{"total":47,"items":[{"citing_arxiv_id":"2607.00684","ref_index":3,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"AdaBoosting Text Prompts for Vision-Language Models","primary_cat":"cs.LG","submitted_at":"2026-07-01T09:28:55+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"TPB is an AdaBoost-style ensemble method for text prompts in VLMs that improves few-shot accuracy by targeting hard examples and maintains gains across model transfers.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.25329","ref_index":183,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"State Space Models Meet Remote Sensing: A Survey","primary_cat":"cs.CV","submitted_at":"2026-06-24T02:50:31+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":2.0,"formal_verification":"none","one_line_summary":"A literature survey of State Space Model methods applied to remote sensing tasks, architectures, and challenges since their introduction to the field.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.21734","ref_index":297,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"HPP: Hierarchical Programmatic Probing for Long Video Understanding by Decoupling Perception and Reasoning","primary_cat":"cs.CV","submitted_at":"2026-06-19T20:43:49+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"HPP decouples perception from reasoning in long-video VLMs by having an LLM run iterative programmatic probes on hierarchically segmented video, reporting gains on LongVideoBench, EgoSchema, VideoMME, and MLVU.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.19706","ref_index":77,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"NEST: Narrative Event Structures in Time for Long Video Understanding","primary_cat":"cs.CV","submitted_at":"2026-06-18T02:05:14+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"NEST is a new benchmark dataset for narrative event structures in long videos, with baselines reporting ETD below 8%, EL under 6%, EAE below 11%, and ERE at 35-44% F1.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.19584","ref_index":4,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Language-Instructed Vision Embeddings for Controllable and Generalizable Perception","primary_cat":"cs.CV","submitted_at":"2026-06-17T20:39:04+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"LIVE uses language to generate task-centric vision embeddings at inference, reducing hallucinations by 34 points on MMVP, outperforming larger VLMs on VQA, and generalizing to unseen tasks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.17030","ref_index":64,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation","primary_cat":"cs.CV","submitted_at":"2026-06-15T17:52:31+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Qwen-RobotWorld is a language-conditioned video world model using Double-Stream MMDiT, an 8.6M-frame embodied corpus, and progressive curriculum training that ranks first on EWMBench and DreamGen Bench.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.13275","ref_index":18,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Zero-Shot Captioning for Cultural Heritage: Automated Image Analysis of Traditional Indonesian Clothing","primary_cat":"cs.CV","submitted_at":"2026-06-11T12:29:39+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"Custom ZeroCLIP uses retrieval from seen provinces to caption traditional Indonesian clothing images from 8 unseen provinces, achieving CLIPScore 0.8536, BLEU-4 0.3342, and METEOR 0.4859 while outperforming baselines.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.07032","ref_index":38,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Never Seen Before: Benchmarking Genuine Zero-Shot Composed Image Retrieval with Consistent Video-Sourced Datasets","primary_cat":"cs.CV","submitted_at":"2026-06-05T08:23:25+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"ZeroSight supplies a video-derived dataset and evaluation protocol for genuine zero-shot composed image retrieval plus the SC4CIR consistency method, demonstrating that prior benchmarks inflate reported performance across 27 tested approaches.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.03793","ref_index":11,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Exploring Adversarial Robustness and Safety Alignment in Multilingual Multi-Modal Large Language Models","primary_cat":"cs.CL","submitted_at":"2026-06-02T15:42:10+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Adversarial images transfer across languages in MLLMs while apparent safety in weaker languages stems from comprehension and visual-grounding failures rather than genuine alignment.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.01049","ref_index":6,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"PMC-InterCPT: Rethinking Biomedical Interleaved Data for Multimodal Continued Pretraining","primary_cat":"cs.CL","submitted_at":"2026-05-31T06:38:30+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"PMC-InterCPT builds a context-grounded biomedical interleaved corpus from PMC literature and shows it improves multimodal performance on Qwen3.5-4B-Base after CPT and SFT while using fewer tokens.