时间线 · 通用视觉自监督研究报

7 月 21 日

P0AV-JEPAarXiv new/cross; audio-visual JEPA SSL

P0Distributional Matching for Vector QuantizationarXiv new; visual tokenizer / VQ representation

P1Ask Twice, Look TwicearXiv new/cross; VLM prompt/attention mechanism

P1How Do VLMs Fail? Vision-OperationarXiv new; ACM MM 2026; VLM failure mechanism

P1Von Mises-Fisher Mixture Model witharXiv new; ICML 2026; VLM test-time adaptation

P2Audio-Visual FlamingoarXiv cross; open long audio-visual MLLM

P2Efficient Frame Selection for LongarXiv new; training-free long-video frame selection

P2MotionForesightarXiv new; video priors to 3D motion forecasting

P2MDNDarXiv new; AAAI 2026 supplementary complete version

P2IoUPDarXiv new; privileged distillation for MLLM grounding

P3Reasoning-Guided Part-Level Visual Grounding viaarXiv new/cross; ECCV 2026; part grounding RL

P3MLLM-DataEnginearXiv new/cross; ICME 2026; MLLM data engine

扫读Are All Tokens Necessary forarXiv new; ViT/foundation-model token reduction

7 月 20 日

状态补录Emergent Region-Level Facial Correspondence in周一早间复核；7 月 18 日 P0；arXiv new/cross 同批

状态补录SeeSE3周一早间复核；7 月 18 日 P0；arXiv new/cross 同批

状态补录Self-Supervised Learning as Discrete CommunicationOpenReview/ICML 状态背景；旧 arXiv revision；不计新增

7 月 19 日

状态补录Emergent Region-Level Facial Correspondence in周末复核；昨天 P0；arXiv new/cross 同批

状态补录SeeSE3周末复核；昨天 P0；arXiv new/cross 同批

状态补录Self-Supervised Learning as Discrete CommunicationOpenReview/ICML 状态核查；ICML 2026 Poster；旧 arXiv revision

7 月 18 日

P0Emergent Region-Level Facial Correspondence inarXiv new/cross; frozen self-supervised VFM correspondence

P0SeeSE3arXiv new/cross; VFM latent geometry probe

P1AlphaWiSEarXiv new/cross; continual multimodal representation

P1FlashDecoderarXiv new/cross; CVPR 2026; video latent decoder

P1U-shaped Multi-granularity Learning for Vision-LanguagearXiv new; VLM prompt representation

P2VideoChat3arXiv new; open video MLLM / I3D-ViT

P2Hierarchical Denoising For Multi-Step VisualarXiv new; hierarchical video diffusion reasoning

P2SymbalarXiv new/cross; ICML 2026; caption-data QA

P2GeoDetectarXiv new/cross; ECCV 2026; VLP embedding geometry

P2Beyond Single ExpertarXiv new; multi-VFM prior fusion

P3Uni-AdaVDarXiv new; visual generation concept erasure

P3GlobalForgearXiv new; robust generated-image detection

扫读VideoSEMAarXiv new/cross; efficient video attention

7 月 17 日

P0VideoRAEarXiv new; VFM representation autoencoder

P0Fine-grained CLIP fine-tuning with self-annotatedarXiv new; self-annotated CLIP region alignment

P1AspectCLIParXiv new/update; PRCV 2026; CLIP consistency regularization

P1Attention-Free and Lightweight Token ReductionarXiv new; VLM visual token reduction

P1The Hyperspherical Geometry of CLIParXiv new; CLIP latent geometry

P2Beyond Color GeometryarXiv new/cross; ViT/MAE color representation analysis

P2DP-BOAarXiv new; ECCV 2026; category discovery

P2Accuracy Without GroundingarXiv new/cross; ACM MM 2026; video LLM diagnostic

P2Groc-POarXiv new/cross; ACM MM 2026; grounded MLLM alignment

P3CLIP-Guided Label-Free Discriminative Region ScoringarXiv new; CLIP/SAM label-free region scoring

P3ScanFocusarXiv new/cross; ECCV 2026; spatio-temporal video grounding

P3Symbiosis-Inspired Knowledge Distillation for IncrementalarXiv new/cross; ICML 2026; knowledge distillation for detection

扫读VGIF-ScorearXiv new; PRCV 2026; video generation evaluation

7 月 16 日

P0Contrastive-Augmented Flow Matching for Style-ContentarXiv new; contrastive + flow representation

P0VisCoarXiv new; VLM visual token self-compression

P1Let RGB Be the Language of VisionarXiv new; unified RGB-to-RGB vision formulation

P1UMSSarXiv new; unsupervised multimodal semantic segmentation

P1MobileSAM2arXiv new + ECCV 2026; SAM2 distillation

P2AVQ-AttentionarXiv new + ECCV 2026; efficient attention

P2Inhibited Self-AttentionarXiv new; ViT attention/object-centric robustness

P2Gaussian Mixture Modeling for Event-AwarearXiv new + PRCV 2026; long-video visual allocation

P2UniVRarXiv new; pure-visual reasoning/RL

P2DynTracearXiv new + ACM MM 2026; 4D spatiotemporal reasoning

P3CoRearXiv new + ACM MM 2026; cross-image VLM reasoning

P3The Seriality Gap in VideoarXiv new; video diffusion analysis

扫读Visual Access Boundaries in Vision-LanguagearXiv new; VLM visual-token access diagnostic

