Tagged articles
Vision-Language Models
41 articles · Page 1 of 1
HUST and Horizon Robotics propose Multimodal Flow, a fully continuous flow-matching architecture that unifies language and vision in embedding spaces without tokenization or dual objectives, achieving state-of-the-art results with a 1.6B model trained on only 150B tokens.
Continuous RepresentationsFlow MatchingGenerative AI
0 likes · 11 min read

Researchers from Duke, Google, USC, UPenn, UNC and others present a systematic survey of memory mechanisms in Vision-Language Models, introducing a 4D taxonomy (When, Where, What, How) covering 300+ papers, 70 benchmarks, 8 memory forms, and 8 application scenarios including long-video understanding and embodied navigation.
4D TaxonomyEmbodied NavigationLong Video Understanding
0 likes · 12 min read

Shanghai Jiao Tong University researchers introduce ZwZ2 and the EvidenceConnect framework, which synthesizes 20.4K single-image training samples requiring models to connect evidence across multiple regions, yielding 4B, 9B, and 27B models that surpass base models on 25 benchmarks and transfer gains to multi-image, spatial reasoning, and agent tasks.
Data SynthesisEvidenceConnectGRPO
0 likes · 14 min read


JD TechOct 27, 2026 · Artificial Intelligence
JD Retail's AI Infra team introduces FlowTrain, a decoupled training framework for vision-language models that eliminates batch-level synchronization via a unified memory pool, heterogeneous parallel allocation, and dynamic packing scheduler, achieving up to 1.7x throughput improvement and over 50% MFU on domestic super-node clusters, validated on industrial e-commerce models.
FlowTrainMFUVision-Language Models
0 likes · 15 min read

University of Pittsburgh researchers show that vision-language models internally represent 3D scene topology in middle-layer latent spaces, revealed by separating appearance attributes and visualizing spatial structure; Dirichlet energy quantifies neighborhood preservation, and fine-tuning with this loss improves spatial reasoning on benchmarks like VSI-Bench.
3D Scene UnderstandingDirichlet EnergyInterpretability
0 likes · 12 min read

Show-Harness introduces an Embodied Harness — a semantic action interface enabling vision-language models to control diverse robots through interpretable, composable actions like move, rotate, grasp, and release, achieving 89–100% zero-shot success across tasks, environments, and embodiments while unifying human teleoperation, GUI agents, and VLM control via the GUMI interface.
Embodied AIGUMISemantic Action Interface
0 likes · 16 min read

This survey reviews vision-language models for egocentric video, analyzing why standard VLMs fail on first-person footage and how hand-object interaction, spatiotemporal reasoning, graph-based representations, and semantic alignment bridge the gap toward embodied AI applications.
Embodied AIGraph Neural NetworksHand-Object Interaction
1 likes · 23 min read

Meta's HumanCLAW benchmark evaluates nine vision-language models on embodied action intelligence, separating high-level decisions from low-level control; the best model completes full interactions at just 16.8% success, revealing critical gaps in embodied self-awareness and closed-loop reasoning.
Action IntelligenceEmbodied AIHumanCLAW
0 likes · 12 min read

EASE introduces evidence-anchored spatial attention supervision to multimodal reinforcement learning with verifiable rewards, using annotated evidence bounding boxes to guide model attention toward relevant image regions during training, improving accuracy on visual reasoning benchmarks without inference overhead.
Attention MechanismEMNLP 2026Evidence Grounding
0 likes · 13 min read


AntTechSep 2, 2026 · Artificial Intelligence
This article summarizes four ECCV 2026 papers: a dynamic cross-layer injection framework for deep vision-language fusion, an event-augmented VLA model enabling robot operation in extreme darkness and blur, a predictive differentiable rendering method using 2D Gaussians for high-fidelity video prediction, and a data-scaling approach for high-resolution weather forecasting that demonstrates clear scaling laws.
Data ScalingECCV 2026Embodied AI
0 likes · 10 min read

The article introduces Code-as-World, a framework that turns visual observations into executable, verifiable code describing objects, dynamics, and appearance, enabling agents to iteratively propose, simulate, and refine world hypotheses and achieving state‑of‑the‑art results on physical reasoning benchmarks.
Agentic DiscoveryCode-as-WorldExecutable World
0 likes · 13 min read

The article identifies the hidden “privilege illusion” that degrades on‑policy distillation when privileged information is injected, and introduces Dual On‑policy Distillation (DOPD), a dynamic two‑stream approach that separates true ability gaps from information gaps, achieving superior performance and stability across LLM and VLM benchmarks.
DOPDOPDVision-Language Models
0 likes · 13 min read

This paper introduces a semantic robustness certification framework for vision‑language models that leverages paired text prompts as semantic proxies to define a continuous transformation in the shared embedding space, derives closed‑form interval bounds where predictions remain unchanged, and validates the method on CLIP ViT‑B/32 with both synthetic and real‑world datasets.
CLIPVision-Language Modelsembedding geometry
0 likes · 13 min read