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.30126","ref_index":32,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"PARCEL: Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding","primary_cat":"cs.CV","submitted_at":"2026-05-28T15:57:31+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"PARCEL is a new visual tokenization architecture combining pool-anchored resampling with conditioned elastic queries to enhance performance-efficiency tradeoffs in LVLMs over prior matryoshka methods.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.26601","ref_index":6,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"FTibSuite: A Comprehensive Resource Suite for Tibetan Vision-Language Modeling","primary_cat":"cs.CV","submitted_at":"2026-05-26T06:36:27+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"FTibSuite provides human-verified multimodal corpora, Tibetan-adapted benchmarks with quality controls, and a baseline VLM showing gains on tasks like MMBench while preserving Chinese capabilities.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.24456","ref_index":10,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"EgoProx: Evaluating MLLMs on Egocentric 3D Proximity Reasoning Across a Cognitive Hierarchy","primary_cat":"cs.CV","submitted_at":"2026-05-23T08:07:45+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"EgoProx benchmark shows MLLMs have some spatial knowledge but struggle to leverage it for egocentric 3D proximity reasoning VQA.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.16972","ref_index":22,"ref_count":2,"confidence":0.98,"is_internal_anchor":true,"paper_title":"WhiteTesseract: Reframing the Interpretation of Cultural Heritage through XR and Conversational AI","primary_cat":"cs.HC","submitted_at":"2026-05-16T12:50:37+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":3.0,"formal_verification":"none","one_line_summary":"WhiteTesseract integrates XR diminished reality and LLM dialogue to increase viewing duration and interaction depth in physical cultural heritage exhibitions, shown in a 26-participant Monet exhibition study with statistically significant results.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.16834","ref_index":18,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Learning Relative Representations for Fine-Grained Multimodal Alignment with Limited Data","primary_cat":"cs.CV","submitted_at":"2026-05-16T06:33:38+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"A new post-hoc alignment technique uses learnable anchors to capture token-level relative similarities between modalities, outperforming global alignment baselines on zero-shot classification, retrieval, and segmentation with scarce paired examples.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.16603","ref_index":7,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Controlla: Learning Controllability via Graph-Constrained Latent Geometry","primary_cat":"cs.CV","submitted_at":"2026-05-15T20:06:02+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Controlla learns identity and attribute factors from multimodal inputs and aligns them with graph priors using graph-constrained optimal transport to enforce consistent attribute trajectories while preserving reference identity.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.09948","ref_index":6,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"LoopVLA: Learning Sufficiency in Recurrent Refinement for Vision-Language-Action Models","primary_cat":"cs.AI","submitted_at":"2026-05-11T03:51:22+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"LoopVLA adds recurrent refinement and learned sufficiency estimation to VLA models, cutting parameters 45% and raising throughput 1.7x while matching baseline task success on LIBERO and VLA-Arena.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"InFirst Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024. [5] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748-8763. PmLR, 2021. [6] Xi Chen, Xiao Wang, Soravit Changpinyo, Anthony J Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model.arXiv preprint arXiv:2209.06794, 2022. [7] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson,"},{"citing_arxiv_id":"2604.24642","ref_index":2,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Probing CLIP's Comprehension of 360-Degree Textual and Visual Semantics","primary_cat":"cs.CV","submitted_at":"2026-04-27T16:10:00+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"CLIP models understand 360-degree textual semantics via explicit identifiers but show limited comprehension of visual semantics under horizontal circular shifts, which a LoRA fine-tuning approach improves with a noted trade-off in original task performance.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.13565","ref_index":3,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"UHR-BAT: Budget-Aware Token Compression Vision-Language model for Ultra-High-Resolution Remote Sensing","primary_cat":"cs.CV","submitted_at":"2026-04-15T07:21:37+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"UHR-BAT is a budget-aware framework that uses text-guided multi-scale importance estimation plus region-wise preserve and merge strategies to compress visual tokens in ultra-high-resolution remote sensing vision-language models.