7 月 15 日

P0Slot-RAEarXiv new

P0On the modality gap andarXiv new/cross; CLIP/InfoNCE theory

P1Lifelong RepresentationsarXiv new + IEEE EAIS 2026

P1Self-supervised Automatic MattingarXiv new

P1Spectral Heat Flow for ConservativearXiv new + ICML 2026 / OpenReview

P1Generalize LMMs to Versatile VisualarXiv new + ECCV 2026

P2FoundationGeoarXiv new + ECCV 2026

P2Label-Free Target-Domain Adaptation for UnconstrainedarXiv new + ACM MM 2026

P2MonkeyOCRv2arXiv new

P3Unified Backbone Refinement for DiffusionarXiv new + ECCV 2026

扫读Single-Teacher View AugmentationarXiv new

7 月 14 日

P0Video Generation Models are General-PurposearXiv new/cross + ECCV 2026

P0SigLIP-HD by Fine-to-Coarse SupervisionarXiv new + ICLR 2026 / OpenReview

P0Scalable Visual Pretraining for LanguagearXiv new/cross

P1Probing Diffusion Denoising Dynamics forarXiv new

P1Vision Transformers Learn Gestalt-Like Figure-GroundarXiv new

P1Subtoken Vision Transformer for Fine-grainedarXiv new

P1Similarity search generalisation in contrastivearXiv new/cross; contrastive theory

P2What VGGT Knows About OverlaparXiv new/cross; VFM probing

P2Robustifying Vision-Language Models via Test-TimearXiv new/cross + ICML 2026 regular

P2On Locality and Length GeneralizationarXiv new/cross + ECCV 2026

P3Weaving Light and TimearXiv new; RGB-event representation

P3The Count Is There, but MisalignedarXiv new/cross; VLM representation probing

扫读REBASEarXiv new; training-free in-context segmentation

7 月 12 日

P0Human-like Object Grouping in Self-supervisedarXiv replacement/update; self-supervised ViT analysis

P1Dive Into the Implicit BiasesarXiv new; ECCV 2026; V-L alignment analysis

P2VocaDetarXiv new/cross; visual tokenization; retrieval memory

P2Deep Sprite-based Image ModelsarXiv replacement/update; unsupervised image decomposition

P2ZipDeptharXiv new; ECCV 2026; foundation-model distillation

扫读XOV-ActionarXiv replacement/update; TPAMI accepted; video open-vocabulary

7 月 11 日

P0A Theory of Contrastive LearningarXiv new; ICML 2026; contrastive SSL theory

P0Gen4UarXiv Thu batch; frozen video diffusion representations

P0Converge to SurprisearXiv Thu batch; self-supervised image clustering

P1Scaling Mixture-of-Experts Video Pretraining forarXiv Thu batch; MoE video foundation pretraining

P1Attention-Guided Cross-Temporal Clustering for Self-SupervisedarXiv Thu batch; MIR journal; self-supervised VOS

P1Sparse Attention for Dense Open-VocabularyarXiv Thu batch; CLIP dense prediction

P1LoCAarXiv Thu batch; ECCV 2026; VFM PEFT

P2Dive Into the Implicit BiasesarXiv Fri batch; ECCV 2026; V-L alignment analysis

P2When Structured Sparse Autoencoders LearnarXiv Fri batch; VLM mechanistic interpretability

P2Texture Representations in Deep VisionarXiv Fri batch; representation analysis

P2BUSarXiv Thu batch; unlabeled VLM training

P3Wat3RarXiv Fri batch; ECCV 2026; semi/self-supervised 3D geometry

7 月 9 日

P0GaussFusionarXiv new; multimodal 3D Gaussian pretraining; masked Gaussian modeling

P0Breaking Spurious Correlations via GenerativearXiv new; CVPR Workshop 2026 GCV; cross-variant SSL

P1Vision as Unified Multimodal GenerationarXiv new; unified multimodal vision foundation model

P1From RGB Generation to Dense Field ReadoutarXiv new; generative pretraining for dense prediction

P1SAMPLearXiv new; ECCV main; VLM prompt learning optimizer

P2Analysis-by-ProxyarXiv new/cross; ICML 2026 Mechanistic Interpretability Workshop Spotlight; VLM localization analysis

P2ELSA3DarXiv new/cross; 3D foundation model; cross-modal semantic anchoring

P2MoWorldarXiv new; video/world model pretraining and distillation

P2Segmentation before AnsweringarXiv new/cross; visual-token grounding for MLLM reasoning

P3Generalized Synthetic Image Detection witharXiv new; CLIP/noise representation learning for synthetic image detection

扫读What Images Cannot SayarXiv new/cross; ECCV 2026; multimodal olfactory representation

7 月 8 日

P0Vision Pretraining for Dense SpatialarXiv new; masked boundary modeling; dense spatial pretraining

