🤗 Daily Paper(2025-05-12)

[deep.di...@gmail.com]

🤗 Daily Paper Newsletter Hope you found some gems! This newsletter delivers you the curated list of papers by 🤗 Daily Papers. project page 🤗 daily paper Multiview Point Cloud Registration via Optimization in an Autoencoder Latent Space Published at 2025-04-30 #ML The authors present POLAR, a multiview registration method that efficiently handles many views and is robust to high levels of degradation and large initial angles. POLAR converts the registration problem into the latent space of a pretrained autoencoder, uses a custom loss to account for degradations, and employs an efficient multistart optimization strategy, outperforming existing methods on both synthetic and real data.... Read More Bielik 11B v2 Technical Report Published at 2025-05-05 #ML The Bielik 11B v2 model is a highly efficient language model tailored for Polish text processing, built on Mistral 7B v0.2 architecture and enhanced with two innovative techniques: Weighted Instruction Cross-Entropy Loss and Adaptive Learning Rate. This model excels in Polish language benchmarks, outperforming many larger models, and offers extensive quantization options for deployment across different hardware configurations.... Read More Bielik v3 Small: Technical Report Published at 2025-05-05 #ML Researchers have created two new generative text models, Bielik v3 Small, tailored for Polish language processing that perform as well as larger models but use less computational resources. They used a custom Polish tokenizer, a special loss function, and an adaptive learning rate to train these models on a large corpus of Polish texts.... Read More Sailing AI by the Stars: A Survey of Learning from Rewards in Post-Training and Test-Time Scaling of Large Language Models Published at 2025-05-05 #ML This survey explores how Large Language Models (LLMs) use rewards to guide their behavior, a method called Learning from Rewards, which helps LLMs adapt from static data to dynamic feedback. The survey categorizes and analyzes this method across different stages and discusses its applications, benchmarks, challenges, and future directions.... Read More A Preliminary Study for GPT-4o on Image Restoration Published at 2025-05-08 #ML This study explores the capabilities of GPT-4o, a powerful image generation model, in the field of image restoration. The researchers found that while GPT-4o's images look good, they sometimes have issues with exact details, but these images can still help improve existing restoration methods, providing a foundation for future advancements in image generation.... Read More G-FOCUS: Towards a Robust Method for Assessing UI Design Persuasiveness Published at 2025-05-08 #ML This study presents WiserUI-Bench, a benchmark for comparing UI design persuasiveness, and G-FOCUS, a method to improve the accuracy of Vision-Language Models in assessing UI design effectiveness. The new approach helps optimize user interactions by reducing bias and improving evaluation, offering a more efficient alternative to costly and time-consuming A/B testing.... Read More Healthy LLMs? Benchmarking LLM Knowledge of UK Government Public Health Information Published at 2025-05-09 #ML The study presents a new benchmark, PubHealthBench, to evaluate Large Language Models' (LLMs) understanding of UK government public health information. The results show that advanced LLMs perform well in multiple-choice questions but struggle with open-ended queries, suggesting the need for additional safeguards when using them for free-form public health responses.... Read More UniVLA: Learning to Act Anywhere with Task-centric Latent Actions Published at 2025-05-09 #ML The authors present UniVLA, a new framework for training robots to perform tasks in various environments using a latent action model derived from videos. UniVLA outperforms existing methods, requiring less computational resources and data, and can be deployed to different robots through efficient latent action decoding.... Read More Tags are generated by Google's Gemini Pro API, and the summary and translation are generated by Upstage's SOLAR mini chat model derived from SOLAR-10.7B open LLM. (Experimental) The full paper is translated in korean with enko-t5-small-v0 model developed by Kim Kihyun. Visit Developer's Social Media