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"dundant tokens consume a significant fraction of the budget. Under a hard cap, this biases the selection toward dominant regions, causing sparse yet critical targets to be missed. To encourage structural diversity and preserve fine-grained de- tails, we partition the token index set I={1, . . . , N} at the original resolution intoRdisjoint regions: P={S m}R m=1, R[ m=1 Sm =I,S m ∩ Sn =∅.(4) Our goal is to group tokens that convey the same semantic information. Therefore, we instantiate the partition using methods that detect such similarities. Options include clus- tering in a joint feature-coordinate space (McQueen, 1967; Hartigan & Wong, 1979; Lloyd, 1982), or using segmen- tation models (e.g., SAM (Kirillov et al., 2023)) followed by a token-level mapping (Appendix C.2). We apply this independently at each scale. 3.3. Scale-Specific Positional Embedding To capture both global context and fine details, we con- struct S resized views {I(s)}S s=1. We designate the lowest- resolution view (s= 1 ) as theanchorto preserve global structure. Higher scales provide progressively finer de- tails. For each scale s, the vision encoder outputs tokens E(s) ={e (s) i }Ns i=1. These tokens form a Us ×V s grid, where Ns =U s ·V s. Tokens from different resolutions must re- main spatially aligned and distinguishable. Therefore, we augment each token to obtain updated features: h(s) i =e (s) i +p (s) i +q (s).(5) Here p(s) i ∈R d is obtained by bilinearly interpolating a base 2D positional embedding (defined on the pretraining grid) to the Us ×V s grid, and q(s) ∈R d is a learned scale embedding (a lookup table overs). Interpolation preserves geometric consistency across resized views, and q(s) pre- vents ambiguity between tokens from different scales. Cross-scale Importance Alignment.Token selection should be comparable across scales under a unified notion of importance. We take the anchor-scale importance as ref- erence: we reshape the anchor token scores {a(1) i } onto "},{"citing_arxiv_id":"2604.10064","ref_index":4,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"On The Application of Linear Attention in Multimodal Transformers","primary_cat":"cs.CV","submitted_at":"2026-04-11T07:06:33+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"Linear attention delivers significant computational savings in multimodal transformers and follows the same scaling laws as softmax attention on ViT models trained on LAION-400M with ImageNet-21K zero-shot validation.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2603.09921","ref_index":5,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"WikiCLIP: An Efficient Contrastive Baseline for Open-domain Visual Entity Recognition","primary_cat":"cs.CV","submitted_at":"2026-03-10T17:18:53+00:00","verdict":null,"verdict_confidence":null,"novelty_score":null,"formal_verification":null,"one_line_summary":null,"context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2603.08942","ref_index":3,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"BiCLIP: Domain Canonicalization via Structured Geometric Transformation","primary_cat":"cs.CV","submitted_at":"2026-03-09T21:26:15+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"BiCLIP recovers a structured geometric transformation from few-shot anchors to canonicalize domain features in VLMs and reports state-of-the-art results on 11 benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2603.00655","ref_index":6,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Mema: Memory-Augmented Adapter for Enhanced Vision-Language Understanding","primary_cat":"cs.CV","submitted_at":"2026-02-28T13:57:19+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Mema adds a stateful memory module to vision encoders that accumulates hierarchical visual features across layers and selectively injects portions back via feedback to preserve fine-grained cues, yielding consistent gains on multimodal benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2602.01738","ref_index":7,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Simplicity Prevails: The Emergence of Generalizable AIGI Detection in Visual Foundation Models","primary_cat":"cs.CV","submitted_at":"2026-02-02T07:20:02+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Frozen features from vision foundation models enable a linear probe to outperform specialized AIGI detectors by over 30% on in-the-wild data due to emergent forgery knowledge from pre-training.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2512.10821","ref_index":3,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Agile Deliberation: Concept Deliberation for Subjective Visual Classification","primary_cat":"cs.AI","submitted_at":"2025-12-11T17:13:09+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Agile Deliberation improves F1 scores by 7.5% over automated baselines and 3% over manual deliberation in 18 user sessions by supporting iterative refinement of subjective visual concepts.