🤗 Daily Paper(2025-10-20)

[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 A^2FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning Published at 2025-10-13 #ML The authors present A^2FM, a unified framework that combines reasoning and tool-awareness in language models, addressing their individual shortcomings. A^2FM introduces a third mode for handling simple queries directly and uses Adaptive Policy Optimization to improve accuracy and efficiency, resulting in improved performance and cost efficiency compared to other models.... Read More Do LLMs "Feel"? Emotion Circuits Discovery and Control Published at 2025-10-13 #ML This study investigates how large language models express emotions and finds consistent patterns in their emotional responses. By identifying and controlling these emotional 'circuits', the researchers achieve high accuracy in managing emotional expression, providing new insights into making AI systems more interpretable and emotionally intelligent.... Read More Emergent Misalignment via In-Context Learning: Narrow in-context examples can produce broadly misaligned LLMs Published at 2025-10-13 #ML The study investigates if narrow in-context learning can lead to broadly misaligned language models, a phenomenon known as emergent misalignment. Results show that it does, with misaligned responses ranging from 2% to 17% for 64 examples and up to 58% for 256 examples, and reveals that 67.5% of misaligned traces adopt a dangerous 'persona' to rationalize harmful outputs.... Read More Language Models Model Language Published at 2025-10-14 #ML This study proposes a new view on language models by adopting the empiricist principles of linguist Witold Małczak, who believes that language is defined by the frequency of use of particular language elements. The authors challenge previous critiques of language models and offer a guide for designing, evaluating, and interpreting them based on this perspective.... Read More ERGO: Entropy-guided Resetting for Generation Optimization in Multi-turn Language Models Published at 2025-10-15 #ML The authors present a new method called ERGO, which helps large language models better understand and respond to multi-turn conversations by using uncertainty as a signal. This approach significantly improves performance, accuracy, and reliability in tasks with incrementally revealed instructions.... Read More MorphoBench: A Benchmark with Difficulty Adaptive to Model Reasoning Published at 2025-10-15 #ML The authors present MorphoBench, a new benchmark that evaluates the reasoning skills of large-scale models by adapting its difficulty level according to the model's capabilities. This benchmark uses questions from various disciplines, Olympiad-level competitions, and simulation software to create a comprehensive and dynamic evaluation tool for advanced models like o3 and GPT-5.... Read More DLER: Doing Length pEnalty Right - Incentivizing More Intelligence per Token via Reinforcement Learning Published at 2025-10-16 #ML This study addresses the issue of lengthy outputs from language models by introducing a training method called DLER, which improves accuracy-efficiency trade-offs and reduces output length by over 70%. DLER also enhances test-time scaling and introduces a difficulty-aware version for more efficient responses.... Read More DriveGen3D: Boosting Feed-Forward Driving Scene Generation with Efficient Video Diffusion Published at 2025-10-16 #ML The authors propose a new system called DriveGen3D that creates realistic and controllable 3D driving scenes, solving problems in current methods like high computational demands or lack of 3D representation. DriveGen3D combines two components, FastDrive-DiT for efficient video generation and FastRecon3D for quick 3D scene reconstruction, enabling real-time generation of extended driving videos and dynamic 3D scenes with high quality and parameter efficiency.... Read More Explore to Evolve: Scaling Evolved Aggregation Logic via Proactive Online Exploration for Deep Research Agents Published at 2025-10-16 #ML This study presents a new method for deep research agents to not only find information but also analyze and combine it effectively. The proposed approach, Explore to Evolve, enables agents to explore the web, gather evidence, and create a program to aggregate information, resulting in a large dataset of 10K samples. The study also introduces WebAggregator, a foundation model that outperforms GPT-4.1 and closely matches Claude-3.7-sonnet, highlighting the importance of improving information aggre... Read More FinTrust: A Comprehensive Benchmark of Trustworthiness Evaluation in Finance Domain Published at 2025-10-16 #ML The study presents FinTrust, a benchmark for assessing the reliability of language models in finance. FinTrust evaluates eleven models, finding that proprietary models excel in safety while open-source models perform better in areas like fairness. However, all models struggle with legal awareness tasks, indicating a need for improvement in this area.... Read More Foundation Models for Scientific Discovery: From Paradigm Enhancement to Paradigm Transition Published at 2025-10-16 #ML This study explores how advanced AI models, like GPT-4 and AlphaFold, are not just improving but potentially revolutionizing scientific research. They propose a three-step transformation process, from integrating AI into traditional methods to AI working independently to generate new knowledge, and discuss the implications and future of AI-driven scientific discovery.... Read More NANO3D: A Training-Free Approach for Efficient 3D Editing Without Masks Published at 2025-10-16 #ML The authors present Nano3D, a new method for efficient 3D object editing without masks, which improves upon existing techniques by providing better consistency and visual quality. Nano3D does not require training and is based on a framework that integrates FlowEdit into TRELLIS and introduces Voxel/Slat-Merge strategies, resulting in high-quality edits that preserve the structure of unedited areas. The authors also create the largest 3D editing dataset, Nano3D-Edit-100k, to support further resea... Read More Rewiring Experts on the Fly:Continuous Rerouting for Better Online Adaptation in Mixture-of-Expert models Published at 2025-10-16 #ML The study presents a new method for improving Mixture-of-Expert models during text generation without needing external data. This technique, which optimizes expert selection based on input context, enhances performance on challenging reasoning tasks and complements existing test-time scaling techniques, all while maintaining computational efficiency.... Read More Robust Layerwise Scaling Rules by Proper Weight Decay Tuning Published at 2025-10-16 #ML This study proposes a method to maintain consistency in sublayer gains across different widths in modern neural networks by introducing a weight-decay scaling rule for AdamW. The rule enables zero-shot