P0SiamJEPAarXiv new/cross; JEPA; predictive representation learning

P1VLRCarXiv new; V-L feature reprojection; 3D pretraining

P1Text as Partial ConstraintarXiv new/cross; contrastive V-L alignment

P1TORINOarXiv new/cross; SAE; visual-token reduction

P2RADIO1DarXiv new/cross; ICML 2026; compact visual tokens

P2Pathways of Visual Information FlowarXiv new/cross; VLM mechanistic analysis

P2SeeMearXiv new/cross; training-free visual token engineering

P2Does It Fail to SeearXiv new/cross; VLM failure attribution

P2PixConarXiv new/cross; DINOv2 teacher; pixel contrastive learning

P3TrustCLIParXiv new; CLIP privacy; adversarial reconstruction

P3Geometric ReciprocityarXiv new; ICML 2026; self-supervised video geometry

扫读Is the Geometry Doing thearXiv new/cross; TMLR under review; V-L representation audit

7 月 4 日

P0Object-centric LeJEPAarXiv new/cross; JEPA; object-centric SSL

P0Understanding Geometric Representations in Self-SupervisedarXiv new; ECCV 2026; SSL ViT interpretability

P1Understanding the Robustness of DistributedarXiv new; ICLR 2026; SSL theory/MIM vs contrastive

P1Towards Robustness against Typographic AttackarXiv new/cross; ECCV 2026; CLIP mechanistic interpretability

P1Combating Textual Noise and RedundancyarXiv new/cross; ECCV 2026; VLM visual-token compression

P2LASERarXiv new; ECCV 2026; LVLM visual attention

P2DRDNarXiv new; MIM; continual ViT

P2Representation Distribution Matching for One-SteparXiv new; frozen encoder distribution matching

P2Visually Grounded Self-Reflection for Vision-LanguagearXiv new/cross; VLM grounding/RL

P3Efficient PEFT Methods with AdaptivearXiv new; deployment/PEFT; DINOv2/OpenCLIP baseline

P3PointDiTarXiv new; ICML 2026; DINOv3-conditioned geometry

P3Domain Generalization via Text-Anchored InformationarXiv new; ECCV 2026; VLM-guided invariance

7 月 3 日

P2Selective Test-Time Debiasing for CLIParXiv new; ACL 2026 Long; OpenReview/ACL ARR; CLIP/VLM test-time adaptation

P2LongVQUBencharXiv new; ECCV 2026; long-video VLM benchmark

P2RetailSMVarXiv new; foundation video world-model adaptation

P3OnPointarXiv new; ECCV 2026; video distillation

P3ValdiarXiv new/cross; RLC 2026 WMW; latent world model

7 月 2 日

P0LeVLJEPAarXiv new/cross; non-contrastive vision-language pretraining

P0MEPAarXiv new/cross; ECCV 2026; visual AR representation alignment

P0Information-Regularized Attention for Visual-Centric ReasoningarXiv new/cross; ECCV 2026; VLM visual representation regularization

P0SpiralFoveaarXiv new; input-adaptive visual tokenization

P1Cross4D-JEPAarXiv new/cross; 4D point-cloud JEPA distillation

P1MoVAarXiv new/cross; ECCV 2026; long video-text alignment

P1HyFL-CLIParXiv new; ECCV 2026; CLIP fine-tuning

P1Identifying Latent Concepts and StructuresarXiv new/cross; ICML 2026; generalized category discovery

P1Active Spatial GuidancearXiv new; ViT spatial representation

P1RosettaarXiv new/cross; composable multimodal pretraining

P2Mirror-Fusion Attention for Reflection-Aware Self-SupervisedarXiv new/cross; ECML PKDD 2026; reflection-aware SSL

P2From Pixels to Temporal CorrelationsarXiv new; ACM MM 2025; video representation for RL pretraining

P2Condensing Large-Scale Datasets Directly witharXiv new; ECCV 2026; dataset distillation

P3Learning to ComposearXiv new/cross; ECCV 2026; zero-shot composed image retrieval

扫读Steal the Patch SizearXiv new/cross; VLM tokenizer security/diagnostics

7 月 1 日

P0ExPLoRearXiv new; ECCV 2026; masked image modeling

P0Delta-JEPAarXiv new; JEPA/world model

P0CLIMBarXiv new/cross; CoLLAs 2026; online continual SSL

P0GEARarXiv new; image tokenizer/autoregressive generation

P1AdaJEPAarXiv new/cross; adaptive latent world model

P1MemLearnerarXiv new; ECCV 2026; video world model memory

P1Mitigating Positional Leakage in 3DarXiv new/cross; 3D masked autoencoder

P1AVTokarXiv new/cross; ECCV 2026; audio-video tokenizer

P1MS-ResamplerarXiv new; visual resampling for MLLMs

P2Visual Semantic EntropyarXiv new/cross; ECCV 2026; VLM visual uncertainty

P2Decodable Is Not GroundedarXiv new; VLM grounding diagnosis

P2ERAarXiv new; visual token pruning

P3Unsupervised Data-Efficient Cross-Modal Retrieval witharXiv cross; unsupervised cross-modal retrieval

P3Cross-Space DistillationarXiv new; ECCV 2026; diffusion distillation

扫读REDIarXiv new; DINOv3 token reduction

6 月 30 日

P0Large Language Model Teaches VisualarXiv new; ICML 2026; language-to-vision distillation

P0ReWorldarXiv new; video/world model representation learning

P0Parallel Rollout Approximation for Pixel-SpacearXiv new; pixel-space autoregressive image generation/representation