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2510.12710","ref_index":28,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Reflection-Based Task Adaptation for Self-Improving VLA","primary_cat":"cs.RO","submitted_at":"2025-10-14T16:44:39+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Reflective Self-Adaptation combines failure-reflective reinforcement learning with success-guided imitation learning to enable faster and more reliable task adaptation for pre-trained Vision-Language-Action models.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2509.19207","ref_index":8,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Long Story Short: Disentangling Compositionality and Long-Caption Understanding in Contrastive VLMs","primary_cat":"cs.CV","submitted_at":"2025-09-23T16:28:51+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Empirical study shows bidirectional but sensitive relationship between compositionality and long-caption understanding in VLMs, promoted by high-quality grounded data and affected by architectural choices like frozen positional embeddings.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2509.10026","ref_index":10,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"LaV-CoT: Language-Aware Visual CoT with Multi-Aspect Reward Optimization for Real-World Multilingual VQA","primary_cat":"cs.CV","submitted_at":"2025-09-12T07:45:44+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"LaV-CoT introduces a multi-stage visual CoT pipeline and GRPO training with language-consistency rewards, delivering up to 9.5% accuracy gains on multilingual VQA benchmarks over similar-sized open models.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2506.01844","ref_index":10,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics","primary_cat":"cs.LG","submitted_at":"2025-06-02T16:30:19+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SmolVLA is a small efficient VLA model that achieves performance comparable to 10x larger models while training on one GPU and deploying on consumer hardware via community data and chunked asynchronous action prediction.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2505.23678","ref_index":7,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Grounded Reinforcement Learning for Visual Reasoning","primary_cat":"cs.CV","submitted_at":"2025-05-29T17:20:26+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"ViGoRL introduces visually grounded RL that anchors reasoning steps to image coordinates and uses multi-turn zooming to outperform standard RL and supervised baselines on spatial and GUI reasoning benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2505.16819","ref_index":14,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Character-Centered Dialogue Generation from Scene-Level Prompts","primary_cat":"cs.CV","submitted_at":"2025-05-22T15:54:42+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"A training-free framework generates expressive, character-grounded dialogue and speech from scene prompts using vision-language encoders, LLMs, and a recursive narrative memory bank for cross-scene consistency.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2502.14786","ref_index":10,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features","primary_cat":"cs.CV","submitted_at":"2025-02-20T18:08:29+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"SigLIP 2 models trained with a unified recipe of captioning, self-supervised losses, and curated diverse data outperform prior SigLIP versions on classification, retrieval, localization, dense prediction, and multilingual understanding at scales from 86M to 1B parameters.","context_count":1,"top_context_role":"dataset","top_context_polarity":"use_dataset","context_text":"representations are pooled using a MAP head (at- 2 SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features tentionpooling)[ 69]. Wesetthetextlengthto64 and use the multilingual Gemma tokenizer [22] with vocabulary size 256k, transforming the text to lower case before tokenization. We use the WebLI dataset [10] containing 10 billion images and 12 billion alt-texts covering 109 languages. To strike a good balance be- tween quality on English and multilingual vision- language benchmarks we compose the mixture such that 90% of the training image-text pairs is sourced from English web pages, and the re- maining 10% from non-English web pages, as recommended in [49]."},{"citing_arxiv_id":"2412.03555","ref_index":14,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"PaliGemma 2: A Family of Versatile VLMs for Transfer","primary_cat":"cs.CV","submitted_at":"2024-12-04T18:50:42+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"PaliGemma 2 is a family of vision-language models that achieves state-of-the-art results on transfer tasks like table structure recognition and radiography report generation by combining SigLIP with Gemma 2 models at various sizes and resolutions.