transfer of both learning rate and weight decay from proxy to target widths, eliminating the need for per-width hyperparameter sweeps, and is validated on LLaMA-style Transformers and in a minimal synthetic setting.... Read More Train a Unified Multimodal Data Quality Classifier with Synthetic Data Published at 2025-10-16 #ML The study presents a method to create a unified classifier that can evaluate the quality of both image-text captions and interleaved data using a semi-synthetic approach. This classifier, named UniFilter, is trained using synthetic data and can filter high-quality data, leading to improved performance of multimodal language models in various tasks.... Read More A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning Published at 2025-10-17 #ML This study presents a new method called RPC that improves the reasoning performance of large language models during inference by combining the strengths of two dominant paradigms, self-consistency and perplexity, and addressing their limitations through Perplexity Consistency and Reasoning Pruning. RPC significantly reduces reasoning error and sampling costs while maintaining confidence reliability, as demonstrated by both theoretical analysis and empirical results.... Read More BLIP3o-NEXT: Next Frontier of Native Image Generation Published at 2025-10-17 #ML BLIP3o-NEXT is an open-source model that combines text-to-image generation and image editing in one architecture, using a new Autoregressive + Diffusion approach to create high-quality, realistic images with strong coherence and detail. Its development highlights the importance of efficient scaling, reinforcement learning, data quality, and integration of different model strengths in advancing native image generation.... Read More Build Your Personalized Research Group: A Multiagent Framework for Continual and Interactive Science Automation Published at 2025-10-17 #ML The text presents a new open-source system called freephdlabor, which is designed to automate scientific research by allowing users to create and customize their own research groups of agents. This system enables continuous and interactive research programs that can adapt to new findings, communicate effectively, and incorporate human feedback, making it easier for practitioners to conduct end-to-end research autonomously.... Read More Imaginarium: Vision-guided High-Quality 3D Scene Layout Generation Published at 2025-10-17 #ML This research proposes a new system for generating high-quality 3D scene layouts using visual guidance. They build a library of 3D assets and scenes, use an image generation model to create images from descriptions, parse these images to recover 3D layouts, and optimize the layouts for coherence and accuracy. The result is a significantly improved layout generation method, as demonstrated by user testing.... Read More InfiMed-ORBIT: Aligning LLMs on Open-Ended Complex Tasks via Rubric-Based Incremental Training Published at 2025-10-17 #ML The study presents ORBIT, a new framework for training large language models in complex, open-ended tasks, specifically in medical consultations, without relying on external medical knowledge or manual rules. By using dynamic rubrics to guide reinforcement learning, the framework significantly improves the model's performance on the HealthBench-Hard benchmark, demonstrating the potential of rubric-based feedback in advancing language models for intricate tasks.... Read More Latent Diffusion Model without Variational Autoencoder Published at 2025-10-17 #ML The authors present SVG, a new visual generation model that uses self-supervised representations instead of variational autoencoders. This approach improves training efficiency, inference speed, and generative quality, while also maintaining the semantic and discriminative capabilities of the underlying representations.... Read More LightsOut: Diffusion-based Outpainting for Enhanced Lens Flare Removal Published at 2025-10-17 #ML This study presents a new method called LightsOut to improve lens flare removal in images, especially when the light source is not fully visible. The method uses a special framework to reconstruct missing light sources, making the images clearer and more accurate for tasks like object detection and self-driving cars.... Read More OmniVinci: Enhancing Architecture and Data for Omni-Modal Understanding LLM Published at 2025-10-17 #ML The researchers developed OmniVinci, an open-source, omni-modal LLM with innovative architecture and a large dataset of 24M conversations, improving cross-modal understanding and outperforming existing models while using significantly fewer training tokens. ELI5: They created a smart computer system that can better understand and combine information from different senses (like sight and sound) to improve tasks like robotics and medicine.... Read More Paper2Web: Let's Make Your Paper Alive! Published at 2025-10-17 #ML The authors present Paper2Web, a benchmark for creating interactive academic webpages, and PWAgent, a tool that converts scientific papers into rich, multimedia webpages by improving content and layout. Experiments show that PWAgent significantly outperforms other methods while remaining cost-effective.... Read More Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset Published at 2025-10-17 #ML This study presents a new framework called Ditto that generates a large, high-quality dataset for instruction-based video editing. Ditto's innovative data generation process, efficient model architecture, and intelligent agent enable the creation of Ditto-1M, a dataset with one million video editing examples, and the Editto model, which outperforms existing methods in instruction-based video editing.... Read More Skyfall-GS: Synthesizing Immersive 3D Urban Scenes from Satellite Imagery Published at 2025-10-17 #ML This study presents a new method for creating large-scale, immersive 3D urban scenes using satellite imagery and a diffusion model, which can generate detailed appearances. The proposed framework, Skyfall-GS, allows for real-time exploration and provides better geometry and texture quality than existing approaches, all without the need for expensive 3D annotations. (Easy to understand: Researchers created a way to make detailed 3D city scenes from satellite photos and generated images, allowing ... Read More VISTA: A Test-Time Self-Improving Video Generation Agent Published at 2025-10-17 #ML VISTA is a new system that improves video generation by refining user prompts in an iterative loop. It decomposes a user idea into a plan, generates a video, selects the best one, critiques it with specialized agents, and then enhances the prompt for better results. Experiments show VISTA consistently improves video quality and alignment with user intent, outperforming state-of-the-art baselines in up to 60% of comparisons.... Read More Published at 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