P1VLM-Aware Meta-Optic Front-End Design forarXiv new; frozen VLM/CLIP-aware front-end

P1CascadeOccarXiv new; 3D occupancy world model; VQ representation

P1PerceptionRubricsarXiv new; ICML 2026; multimodal perception evaluation

P1Vision-Default, Prior-OverridearXiv new; VLM mechanistic interpretability

P2DMV-BencharXiv new; multimodal agent visual memory benchmark

P2Reflect-R1arXiv new; ECCV comment; long-video evidence grounding

P2A Unified Framework for VisionarXiv new; equivariant ViT architecture

P3Beyond MoCaparXiv new; motion tokenizer; synthetic data scaling

P3SIFTarXiv new; ECCV 2026; self-imagination video diffusion fine-tuning

P3Qwen-Image-2.0-RL Technical ReportarXiv new; image generation RLHF/on-policy distillation

扫读ScaLe-INRarXiv new; implicit neural representation

6 月 27 日

P0ViQarXiv new; ECCV 2026; visual tokenizer

P0ReasonCLIP-58MarXiv new; ECCV 2026; CLIP continual pretraining

P0Neural Voxel DynamicsarXiv new; V-JEPA/video self-supervised world model

P1Paying More Attention to VisualarXiv new; ECCV 2026; self-evolving LMM

P1Ask, Solve, GeneratearXiv new; unlabeled image self-evolution

P1See & SniffarXiv new; ECCV 2026; cross-modal SSL

P1A Generalization Theory for JEPA-BasedarXiv new; JEPA theory

P2TOPSarXiv new; visual token pruning

P2Bridging Vision and Language ConceptsarXiv new; concept alignment

P2ProtoKVarXiv new; ICML 2026; streaming video memory

P2Staying VIGILantarXiv new; ECCV 2026; counterfactual visual alignment

P3Safe Autoregressive Image Generation witharXiv new; ICML 2026; visual codebook

P3Beyond the Hard BudgetarXiv new; SAE/VFM interpretability

6 月 25 日

P0MJEPAarXiv 新增 · audio-visual JEPA / unified encoder

P0Semantic Allocation in Ordered BottlenecksarXiv 新增 · ordered bottleneck / visual representation

P1V-ZeroarXiv 新增 · VLM self-distillation / contrastive evidence

P1TACOarXiv 新增 + OpenReview 旧状态 · CLIP video adaptation

P1Steering Vision-Language Models with JointarXiv 新增 · VLM interpretability / sparse autoencoder

P1MIMFlowarXiv 新增 · ECCV 2026 · masked image modeling / generation

P2Structuring SparsityarXiv 新增 · visual concept manifolds / interpretability

P2Dual Distribution Estimation for Zero-shotarXiv 新增 · ECCV 2026 · VLM test-time adaptation

P2Brevity is the Soul ofarXiv 新增 · VLM pretraining data curation

P2Causal-rCMarXiv 新增 · video diffusion distillation / world models

P3Curvature-Guided Mixing for MLLM AdaptationarXiv 新增 · ECCV 2026 · MLLM adaptation

P3ROAD-VLAarXiv 新增 · VLA self-distillation / robotics

扫读Wan-Streamer v0.1arXiv 新增 · streaming multimodal foundation model

6 月 24 日

P0P-JEPAarXiv 新增 · video JEPA / long-duration representation

P0Radial Basis Function Networks asarXiv 新增 · SSL projector / label-free representation metric

P1MultiMemarXiv 新增/更新 · multimodal contrastive learning

P1Rethinking Object-Centric Representations for VideoarXiv 新增 · object-centric video / slots

P1Data Selection Through Iterative Self-FilteringarXiv 新增 · CLIP data filtering / bootstrapped pretraining

P1Extraction and Analysis of MultimodalarXiv 新增 · VLM interpretability / sparse autoencoder

P2LEViLarXiv 新增 · video distillation / pseudo labels

P2SPARarXiv 新增 · semantic-pixel alignment / unified tokenizer

P2Structural Assessment for Understanding andarXiv 新增 · dataset distillation / visual tokens

P2Lightweight 3D Feature Pretraining byarXiv 新增 · 2D VFM to 3D representation

P3Structured Hyperedge Adaptation for Parameter-EfficientarXiv 新增 · ViT adaptation / token structure

P3T-VSSarXiv 新增 · VLM visual feature steering

扫读Beyond Flat LabelsarXiv 新增 · hierarchical contrastive learning

6 月 23 日

状态补录3D-DLP连续无新增日复盘 + ICML 2026 + object-centric SSL

状态补录Language-Instructed Vision Embeddings for Controllable连续无新增日复盘 + ICLR 2026 Poster + language-steered VFM

状态补录LEAP连续无新增日复盘 + ViT/VFM feature distillation

状态补录Spatial-Aware Reduction Framework连续无新增日复盘 + ICML 2026 + visual token reduction

状态补录UNIEGO连续无新增日复盘 + video representation + multi-teacher distillation

6 月 22 日

状态补录3D-DLP连续无新增日复盘 + ICML 2026 + object-centric SSL

状态补录Language-Instructed Vision Embeddings for Controllable连续无新增日复盘 + ICLR 2026 Poster + language-steered VFM