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2408.16500","ref_index":13,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"CogVLM2: Visual Language Models for Image and Video Understanding","primary_cat":"cs.CV","submitted_at":"2024-08-29T12:59:12+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"CogVLM2 family achieves state-of-the-art results on image and video understanding benchmarks through improved visual expert architecture, higher resolution inputs, and automated temporal grounding for videos.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2408.04840","ref_index":32,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models","primary_cat":"cs.CV","submitted_at":"2024-08-09T03:25:42+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"mPLUG-Owl3 introduces hyper attention blocks to integrate vision and language for long image-sequence understanding and reports SOTA results on single-image, multi-image, and video benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2407.07726","ref_index":27,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"PaliGemma: A versatile 3B VLM for transfer","primary_cat":"cs.CV","submitted_at":"2024-07-10T14:57:46+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"PaliGemma is an open 3B VLM based on SigLIP and Gemma that achieves strong performance on nearly 40 diverse open-world tasks including benchmarks, remote-sensing, and segmentation.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"placingafixedandlargesetofclassesbyacaption embedding instead. The caption embeddings are mostly obtained using language encoders (sim- ilar to BERT [33]) and allow to open up the vocabulary of classification and retrieval tasks. The second generation, akin to T5 [95] in lan- guage, is a unification of captioning and question- answering tasks via generative encoder-decoder modeling [27, 111, 120, 138], often backed by the progress in generative language models. Corresponding author(s): lbeyer,xzhai@google.com © 2024 Google DeepMind. All rights reserved arXiv:2407.07726v2 [cs.CV] 10 Oct 2024 PaliGemma: A versatile 3B VLM for transfer Gemma: 2B Language Model SigLIP: 400M Vision Model Contrastive Vision Encoder Transformer Decoder"},{"citing_arxiv_id":"2406.09246","ref_index":19,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"OpenVLA: An Open-Source Vision-Language-Action Model","primary_cat":"cs.RO","submitted_at":"2024-06-13T15:46:55+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"OpenVLA achieves 16.5% higher task success than the 55B RT-2-X model across 29 tasks with 7x fewer parameters while enabling effective fine-tuning and quantization without performance loss.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"models for robotic representation learning [12-14] and as a component in modular systems for task planning and execution [15, 16]. More recently, they have been used for directly learning vision- language-action models [VLAs; 1, 7, 17, 18] for control. VLAs provide a direct instantiation of using pretrained vision-and-language foundation models for robotics, directly fine-tuning visually- conditioned language models (VLMs) such as PaLI [19, 20] to generate robot control actions. By building off of strong foundation models trained on Internet-scale data, VLAs such as RT-2 [ 7] demonstrate impressive robustness results, as well as an ability to generalize to novel objects and tasks, setting a new standard for generalist robot policies. Yet, there are two key reasons preventing the widespread use of existing VLAs: 1) current models [ 1, 7, 17, 18] are closed, with limited"},{"citing_arxiv_id":"2404.18416","ref_index":15,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Capabilities of Gemini Models in Medicine","primary_cat":"cs.AI","submitted_at":"2024-04-29T04:11:28+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Med-Gemini sets new records on 10 of 14 medical benchmarks including 91.1% on MedQA-USMLE, beats GPT-4V by 44.5% on multimodal tasks, and surpasses humans on medical text summarization.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2312.16886","ref_index":19,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"MobileVLM : A Fast, Strong and Open Vision Language Assistant for Mobile Devices","primary_cat":"cs.CV","submitted_at":"2023-12-28T08:21:24+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"MobileVLM achieves on-par performance with much larger vision-language models on standard benchmarks while delivering state-of-the-art inference speeds of 21.5 tokens per second on Snapdragon 888 CPU and 65.3 on Jetson Orin GPU.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"millions of text-image pairs in the line of VLMs, where the new datasets are usually released alongside their cor- responding new models. To name a few, apart from an enhanced visual receptor and novel language model called Qwen-LM [4], the multilingual multimodal Qwen-VL [5] additionally aligns the image with caption and box tu- ples, which sets a new record of generalist models. PALI [19] and PALI-X [18] consume an internal multi-language image-text dataset called WebLI at a scale of 12 billion. Most recently, observing the constraints of current image- text datasets like hallucination and inaccurate descriptions, ShareGPT4V [16] exploits GPT-4V [90] for generating 1.2M high-quality image-text pairs with which can surpass the LLaV A series."