving superior performance across a broad range of multimodal benchmarks. It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B) and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency-quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger pure-text understanding, surpassing comparable text-only backbones in several cases; (ii) robust long-con
authors
co-cited works
representative citing papers
CiteVQA requires models to cite specific document regions with bounding boxes alongside answers and finds that even the strongest MLLMs frequently cite the wrong region, with top SAA scores of only 76.0 for closed models and 22.5 for open-source ones.
SenseBench is the first physics-based benchmark with 10K+ instances and dual protocols to evaluate VLMs on remote sensing low-level perception and diagnostic description, revealing domain bias and specific failure modes.
EgoMemReason is a new benchmark showing that even the best multimodal models achieve only 39.6% accuracy on reasoning tasks that require integrating sparse evidence across days in egocentric video.
RuleSafe-VL creates 2,166 rule-conditioned cases from 93 atomic rules and 92 relations across three policy families to diagnose where VLMs fail at rule-based content moderation reasoning.
TraceAV-Bench is the first benchmark for multi-hop trajectory reasoning over long audio-visual videos, showing top models reach only 51-68% accuracy with substantial room for improvement.
PureDocBench shows document parsing is far from solved, with top models at ~74/100, small specialists competing with large VLMs, and ranking reversals under real degradation.
MedHorizon benchmark reveals current multimodal LLMs achieve only 41.1% accuracy on long medical videos due to failures in sparse evidence retrieval and procedural reasoning.
WindowsWorld benchmark shows leading GUI agents achieve under 21% success on multi-application professional tasks, with failures especially on conditional judgment across three or more apps and inefficient execution.
MM-JudgeBench shows substantial cross-lingual performance variance in 22 LVLM judges, with model size and architecture as poor predictors of multilingual robustness.
EVE enables verifiable self-evolution of MLLMs by using a Challenger-Solver architecture to generate dynamic executable visual transformations that produce VQA problems with absolute execution-verified ground truth.
VLMs hallucinate by prioritizing contradictory on-screen text over visual content, addressed via the VisualTextTrap benchmark with 6,057 human-validated samples and the VTHM-MoE dual-encoder framework using dimension-specific experts and adaptive routing.
RefereeBench shows that even the strongest video MLLMs reach only around 60% accuracy on multi-sport refereeing tasks and struggle with rule application and temporal grounding.
VLM-UnBench demonstrates that prompt-based training-free unlearning in VLMs leaves forget accuracy near the no-instruction baseline except under oracle conditions that reveal the target concept.
ScreenParse dataset and ScreenVLM model deliver dense screen parsing that outperforms larger VLMs on PageIoU and transfers to better UI grounding.
GUIGuard-Bench is a new benchmark with annotated GUI screenshots that measures privacy recognition, planning fidelity under protection, and utility impact for trajectory-based GUI agents.
Cultural commonsense in India is mostly regional, with only 39.4% agreement across five regions, and LLMs achieve just 13.4-20.9% accuracy while over-representing North and Central areas.
Molmo2 delivers state-of-the-art open-weight video VLMs with new grounding datasets and training methods that outperform prior open models and match or exceed some proprietary ones on pointing and tracking tasks.
S1-MMAlign is a new large-scale dataset of 15.5 million semantically enhanced scientific image-text pairs created via an AI recaptioning pipeline to improve multimodal understanding.
ToG-Bench is the first benchmark for task-oriented spatio-temporal video grounding in egocentric videos, with explicit-implicit dual grounding and one-to-many object scenarios across 100 ScanNet clips and 2704 instructions.
GaussDet enables open-vocabulary and referring segmentation in 3D Gaussians by learning instance features and aggregating votes from 2D detectors, improving referential grounding by 16.7% mIoU in zero-shot setting.
Goku supplies a 2M-scale dataset, synthesis pipeline, decoupled dual-branch model, and 1000-case benchmark for multi-task instruction-based video editing, reporting up to 8% gains in instruction following.
OmniCoT is a new panoramic reasoning benchmark with 6.7K eval, 1K real, and 14.3K training examples plus a two-stage SFT+GRPO training method to enforce global 360-degree consistency.
MuseBench shows state-of-the-art MLLMs achieve only 48.29% accuracy on intent-level audiovisual arts understanding versus 87.18% for human experts.
citing papers explorer
-
RoboGaze: Evaluating Robot World Models via Structured Vision-Language Analysis
RoboGaze presents a structured multi-agent VLM pipeline and robotics-specific error taxonomy that improves video evaluation metrics by up to 43 F1 points over zero-shot baselines on a 382-clip dataset.