状态补录LEAP连续无新增日复盘 + ViT/VFM feature distillation

状态补录Spatial-Aware Reduction Framework连续无新增日复盘 + ICML 2026 + visual token reduction

状态补录UNIEGO连续无新增日复盘 + video representation + multi-teacher distillation

6 月 21 日

状态补录3D-DLP周末复盘 + ICML 2026 + object-centric SSL

状态补录Language-Instructed Vision Embeddings for Controllable周末复盘 + ICLR 2026 Poster + language-steered VFM

状态补录LEAP周末复盘 + ViT/VFM feature distillation

状态补录Spatial-Aware Reduction Framework周末复盘 + ICML 2026 + visual token reduction

状态补录UNIEGO周末复盘 + video representation + multi-teacher distillation

6 月 20 日

P03D-DLParXiv 新增 + ICML 2026 + object-centric SSL

P0Language-Instructed Vision Embeddings for ControllablearXiv 新增 + ICLR 2026 Poster + language-steered VFM

P0LEAParXiv 新增 + ViT/VFM feature distillation

P0Spatial-Aware Reduction FrameworkarXiv 新增 + ICML 2026 + visual token reduction

P1UNIEGOarXiv 新增 + video representation + multi-teacher distillation

P1The Hidden Evolution of DisguisedarXiv 新增 + VLM representation diagnosis

P1SPOT-EarXiv 新增 + frozen VLM grounding

P1SpatialSVarXiv 新增 + IJCAI 2026 + 3D visual supervision

P2NESTarXiv 新增 + long video benchmark

P2Multimodal Concept Bottleneck ModelsarXiv 新增 + NeurIPS 2025 MI workshop + CLIP interpretability

P2TimagearXiv 新增 + ECCV + input-level VLM alignment

P2ELVAarXiv 新增 + ECCV 2026 + multimodal retrieval

P2Current World Models Lack aarXiv 新增 + world-model diagnostic benchmark

P3Through the PRISMarXiv 新增 + video diffusion latent preference

P3Variable-Length Tokenization via Learnable GlobalarXiv 新增 + diffusion tokenizer / visual tokens

6 月 19 日

P0Seeing Before ReasoningarXiv 新增 + MLLM on-policy self-distillation

P0Visual-OPSDarXiv 新增 + cross-modal on-policy self-distillation

P0Moving Beyond DiversityarXiv 新增 + ECCV 2026 under review + VLM token pruning

P1Beyond the Linear Separability CeilingarXiv replacement/update + TMLR accepted

P1Reasoning as IntersectionarXiv 新增 + video evidence alignment

P1Improving Visual Token Reduction viaarXiv replacement/update + ICML 2026 accepted + visual token reduction

P1MUFASAarXiv replacement/update + CVPR 2026 + object-centric SSL

P2DREAMarXiv 新增 + video-text retrieval representation

P2Benchmarking Large Vision-Language Models onarXiv 新增 + LVLM fine-grained diagnosis

P2Semantic Robustness Certification for Vision-LanguagearXiv 新增 + ICML accepted + VLM robustness

P2LAREarXiv 新增 + ICML 2026 EMM-QA workshop + text-image retrieval

P2APTarXiv 新增 + causal video-language understanding

P3UniTemparXiv 新增 + video diffusion bidirectional distillation

P3Data-Forcing DistillationarXiv 新增 + few-step video generation distillation

6 月 18 日

P0Expanding SPHERE-JEPAarXiv 新增 + SSL regularizer / JEPA collapse prevention

P0Unified Multimodal Autoregressive Modeling witharXiv 新增 + ICML 2026 Poster + unified visual tokenizer

P0See First, Answer LaterarXiv 新增 + MLLM visual evidence pre-alignment

P1TivTokarXiv 新增 + video tokenizer / latent factorization

P1ThinkJEPAarXiv replacement + VLM-guided JEPA latent world model

P1Visuals Lie, Consistency SpeaksarXiv 新增 + ICLR 2026 Workshop on Multimodal Intelligence

P1Future Dynamic 3D ReconstructionarXiv 新增 + ICML 2026 Poster + 3D latent world model

P2Rethinking Cross-Layer Information Routing inarXiv replacement + diffusion transformer representation routing

P2Reinforcing Dual-Path Reasoning in SpatialarXiv 新增 + spatial VLM RL / 3D grounding

P2LiveStarProarXiv 新增 + streaming video memory

P2RAIGenarXiv replacement + ICML 2026 Poster + sparse autoencoder interpretability

P3Pulling The REINSarXiv 新增 + video diffusion representation steering

P3Last But Not LeastarXiv replacement + MLLM visual token retention

扫读Beyond Native SuccessarXiv 新增 + CLIP security / deployment interface audit

6 月 17 日

P0You Don't Need Strong AssumptionsarXiv 新增 + video SSL / JEPA-adjacent predictive representation

P0Momentum-Guided Semantic Forecasting (MoFore) forarXiv 新增 + self-supervised video representation

P0Distilling Drifting Transformers with RepresentationarXiv 新增 + representation autoencoder distillation

P1Selective Synergistic Learning for VideoarXiv 新增 + video object-centric learning