},{"citing_arxiv_id":"2312.11805","ref_index":11,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Gemini: A Family of Highly Capable Multimodal Models","primary_cat":"cs.CL","submitted_at":"2023-12-19T02:39:27+00:00","verdict":"CONDITIONAL","verdict_confidence":"MODERATE","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Gemini Ultra reaches human-expert performance on MMLU for the first time and sets new state-of-the-art results on 30 of 32 benchmarks, including all 20 multimodal ones tested.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2311.04257","ref_index":11,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"mPLUG-Owl2: Revolutionizing Multi-modal Large Language Model with Modality Collaboration","primary_cat":"cs.CL","submitted_at":"2023-11-07T14:21:29+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"mPLUG-Owl2 presents a modular MLLM architecture that enables modality collaboration via shared functional modules and modality-adaptive components, achieving SOTA on both text and multi-modal tasks with one generic model.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2310.06114","ref_index":112,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Learning Interactive Real-World Simulators","primary_cat":"cs.AI","submitted_at":"2023-10-09T19:42:22+00:00","verdict":"CONDITIONAL","verdict_confidence":"MODERATE","novelty_score":7.0,"formal_verification":"none","one_line_summary":"UniSim learns a universal real-world simulator from orchestrated diverse datasets, enabling zero-shot deployment of policies trained purely in simulation.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2308.01390","ref_index":6,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"OpenFlamingo: An Open-Source Framework for Training Large Autoregressive Vision-Language Models","primary_cat":"cs.CV","submitted_at":"2023-08-02T19:10:23+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"OpenFlamingo provides open-source autoregressive vision-language models that achieve 80-89% of Flamingo performance on seven vision-language datasets.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"Model Language model Cross-attention interval <image> and <|endofchunk|> OpenFlamingo-3B MPT-1B [27] 1 Trainable OpenFlamingo-3B (Instruct) MPT-1B (Instruct) [27] 1 Trainable OpenFlamingo-4B RedPajama-3B [35] 2 Frozen OpenFlamingo-4B (Instruct) RedPajama-3B (Instruct) [35] 2 Frozen OpenFlamingo-9B MPT-7B [27] 4 Trainable 2 and LLaVa, can incorporate only one image in their context [6, 16, 22, 25, 39, 41], autoregressive vision-language models accept interleaved image- text sequences, enabling in-context learning. We chose to replicate Flamingo because of its strong in-context learning abilities. Aggregated across evaluation sets, Flamingo models see steady performance improvements up to 32 in- context examples [ 3]. This is in contrast with"},{"citing_arxiv_id":"2303.15343","ref_index":13,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Sigmoid Loss for Language Image Pre-Training","primary_cat":"cs.CV","submitted_at":"2023-03-27T15:53:01+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SigLIP replaces softmax-based contrastive loss with a simple pairwise sigmoid loss for vision-language pre-training, decoupling batch size from normalization and reaching strong zero-shot performance with limited compute.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2303.03378","ref_index":7,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"PaLM-E: An Embodied Multimodal Language Model","primary_cat":"cs.LG","submitted_at":"2023-03-06T18:58:06+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"PaLM-E is a single 562B-parameter multimodal model that performs embodied reasoning tasks like robotic manipulation planning and visual question answering by interleaving vision, state, and text inputs with positive transfer from joint training on language and robotics data.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2302.11550","ref_index":7,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Scaling Robot Learning with Semantically Imagined Experience","primary_cat":"cs.RO","submitted_at":"2023-02-22T18:47:51+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Augmenting robot datasets via diffusion-based semantic inpainting enables manipulation policies to solve unseen tasks with new objects and improves robustness to novel distractors.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2301.12597","ref_index":2,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models","primary_cat":"cs.CV","submitted_at":"2023-01-30T00:56:51+00:00","verdict":"UNVERDICTED","verdict_confidence":"MODERATE","novelty_score":7.0,"formal_verification":"none","one_line_summary":"BLIP-2 bootstraps vision-language pre-training from frozen image encoders and LLMs via a lightweight two-stage Querying Transformer, delivering SOTA results with 54x fewer trainable parameters than Flamingo80B on zero-shot VQAv2.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null}],"limit":50,"offset":0}