-
POINav: Benchmarking and Enhancing Final-Meters Arrival in Real-World Vision-Language Navigation
POINav-Bench provides the first high-fidelity real-world benchmark for POI-goal VLN using 3DGS reconstructions of 126k m² with 163 POIs, supported by a Brain-Action framework and 70K real signage-entrance dataset.
-
MemCompiler: Compile, Don't Inject -- State-Conditioned Memory for Embodied Agents
MemCompiler reframes memory use as state-conditioned compilation, delivering relevant guidance via text and latent channels to improve embodied agent performance up to 129% and cut latency 60% versus static injection.
-
OA-WAM: Object-Addressable World Action Model for Robust Robot Manipulation
OA-WAM uses persistent address vectors and dynamic content vectors in object slots to enable addressable world-action prediction, improving robustness on manipulation benchmarks under scene changes.
-
Action Agent: Agentic Video Generation Meets Flow-Constrained Diffusion
Action Agent pairs LLM-driven video generation with a flow-constrained diffusion transformer to produce velocity commands, raising video success to 86% and delivering 64.7% real-world navigation on a Unitree G1 humanoid.
-
RoboWM-Bench: A Benchmark for Evaluating World Models in Robotic Manipulation
RoboWM-Bench evaluates video world models by converting their manipulation video predictions into executable actions validated in simulation, showing that visual plausibility does not guarantee physical executability.
-
ViVa: A Video-Generative Value Model for Robot Reinforcement Learning
ViVa turns a video generator into a value model for robot RL that jointly forecasts future states and task value, yielding better performance on real-world box assembly when integrated with RECAP.
-
JailWAM: Jailbreaking World Action Models in Robot Control
JailWAM is the first dedicated jailbreak framework for World Action Models, achieving 84.2% attack success rate on LingBot-VA in RoboTwin simulation and enabling safety evaluation of robotic AI.
-
QuadAgent: A Responsive Agent System for Vision-Language Guided Quadrotor Agile Flight
QuadAgent uses an asynchronous multi-agent architecture with an Impression Graph for scene memory and vision-based avoidance to enable training-free vision-language guided agile quadrotor flight, outperforming baselines in simulations and achieving real-world speeds up to 5 m/s.
-
PhysMem: Scaling Test-Time Memory for Embodied Physical Reasoning
PhysMem enables VLM-based robot planners to learn and verify physical properties through test-time interaction and hypothesis testing, raising success on a brick insertion task from 23% to 76%.
-
GROW$^2$: Grounding Which and Where for Robot Tool Use
GROW² hierarchically grounds open-world tool affordances by using VLMs for semantic selection of objects and parts followed by geometric localization with vision foundation models.
-
Training Vision-Language-Action Models with Dense Embodied Chain-of-Thought Supervision
ZR-0 trains a dual-stream VLA model on 60 million frames with dense ECoT annotations so that high-level reasoning transfers across single-arm, bimanual, and humanoid embodiments while skipping reasoning at inference.
-
SA-VLA: State-aware tokenizer for improving Vision-Language-Action Models' performance
SA-VLA adds state conditioning to VQ-based action tokenization in VLA policies, expanding each discrete token's effective support to state-dependent actions and raising average success rates from 0.29 to 0.56 on 12 sim tasks and 0.15 to 0.33 on 3 real tasks.
-
Keypose Exploration: Efficient Automatic Trajectory Labelling and Cross-Embodiment Policy Transfer
An automatic single-demo VLM trajectory labelling pipeline enables keypose-guided diffusion policies that match baseline performance and show preliminary benefits for cross-embodiment transfer on robomimic tasks.
-
LocalNav: Distilling Frontier VLMs and Embodied RL for On-Device Object Goal Navigation
Distillation from frontier VLMs plus E-RLVR regularization produces a 4B local model that achieves 34.5% SR on OVON while cutting inference latency by 82.8%.
-
Direct Action-Head Injection of A Grounded 3D Point Unlocks Spatial and Task Generalization
Direct 3D point grounding injected into the action head via a two-layer MLP and adaptive layer norm boosts VLA success rates by 32-46 points on spatial and task perturbations in LIBERO-PRO.
-
LA4VLA: Learning to Act without Seeing via Language-Action Pretraining
LA4VLA creates a 33K language-action dataset from existing demos and shows that pretraining on language-action pairs before or alongside vision-language-action training boosts success rates in sim and real robot tasks.