P1Task-Instructed Causal Routing of VisionarXiv 新增 + frozen VFM routing / dense transfer

P1Beyond Scalar DistancesarXiv 新增 + MLLM-derived visual embedding supervision

P1Cascaded Sparse Autoencoders Learn Multi-LevelarXiv 新增 + MLLM visual concept interpretability

P2MVEBarXiv 新增 + video embedding benchmark

P2What Should a Streaming VideoarXiv 新增 + streaming video latent memory

P2Thinking with Visual GroundingarXiv 新增 + visually grounded reasoning supervision

P2DifFRACTarXiv 新增 + diffusion transformer feature circuits

P3LOCUSarXiv 新增 + MLLM local cue self-improvement

扫读PositionarXiv 新增 + ICML 2026 accepted position

6 月 16 日

P0RepFusionarXiv 新增 + representation autoencoder

P0TSAarXiv 新增 + object-centric video SSL

P0S$^2$COPEarXiv 新增 + self-supervised concept discovery

P1RATS! Patches Talk Through RegistersarXiv 新增 + register/part representation

P1Learning What to Predictself-supervised objective design

P1ViT-UparXiv 新增 + DINO/VFM dense features

P1Temporal Straightening for Latent PlanningICML 2026 + JEPA/world model

P2Self-Evolving Visual QuestionerarXiv 新增 + VLM self-evolution

P2Mirage ProbesarXiv 新增 + VLM representation diagnostics

P2Gaze HeadsarXiv 新增 + VLM mechanism

P3What Drives Test-Time Adaptation forarXiv 新增 + CLIP adaptation benchmark

扫读CausalMotionarXiv 新增 + video generation representation

6 月 14 日

P0V-JEPA 2.1arXiv 更新 (v3)

