✨Don't Blind Your VLA: Aligning Visual Representations for OOD Generalization

📝 Summary:
Naive action fine-tuning degrades visual representations in Vision-Language-Action models. This study analyzes this degradation and introduces a simple method to align representations, improving out-of-distribution generalization.

🔹 Publication Date: Published on Oct 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.25616
• PDF: https://arxiv.org/pdf/2510.25616
• Project Page: https://blind-vla-paper.github.io
• Github: https://github.com/CognitiveAISystems/BlindVLA

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#VLA #OODGeneralization #ComputerVision #MachineLearning #RepresentationLearning

✨Dynamic Reflections: Probing Video Representations with Text Alignment

📝 Summary:
This work presents the first comprehensive study on video-text representation alignment. It reveals alignment depends on data richness and correlates with downstream task performance, suggesting its value for general video understanding. This introduces video-text alignment as a zero-shot method ...

🔹 Publication Date: Published on Nov 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.02767
• PDF: https://arxiv.org/pdf/2511.02767
• Github: https://video-prh.github.io/

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#VideoUnderstanding #TextAlignment #VideoTextAI #ZeroShotLearning #RepresentationLearning

✨FedRE: A Representation Entanglement Framework for Model-Heterogeneous Federated Learning

📝 Summary:
FedRE is a federated learning framework for model-heterogeneous environments. Clients create and upload entangled representations and entangled-label encodings to train a global classifier. This method enhances performance, protects privacy, and reduces communication overhead.

🔹 Publication Date: Published on Nov 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.22265
• PDF: https://arxiv.org/pdf/2511.22265

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#FederatedLearning #MachineLearning #AI #PrivacyPreservingAI #RepresentationLearning

✨In-Context Representation Hijacking

📝 Summary:
Doublespeak is an in-context attack that hijacks LLM representations. It replaces harmful keywords with benign ones in examples, making LLMs interpret innocuous prompts as harmful, bypassing safety. This highlights a need for representation-level alignment.

🔹 Publication Date: Published on Dec 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.03771
• PDF: https://arxiv.org/pdf/2512.03771

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #AISafety #AIsecurity #InContextLearning #RepresentationLearning

✨The Prism Hypothesis: Harmonizing Semantic and Pixel Representations via Unified Autoencoding

📝 Summary:
The Prism Hypothesis posits semantic encoders capture low-frequency meaning, while pixel encoders retain high-frequency details. Unified Autoencoding UAE leverages this with a frequency-band modulator to harmonize both into a single latent space. This achieves state-of-the-art performance on imag...

🔹 Publication Date: Published on Dec 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.19693
• PDF: https://arxiv.org/pdf/2512.19693
• Github: https://github.com/WeichenFan/UAE

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#DeepLearning #ComputerVision #Autoencoders #RepresentationLearning #AIResearch

✨Dynamic Large Concept Models: Latent Reasoning in an Adaptive Semantic Space

📝 Summary:
DLCM shifts computation from individual tokens to a compressed concept space, enabling more efficient reasoning. This hierarchical approach learns semantic boundaries end-to-end and improves performance on benchmarks by reallocating compute.

🔹 Publication Date: Published on Dec 31, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.24617
• PDF: https://arxiv.org/pdf/2512.24617

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#AI #MachineLearning #LargeModels #RepresentationLearning #EfficientAI

✨Implicit Neural Representation Facilitates Unified Universal Vision Encoding

📝 Summary:
This paper unifies image representation learning for both recognition and generation. It uses a hyper-network for implicit neural representation with knowledge distillation to create compressed embeddings. The model achieves state-of-the-art results and enables generative capabilities.

🔹 Publication Date: Published on Jan 20

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.14256
• PDF: https://arxiv.org/pdf/2601.14256
• Github: https://github.com/tiktok/huvr

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#ComputerVision #DeepLearning #GenerativeAI #RepresentationLearning #VisionEncoding

✨Communication-Inspired Tokenization for Structured Image Representations

📝 Summary:
COMiT introduces a framework for learning structured, object-centric visual tokens through iterative encoding and flow-matching decoding. This single-transformer approach improves compositional generalization and relational reasoning by creating interpretable token structures.