The paper identifies Contextual Variable Overestimation in vision‑language models, introduces the MACP multi‑country ad preference dataset, proposes the three‑gate GeoReward framework to restore sensitivity to sparse country cues, and demonstrates superior accuracy, sensitivity, and controllable ad generation across ten markets.
Cross-Market PreferenceDataset MACPGeoReward
0 likes · 19 min read

The paper introduces Deep Pre‑Alignment (DPA), a novel Vision‑Language Model architecture that inserts a perceiver VLM to pre‑align visual features with the LLM’s text space, reducing alignment cost, preserving language ability, and delivering consistent multimodal performance gains across multiple benchmarks with minimal inference overhead.
Deep Pre-AlignmentLLMPerceiver VLM
0 likes · 10 min read

The article introduces Visual Para-Thinker, a parallel reasoning framework for large vision‑language models that mitigates attention drift and visual hallucination by employing path‑aware attention, learnable parallel rotary position embeddings, and hybrid block‑and‑scan visual token partitions, and validates the approach with extensive multimodal benchmarks.
LPRoPEMultimodal BenchmarksParallel Attention
0 likes · 10 min read

Meta and Princeton introduce VLM³, a unified vision‑language framework built on Qwen3‑VL‑4B that models depth estimation, object‑level 3D understanding, pixel matching and camera pose estimation without extra encoders, achieving up to 0.90 depth accuracy and outperforming larger specialist models on multiple benchmarks.
3D PerceptionQwen3-VL-4BVision-Language Models
0 likes · 15 min read

The article reviews three major recent breakthroughs in computer vision—self‑supervised visual foundation models, feed‑forward 3D reconstruction, and unified multimodal models—detailing their underlying methods, key papers, performance characteristics, and practical implications for real‑world AI applications.
3D reconstructionComputer VisionMultimodal Models
0 likes · 22 min read

The article introduces Visual Para-Thinker, the first parallel reasoning framework tailored for large‑scale vision‑language models, explains its block and scan visual path divisions, details the Path‑aware Attention and Learnable Parallel Rotary Position Embedding mechanisms, and presents experimental results showing significant gains on visual perception benchmarks.
LPRoPEPath-aware AttentionVision-Language Models
0 likes · 9 min read

Laser introduces a latent‑superposition paradigm that replaces step‑by‑step token prediction with dynamic windowed alignment, achieving over 97% token‑consumption reduction, new SOTA performance on six visual benchmarks, and improved interpretability for multimodal large models.
ACL 2026Dynamic Window AlignmentLatent Superposition
0 likes · 13 min read

By learning a compact latent‑action space from paired image‑text and large‑scale text data, the authors reduce the RL search space from a vocabulary of over 150 k tokens to a 128‑codebook, enabling more efficient fine‑tuning of multimodal conversational agents and achieving consistent gains across several RL algorithms.
Reinforcement LearningVision-Language Modelsdialogue agents
0 likes · 11 min read

The article analyzes the prompt‑sensitivity problem of zero‑shot classification in vision‑language models, critiques class‑agnostic prompt weighting, and presents CARPRT—a training‑free, black‑box compatible method that reweights prompts per class using similarity scores and pseudo‑labels, achieving consistent gains across datasets and model architectures.
Black-Box OptimizationClass-Aware ModelingPrompt Reweighting
0 likes · 11 min read

Apple’s recent AI paper probes whether large vision‑language models truly forget user data by examining residual streams and final logits, revealing that hidden image attributes persist in top‑k outputs and exposing significant privacy and security risks.
AI securityVision-Language Modelsinformation bottleneck
0 likes · 11 min read

This paper introduces a novel framework that constructs and aligns dual visual‑textual trees on heterogeneous hyperbolic manifolds, addressing asymmetric modality alignment in hierarchical classification tasks and achieving state‑of‑the‑art performance on benchmarks such as CIFAR‑100, ImageNet and Rare Species datasets.
Vision-Language Modelscross-attentionhierarchical classification
0 likes · 8 min read

The paper introduces TaYS (Think‑as‑You‑See), a streaming chain‑of‑thought framework that replaces the traditional “watch‑then‑think” video inference pipeline with a parallel, real‑time “watch‑and‑think” approach, dramatically reducing latency and improving accuracy on complex video reasoning tasks.
Dual KV-CacheTaYSVision-Language Models
0 likes · 8 min read

An analysis of recent Vision‑Language models shows that removing interleaved images has minimal impact on benchmark performance, suggesting that better priors from RL fine‑tuning and effective context management are the key drivers of success.
Attention RolloutCropping MethodsInterleaved Images
0 likes · 8 min read

Vision‑language models turn continuous images into discrete tokens through patch extraction, encoding, and projection, enabling Transformers to reason jointly over vision and text, but this compression introduces limits in spatial reasoning, counting, and resolution sensitivity that users must understand.
Self-AttentionVision-Language Modelscounting
0 likes · 22 min read

This week’s AI paper roundup highlights five recent studies: a construction‑site vision‑language dataset and safety inspection tasks, a deep CORAL method for unsupervised domain adaptation, the discovery of a new family of unstable singularities in nonlinear PDEs, a reinforcement‑learning approach that boosts reasoning in large language models, and the PANORAMA architecture for omnidirectional vision in embodied AI.
Construction SafetyOmnidirectional VisionPDE Research
0 likes · 6 min read