P0Dual-State Slot AttentionarXiv 新增

P1Cross-Modal Masked Compositional Concept ModelingACL 2026 Main + arXiv

P1HYDRA-XarXiv 新增

P1RepWAMarXiv 新增 + project

P2Modality Forcing for Scalable SpatialarXiv 新增 + project

P2MaskWAMarXiv 新增

P2GeoWorld-VLMarXiv 更新

P3EPM-JEPAarXiv 新增

P3ECAICML 2026 + arXiv

6 月 11 日

P0Latent Spatial Memory for VideoarXiv recent-list补漏 + project

P0MaskAlignarXiv recent-list补漏

P1SMIarXiv recent-list补漏

P1FF-JEPAarXiv recent-list补漏

P1MilliVidarXiv recent-list补漏 + project

P2ATMarXiv recent-list补漏

P2Echo-MemoryarXiv recent-list补漏 + code

P2Prisma-WorldarXiv recent-list补漏 + project

P2Securing Self-supervised Data Curation forarXiv recent-list补漏

P3Thinking Without ImagesarXiv recent-list补漏

扫读Test-Time Scaling in Multimodal FoundationACL 2026 Findings + arXiv

6 月 10 日

P0Next ForcingarXiv 新增

P0IDEALarXiv 新增

P1ARMarXiv 新增

P1Can Image Models Imagine Time? ImageTimearXiv 新增

P1BiWMarXiv 今日公告

P1Let ViT SpeakarXiv 更新

P2When to Align, When to PredictarXiv 新增

P2Cross-Modal Knowledge Distillation without PairedarXiv 新增

P2SARAarXiv 更新

P2Beyond Model SizearXiv 新增

P2Quo Vadis, Visual In-Context Learning?arXiv 新增

P3One Token per Multimodal EvidencearXiv 新增

扫读Mitigating Manifold DepartureICML 2026 + arXiv

6 月 9 日

P0What Makes Video World ModelarXiv 新增

P0Do Video Foundation Models UnderstandarXiv 新增

P1Evaluating the Representation Space ofarXiv 新增 + ICML FoGen Workshop

P1A Mixed Diet Makes DINOCVPR 2026 Highlight

P1Self-Supervised Learning with a Multi-TaskarXiv 更新

P2Neural Field Tokenizations with HierarchyarXiv 新增

P2Vision-Language Asymmetry in Bistable ImagearXiv 新增 + ICML Workshop

P2Late-Layer Fusion is EnougharXiv 新增

P3The Cross-Architecture SubstratearXiv 新增

扫读Look Less, Reason MorearXiv 新增

6 月 8 日

P0AdaTokarXiv 新增

P0TEVIarXiv 新增

P1Inside the Visual MindarXiv 新增 + ICML 2026 Poster

P1Lighting-Aware Representation Learning under ControllablearXiv 新增

P1MotionEnhancerarXiv 新增 + CVPR 2026

P2Diagnosing Visual Ignorance in Vision-LanguagearXiv 新增

P2Generalization of Diffusion Models ArisesICLR 2026

P2Breaking the Lock-inarXiv 新增 + ICML 2026 Poster

P3The Dual Mechanisms of SpatialarXiv 更新

扫读Implicit Data Synthesis for ContrastivearXiv 新增

6 月 7 日

P0From Static to DynamicCVPR 2026 Day 3

P0PyraTokCVPR 2026 Day 3

P0MacTokCVPR 2026 Day 3

P1Toward Diffusible High-Dimensional Latent SpacesCVPR 2026 Day 3

P13D sans 3D ScansCVPR 2026 Day 3

P1E-RayZerCVPR 2026 Day 1 补录

P1VideoSSRCVPR 2026 Day 2 补录

P2Progressive Mask Distillation for Self-supervisedCVPR 2026 Day 3

P2CORECVPR 2026 Day 3

P3SynCLIPCVPR 2026 Day 3

扫读Scaling Dense Event-Stream Pretraining fromCVPR 2026 Day 1 补录

6 月 6 日

P0Learning Geometric Representations from VideosVisual SSL / representation

P0Balancing Image Compression and GenerationVisual SSL / representation

P0Imagine Before You Predictcode

P1The Invisible Hand of PhysicsVisual SSL / representation

P1ActiveMimicproject

P1Learning Visual Spatial Planning fromVisual SSL / representation

P1ViCuRVisual SSL / representation

P2Cosine MisleadsVisual SSL / representation

P2HyperVisVisual SSL / representation

P2Mechanistic Insights into Functional SparsityVisual SSL / representation

P3Knowledge Distillation for Visual AutoregressiveVisual SSL / representation

P3BRepCLIPproject

扫读VideoKRICML 2026 Spotlight

6 月 5 日

P0Geometry-Preserving Unsupervised Alignment for HeterogeneousICML 2026

P0KODAVisual SSL / representation

P0Continual Visual and Verbal LearningVisual SSL / representation

P1The Loss Is Not EnoughICML 2026

P1RePercENTVisual SSL / representation

P1ChannelTokVisual SSL / representation

P2Stateful Visual Encoders for Vision-Languageproject

P2Scene-Centric Unsupervised Video Panoptic SegmentationCVPR 2026

P2Who Needs Labels? Adapting VisionVisual SSL / representation

P2Intra-Modal Neighbors Never LieVisual SSL / representation

P3Video2LoRAVisual SSL / representation

P3Hyper-ICLICML 2026

扫读M$^3$EvalVisual SSL / representation

6 月 4 日

P0$A^2$Visual SSL / representation

P0IdEstICML 2026

P1PRISMICML 2026

P1Beyond CompressionVisual SSL / representation

P1Formalizing the Binding ProblemICML 2026

P2Visual Instruction Tuning Aligns ModalitiesVisual SSL / representation

P2Cosmos 3project/code

P3Where Do We (Not) NeedVisual SSL / representation

P3Attend to AnythingICML 2026

扫读Benchmarking Visual State Tracking inproject

6 月 3 日

P0VISRegVisual SSL / representation

P0UR-JEPAVisual SSL / representation

P1Scaling Parallel Sequence Models toVisual SSL / representation

P1Decomposed On-Policy Distillation for Vision-LanguageICML 2026 Spotlight

P1V-LynXICML 2026

P2AdaCodecVisual SSL / representation

P2EvoCutVisual SSL / representation

P2From Extrinsic to IntrinsicICML 2026 Poster

P3SCAPOVisual SSL / representation

P3InfoMergeVisual SSL / representation

扫读T-CLIPVisual SSL / representation

6 月 2 日

P0Internalizing Temporal Consistency in VideoarXiv backfill

P1SOCOproject

P2Learning from Fine-Grained Visual DiscrepanciesICML 2026 Poster

P2Diversity MattersICML 2026 Poster

P3Toward Identifiable Sparse AutoencodersICML 2026 Poster

P3ReGuLaRVisual SSL / representation

扫读YARDVisual SSL / representation

6 月 1 日

P0RayDerproject

P1How can embedding models bind concepts?ICML 2026

P1Representation Forcing for Bottleneck-Free Unifiedproject