🔹 Publication Date: Published on Feb 24

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.20731
• PDF: https://arxiv.org/pdf/2602.20731
• Project Page: https://araachie.github.io/comit/
• Github: https://github.com/araachie/comit

🔹 Models citing this paper:
• https://huggingface.co/cvg-unibe/comit-xl
• https://huggingface.co/cvg-unibe/comit-l
• https://huggingface.co/cvg-unibe/comit-b

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#ComputerVision #Transformers #ImageRecognition #RepresentationLearning #AIResearch

✨Layer by layer, module by module: Choose both for optimal OOD probing of ViT

📝 Summary:
Intermediate layers in ViTs provide better representations. Performance degradation in deeper layers is caused by distribution shift. Optimal probing depends on shift magnitude: FFN activation for strong shift, MHA output for weak shift.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05280
• PDF: https://arxiv.org/pdf/2603.05280
• Github: https://github.com/ambroiseodt/vit-probing

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#ViT #OOD #DeepLearning #RepresentationLearning #ComputerVision

✨Efficient Universal Perception Encoder

📝 Summary:
EUPE enhances edge device performance through a novel two-stage knowledge distillation approach. It scales up to a large proxy teacher then down to an efficient encoder. This method provides superior, versatile representations for diverse tasks, outperforming prior techniques.

🔹 Publication Date: Published on Mar 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.22387
• PDF: https://arxiv.org/pdf/2603.22387
• Github: https://github.com/facebookresearch/eupe

🔹 Models citing this paper:
• https://huggingface.co/facebook/EUPE-ConvNeXt-S
• https://huggingface.co/facebook/EUPE-ViT-S
• https://huggingface.co/facebook/EUPE-ViT-B

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#KnowledgeDistillation #EdgeAI #ComputerVision #DeepLearning #RepresentationLearning

🔥 Representation Distribution Matching for One-Step Visual Generation

💡 The paper introduces Representation Distribution Matching, a method for one-step visual generation that matches feature distributions under pretrained encoders. The goal is to generate high-quality images by comparing the distributions of generated and reference features. The authors identify two key design axes: how the distributions are compared and the representations they are compared in. They conduct controlled studies and find three main results.

First, they show that the Maximum Mean Discrepancy, a classical method that was previously ineffective, becomes a strong and scalable objective when estimated correctly. Second, they find that the batch size of the generated images has a significant impact on performance, with an optimum batch size above 2048, which is much larger than typical batch sizes. Third, they demonstrate that using a single representation can be gamed, resulting in low scores despite visibly fake images, and instead propose using a balanced set of encoders and evaluating with a Sliced-Wasserstein distance over 14 encoders.

The authors combine these findings to develop an improved Representation Distribution Matching method, which they call iRDM. They evaluate iRDM on the ImageNet dataset and achieve state-of-the-art results, with a Sliced-Wasserstein distance of 1.30. Additionally, they use a human-preference proxy, called PickScore, which shows that iRDM is preferred over the previous best one-step generator on 71.2% of matched samples. They also apply the same method to post-train a four-step generator, called FLUX.2, and achieve better results than the original four-step version, with improved performance on GenEval and PickScore, and requiring only 90 GPU-hours. Overall, the paper presents a new method for one-step visual generation that achieves state-of-the-art results and can be used to improve existing generators.

📅 Published on Jul 2

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2607.02375
• PDF: https://arxiv.org/pdf/2607.02375
• Project Page: https://alan-lanfeng.github.io/rdm/

🤖 Models citing this paper:
• https://huggingface.co/epfl-vita/flux2-klein-1step-rdm

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/epfl-vita/flux2-klein-1step-demo

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#VisualGeneration #RepresentationLearning #DistributionMatching #ImageSynthesis #DeepLearning