AntTechSep 25, 2025 · Artificial Intelligence
The ICCV 2025 live session will deep‑dive into two cutting‑edge papers—PixTrace with CopyNCE for precise image copy detection and Skip‑Vision for dramatically faster training and inference of vision‑language models—showcasing their methods, results, and real‑world impact.
Computer VisionICCV 2025Model Acceleration
0 likes · 5 min read

This article analyzes the MolmoAct large action model, detailing its three‑stage perception‑planning‑control architecture, novel depth‑aware tokenization, extensive pre‑training and fine‑tuning pipelines, and benchmark results that demonstrate superior efficiency and generalization over prior vision‑language‑action systems.
Model TrainingMolmoActRobotics
0 likes · 12 min read

This study introduces EvoVLMA, an evolutionary vision-language model adaptation framework that automatically searches training-free VLM adaptation algorithms using a two-stage LLM-guided evolution, demonstrating superior performance—such as a 1.91 % accuracy gain on 8-shot image classification—and releasing the code publicly.
Evolutionary AlgorithmsLLMVision-Language Models
0 likes · 5 min read

This article compares two ways to add vision to large language models—training a native multimodal model from scratch or attaching a visual module to a pretrained LLM—then details the VoRA approach that uses LoRA adapters to inject visual knowledge without extra inference cost.
ChameleonLLaVALoRA
0 likes · 7 min read

The paper introduces ATPrompt, a method that inserts fixed attribute tokens into learnable prompts for CLIP‑style vision‑language models, enabling the soft prompts to capture generic attribute representations and significantly improve base‑to‑novel generalization without extra regularization losses.
ATPromptVision-Language ModelsZero-shot Generalization
0 likes · 20 min read

The article examines how multimodal AI models repeatedly miscount a six‑finger image, explores the underlying bias revealed in the paper “Vision Language Models are Biased,” and warns that such prior‑knowledge‑driven errors can have serious safety implications in real‑world applications.
AI biasVision-Language Modelsmodel hallucination
0 likes · 10 min read

The CVPR 2025 awards spotlight groundbreaking work—from the VGGT transformer that predicts full 3D scenes in a single feed‑forward pass to neural inverse rendering that reconstructs geometry from time‑resolved light—offering a comprehensive view of emerging trends, novel architectures, and performance breakthroughs across computer‑vision research.
3D reconstructionCVPR 2025Neural Rendering
0 likes · 11 min read


AIWalkerApr 8, 2025 · Artificial Intelligence
AgenticIR combines visual language models and large language models in a multi‑stage reasoning workflow—perception, planning, execution, reflection, and adjustment—to evaluate, plan, and iteratively apply specialized restoration tools, achieving superior results on complexly degraded images compared to baseline methods.
Agentic SystemsICLR 2025Vision-Language Models
0 likes · 15 min read


AntTechMar 18, 2025 · Artificial Intelligence
Researchers from Ant Insurance and Zhejiang University propose MoLE, a Mixture of Layer Experts decoding method that reduces hallucinations in large vision‑language models, demonstrating state‑of‑the‑art performance on LVLM benchmarks and enabling reliable end‑to‑end medical‑record‑to‑claim automation.
AIMixture of ExpertsVision-Language Models
0 likes · 7 min read


AIWalkerMar 17, 2025 · Artificial Intelligence
The paper introduces UNIFIEDREWARD, the first unified reward model for multimodal understanding and generation that supports pairwise ranking and pointwise scoring, builds a 236K human‑preference dataset across image and video tasks, and uses DPO to align VLMs and diffusion models, achieving significant performance gains on both image and video benchmarks.
Direct Preference OptimizationImage GenerationUnified Reward
0 likes · 19 min read


AIWalkerFeb 15, 2025 · Artificial Intelligence
Janus-Pro, the open‑source successor to Janus, introduces a decoupled visual encoder and scaled training data to boost both multimodal understanding and text‑to‑image generation, achieving state‑of‑the‑art results on benchmarks such as GQA, GenEval and DPG‑Bench.
Janus-ProModel ScalingVision-Language Models
0 likes · 13 min read


NewBeeNLPNov 11, 2024 · Artificial Intelligence
This article surveys ten recent multimodal large language model papers, covering vision representation laws, a stricter instruction benchmark, safety impacts of visual adaptation, the Mini‑Gemini architecture, automatic pruning, vision capability boosting, long‑context transfer, efficient token sparsification, math reasoning, and hallucination mitigation.
EfficiencyTraining StrategiesVision-Language Models
0 likes · 18 min read

Over the past year, Vision‑Language Models have surged from early multimodal experiments to competitive open‑source systems rivaling GPT‑4, driven by higher‑resolution processing, richer vision encoders, better projection layers, and larger curated datasets, yet they still face evaluation difficulties, hallucinations, speed limits, and limited multimodal output.
Computer VisionVision-Language Modelsdeep learning
0 likes · 24 min read