-
SpaceVLN: A Zero-Shot Vision-and-Language Navigation Agent with Online Spatial Cognitive Memory and Reasoning
SpaceVLN proposes a stagewise closed-loop framework using Spatial Cognitive Memory and Spatial-CoT for zero-shot vision-and-language navigation and object-goal navigation, reporting SOTA results on R2R-CE, RxR-CE, GN-Bench, and HM3D-OVON plus real-robot tests.
-
Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring
Hide-and-Seek uses contrastive objectives on trajectories to localize failure signals in VLA models from trajectory-level supervision alone.
-
PrimitiveVLA: Learning Reusable Motion Primitives for Efficient and Generalizable Robotic Manipulation
PrimitiveVLA introduces a primitive-centric framework that disassembles demonstrations into reusable motion primitives during fine-tuning and assembles them at inference via VLM planner and LLM switch for improved data efficiency and zero-shot generalization in robotic manipulation.
-
FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies
FineVLA unifies robot datasets into 47k fine-grained trajectories, adds a VLM annotator and benchmark, and shows that mixing fine-grained and goal-level instructions improves steerable control without hurting task success.
-
Semantically Structured Mixture-of-Experts for Compositional Robotic Manipulation
SMoDP routes action chunks in a diffusion policy to semantically specialized experts via a VLM-supervised skill predictor and dual contrastive alignment, achieving better efficiency and compositional transfer than baselines.
-
SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework
SUGAR turns diverse human videos into deployable humanoid loco-manipulation policies via automated prior extraction, physics refinement, and hierarchical distillation, showing scaling with data volume and zero-shot real-world transfer on six tasks.
-
How to Instruct Your Robot: Dense Language Annotations Power Robot Policy Learning
DeMiAn re-annotates robot and egocentric videos with VLM-generated dense labels across motion, scene, pose, and reasoning aspects, then uses a learned instructor to boost policy success by 5 points on RoboCasa over task-only baselines.
-
Guide, Think, Act: Interactive Embodied Reasoning in Vision-Language-Action Models
GTA-VLA conditions VLA models on user spatial priors to produce a unified spatial-visual chain-of-thought, reaching 81.2% success on SimplerEnv WidowX and improving performance under out-of-distribution shifts.
-
SECOND-Grasp: Semantic Contact-guided Dexterous Grasping
SECOND-Grasp integrates semantic contact proposals from vision-language reasoning with geometric refinement to achieve 98%+ lifting success and improved intent-aware grasping on seen and unseen objects.
-
HarmoWAM: Harmonizing Generalizable and Precise Manipulation via Adaptive World Action Models
HarmoWAM unifies predictive and reactive control in world action models via an adaptive gating mechanism to deliver improved zero-shot generalization and precision in robotic manipulation.
-
ALAM: Algebraically Consistent Latent Action Model for Vision-Language-Action Models
ALAM introduces algebraic consistency regularization on latent action transitions from videos, raising VLA success rates from 47.9% to 85.0% on MetaWorld MT50 and 94.1% to 98.1% on LIBERO.
-
Toward Visually Realistic Simulation: A Benchmark for Evaluating Robot Manipulation in Simulation
VISER is a new visually realistic simulation benchmark for robot manipulation tasks that uses PBR materials and MLLM-assisted asset generation, achieving 0.92 Pearson correlation with real-world policy performance.
-
Task-Aware Scanning Parameter Configuration for Robotic Inspection Using Vision Language Embeddings and Hyperdimensional Computing
ScanHD achieves 92.7% exact accuracy and 98.1% Win@1 accuracy in recommending discrete scanning parameters from instructions and images on a new real-world dataset.
-
MolmoAct2: Action Reasoning Models for Real-world Deployment
MolmoAct2 is an open VLA model that outperforms baselines like Pi-05 on 7 benchmarks and whose backbone surpasses GPT-5 on 13 embodied-reasoning tasks through new datasets, specialized training, and architecture changes for lower latency.
-
Affordance Agent Harness: Verification-Gated Skill Orchestration
Affordance Agent Harness is a verification-gated orchestration system that unifies skills via an evidence store, episodic memory priors, an adaptive router, and a self-consistency verifier to improve accuracy-cost tradeoffs in open-world affordance grounding.
-
FreeOcc: Training-Free Embodied Open-Vocabulary Occupancy Prediction
FreeOcc enables training-free open-vocabulary 3D occupancy prediction from RGB-D sequences by combining SLAM, dense Gaussian maps, off-the-shelf vision-language models, and probabilistic projection, achieving over 2x gains on benchmarks and zero-shot transfer to novel scenes.