P1VisionPulseICML 2026

P2Fixed-Point Masked Generative ModelingVisual SSL / representation

P2Variational Adapter for Cross-modal SimilarityICML 2026 + OpenReview

P2MergeTokVisual SSL / representation

P2Beyond ClassificationVisual SSL / representation

P3PEEKcode

P3Interpretability Without TradeoffsVisual SSL / representation

5 月 31 日

P1SLADCVPR Findings 2026

P2UniNoteKDD Ads Track 2026

P2Geometry-Guided Modeling of Foundation FeaturesICML 2026

P2minWMVisual SSL / representation

P3VideoMLAVisual SSL / representation

P3YoCausalVisual SSL / representation

P3A Geometric View of SRCVisual SSL / representation

扫读How Much Is a Datasetdata scaling

5 月 30 日

P0Cycle Consistency in Video Object-CentricVisual SSL / representation

P0LoMoVisual SSL / representation

P0Beyond 3D VQAsCVPR 2026 + arXiv

P0GPICdataset

P1TRACERICML 2026 + arXiv

P1PARCELVisual SSL / representation

P1Geometry MattersVisual SSL / representation

P1Improving CLIP Adaptation by BreakingICML 2026 + arXiv

P2EarlyTomCVPR 2026 + arXiv

P2Deep Psychovisual Image RepresentationsVisual SSL / representation

P2Learning Representations from 3D GaussianVisual SSL / representation

P3VedaICML 2026 + arXiv

P3Grounded 3D-Aware Spatial Vision-Language ModelingCVPR 2026 + arXiv

扫读OccamTokenVisual SSL / representation

扫读Why Far Looks UpVisual SSL / representation

5 月 29 日

P0From Pixels to Words --Visual SSL / representation

P0Which Pretraining Paradigm Better ServesVisual SSL / representation

P0Representation-Conditioned Diffusion Models for GuidedarXiv cross-list

P1Beyond Surrogate GradientsVisual SSL / representation

P1AdaMergearXiv cross-list

P1SIGMAVisual SSL / representation

P1Pix2KeyICML 2026 + arXiv status

P2ROVERarXiv cross-list

P2Self-Prophetic Decoding to Unlock VisualICML 2026

P2Look on DemandVisual SSL / representation

P2Beyond InterpretabilityarXiv update

P3Understanding Self-Supervised Learning via LatentarXiv update

P3ProprioVisual SSL / representation

P3RE-TRIANGLEarXiv cross-list

扫读OmniVerifier-M1arXiv cross-list

扫读A self-supervised learning approach toVisual SSL / representation

5 月 28 日

P0SPHERE-JEPAVisual SSL / representation

P0Gemini Embedding 2Visual SSL / representation

P0DV-SFTVisual SSL / representation

P0FoundObjICML 2026

P1OmniRetrieverVisual SSL / representation

P1JLTVisual SSL / representation

P1JetViTCVPR 2026 Findings

P1FAST-GOALIEEE TIP 2026

P2Q-GeoMemVisual SSL / representation

P2How and What to Imagine?Visual SSL / representation

P2The Rescue EffectVisual SSL / representation

P2Geometry-Aware Representation Denoising for RobustVisual SSL / representation

P3MRTCVPR 2026

P3O-MARCVisual SSL / representation

P3SIMPCICML 2026

扫读Pop-Up Distractions Reveal Bag-of-Events BehaviorVisual SSL / representation

5 月 27 日

P0LEASECVPR 2026

P0DIVAICML 2026

P0UWM-JEPAVisual SSL / representation

P0Unified 3D Scene Understanding ThroughICLR 2026 Poster + arXiv

P1DUELVisual SSL / representation

P1MAGICVisual SSL / representation

P1Channel-wise Vector QuantizationVisual SSL / representation

P1Nano World ModelsProject

P2Causal Route GatingICML 2026 Spotlight

P2AFIPICML 2026

P2FG-CLIP 2arXiv v3 + ICML 2026

P2Residual Connections Harm Generative RepresentationarXiv update + CVPR 2026

P2Reevaluating CLIP Intra-Modal MisalignmentarXiv update + CVPR 2026

P3World Models as Group ActionsVisual SSL / representation

P3Uncertainty-DTW for Visual TokensVisual SSL / representation

P3CLIP-Guided SAMVisual SSL / representation

P3HCL-FFCVPR 2026

5 月 26 日

P1The TIME MachinearXiv 新增

P1LaMoarXiv 新增 / 项目页

P1Vision Transformers Need Better TokenarXiv 新增

P2Good Token HuntingarXiv 新增 / 项目页

P2Transcoders Trace Visual GroundingarXiv 新增

P2Seeing without LookingCVPR 2026 Workshop

P2PGTarXiv 新增 / ICML 列表出现

P2Not Too Generative, Not Too DiscriminativearXiv 新增

P3Dithering DefenseICIP 2026

P3CVSearchICML 2026

5 月 25 日

P1TextTeacherLanguage-guided visual representation

P1MDMVLM dataset distillation

P2EvoVidVideo SSL / self-evolution

P2Ablate-to-ValidateVLM diagnostics

P3Neural Collapse by DesignRepresentation geometry

扫读SESTEvent-camera SSL transfer

5 月 24 日

P1TC-JEPAJEPA / VLM pretraining

P1ResDreamerWorld model / video SSL

P1Cambrian-PVideo MLLM / geometry

P2Self-Supervised Weight Templates for ScalableICML 2026

P2DeltaDirectVideo diagnostics

P2GeoWeaverGeometry-grounded VLM

P3Swift SamplingarXiv 新增

P3The Matching PrinciplearXiv 新增

状态补录DINOv3TMLR Accepted / OpenReview

5 月 23 日

P0DecQarXiv 新增

P0What Does the Caption Really Say?arXiv 新增

P0RiTarXiv 新增

P1Visual-Advantage OPDarXiv 新增

P1Conceptualizing EmbeddingsarXiv 新增

P1Supervised Heads as Semantic PrototypesarXiv 新增

P2ST-SimDiffICLR 2026 + arXiv

P2ST-GridPoolICLR 2026 + arXiv

P2RISEarXiv 新增

P3SegCompassCVPR 2026 + arXiv

5 月 22 日

P1Pairwise ModalitiesMultimodal representation

P1Weighted Reverse Convolution for FeaturearXiv replacement

P2AIRarXiv 新增

P2Capability != InterpretabilityarXiv 新增

P3SGAarXiv 新增

P3FullFlowarXiv 新增

P3Why Latent Actions Fail, andarXiv 新增

5 月 21 日

高UniRefinerCVPR 2026

高Towards Data-Efficient Video Pre-training withCVPR Workshop

中高Efficient coding along the visualVisual SSL / representation

中高Targeted Downstream-Agnostic AttackVisual SSL / representation

中Vision-OPDarXiv 更新