-
MotuBrain: An Advanced World Action Model for Robot Control
MotuBrain jointly models video and action via a three-stream Mixture-of-Transformers UniDiffuser to reach 95.8-96.1% success on RoboTwin 2.0 benchmarks, top EWMScore, and fast 11 Hz inference while adapting to new robots with 50-100 trajectories.
-
BridgeACT: Bridging Human Demonstrations to Robot Actions via Unified Tool-Target Affordances
BridgeACT learns robot manipulation from human videos alone by predicting task-relevant grasp regions and 3D motion affordances that map directly to robot controllers.
-
Long-Horizon Manipulation via Trace-Conditioned VLA Planning
LoHo-Manip enables robust long-horizon robot manipulation by using a receding-horizon VLM manager to output progress-aware subtask sequences and 2D visual traces that condition a VLA executor for automatic replanning.
-
ST-$\pi$: Structured SpatioTemporal VLA for Robotic Manipulation
ST-π structures VLA models by having a spatiotemporal VLM produce causally ordered chunk-level prompts that guide a dual-generator action expert to jointly handle spatial and temporal control in robotic manipulation.
-
CLASP: Closed-loop Asynchronous Spatial Perception for Open-vocabulary Desktop Object Grasping
CLASP achieves 87% success in open-vocabulary desktop grasping via dual-pathway perception, asynchronous closed-loop evaluation, and automated multimodal data synthesis.
-
HEX: Humanoid-Aligned Experts for Cross-Embodiment Whole-Body Manipulation
HEX introduces a state-centric framework with humanoid-aligned representations and mixture-of-experts proprioceptive prediction for coordinated whole-body control on bipedal humanoids.
-
Uncovering Linguistic Fragility in Vision-Language-Action Models via Diversity-Aware Red Teaming
DAERT generates diverse adversarial instructions via a uniform policy in RL to drop VLA task success rates from 93.33% to 5.85% on benchmarks with models like π0 and OpenVLA.
-
RoboPlayground: Democratizing Robotic Evaluation through Structured Physical Domains
RoboPlayground reframes robotic manipulation evaluation as a language-driven process over structured physical domains, letting users author varied yet reproducible tasks that reveal policy generalization failures.
-
FSUNav: A Cerebrum-Cerebellum Architecture for Fast, Safe, and Universal Zero-Shot Goal-Oriented Navigation
FSUNav's dual brain-inspired modules achieve state-of-the-art zero-shot goal navigation across heterogeneous robots with improved speed, safety, and generalization.
-
DIAL: Decoupling Intent and Action via Latent World Modeling for End-to-End VLA
DIAL decouples intent from action in end-to-end VLAs using a latent visual foresight bottleneck and two-stage training, reaching SOTA on RoboCasa with 10x fewer demonstrations and zero-shot real-world transfer.
-
Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons
Robometer combines intra-trajectory progress supervision with inter-trajectory preference supervision on a 1M-trajectory dataset to learn more generalizable robotic reward functions than prior methods.
-
Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning
R&B-EnCoRe uses self-supervised importance-weighted variational inference to distill action-predictive reasoning datasets that improve VLA performance on manipulation, navigation, and driving tasks without external verifiers.
-
Clutter-Robust Vision-Language-Action Models through Object-Centric and Geometry Grounding
OBEYED-VLA improves VLA robustness in cluttered real-world manipulation by disentangling perception into VLM-based object-centric grounding and geometry-aware stages, then fine-tuning the policy only on single-object demonstrations.
-
FutureNav: Unified World-Action Modeling for Vision-and-Language Navigation
FutureNav proposes a 4B-scale VLM that jointly optimizes action prediction, inverse/forward dynamics, and future state generation for VLN and reports SOTA results on multiple benchmarks.
-
OpenSPM: An Environment-Transferable Robotic Key Spatial Pose Memory and Closed-Loop High-Frequency Flow-Matching Action Generation Model
OpenSPM extracts key spatial poses from demonstrations as transferable memory and generates high-frequency actions via flow-matching, achieving 85.6% success rate on ten LIBERO-GOAL tasks at 1033.3 Hz.
-
S$^2$-VLA: State-Space Guided Vision-Language-Action Models for Long-Horizon Manipulation
S²-VLA uses a state-space model to maintain a belief state that produces dynamic gating weights for fusing visual, language, and action features, claiming better long-horizon manipulation than 7B models with only 2B parameters.
-
Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models
Embodied-R1.5 is an 8B EFM achieving SOTA on 16 of 24 embodied VLM benchmarks, fine-tunable to outperform leading VLAs, with claimed zero-shot real-robot generalization.