AI Research Paper Feed

TLDR: This paper presents a novel reinforcement learning framework that uses semantic and token entropy to improve reasoning in large language models, outperforming existing methods across multiple benchmarks.

TLDR: CARL is a reinforcement learning algorithm that focuses on optimizing critical actions in multi-step tasks, leading to improved performance and efficiency.

TLDR: Phase-Preserving Diffusion (φ-PD) enables structure-aligned generation by preserving input phase in the diffusion process, improving spatial consistency in tasks like image-to-image translation.

TLDR: The gpuFLOPBench benchmark evaluates Large Language Models' (LLMs) ability to predict FLOP counts for CUDA kernels, highlighting their challenges in reasoning about code complexity without execution.

TLDR: YingMusic-SVC is a robust zero-shot singing voice conversion system that improves timbre similarity and naturalness in real-world conditions using innovative techniques like Flow-GRPO and singing-specific biases.

TLDR: This study proposes using smartphone speech analysis integrated with Relational Graph Transformers for continuous monitoring of cognitive symptoms in patients with rare neurological diseases, showing promise in phenylketonuria (PKU) cases.

TLDR: This study presents a multi-agent reinforcement learning framework for scheduling operating rooms, outperforming traditional methods in balancing elective and urgent surgeries under uncertainty.

TLDR: The paper introduces a theoretical framework for analyzing reasoning algorithms in large language models, focusing on iterative solution improvement and answer aggregation.

TLDR: The paper introduces Format Reinforcement Learning (FormatRL) to improve structured document translation by optimizing structure-aware rewards, achieving better results on SAP documentation benchmarks.

TLDR: YingMusic-Singer introduces a novel melody-driven singing voice synthesis framework that operates without manual phoneme alignment or melody annotation, improving scalability and performance in zero-shot settings.

TLDR: BabySeg is a deep learning framework for accurate infant brain segmentation from diverse MRI protocols, achieving state-of-the-art performance across various age groups and scan types.

TLDR: The thesis presents novel methods for causal inference and representation learning in high-dimensional, time-varying data, introducing models like CDVAE and frameworks utilizing RNNs and CPC for improved causal analysis.

TLDR: This paper reveals that system prompts in large vision-language model-based text-to-image systems significantly contribute to social biases, and proposes a framework called FairPro to reduce these biases while maintaining text-image alignment.

TLDR: TV2TV is a new framework for video generation that interleaves text and video generation to improve visual quality and control by using a Mixture-of-Transformers architecture.

TLDR: The paper presents DAMASHA, a framework for detecting transitions between human and AI-generated text using stylometric cues and perplexity signals, and introduces a new benchmark for testing its robustness against adversarial text segments.

TLDR: SmartAlert, a machine learning-driven clinical decision support system, successfully reduced unnecessary repeat lab tests in hospitals by 15% without compromising patient safety.

TLDR: AgentBay is a hybrid sandbox for seamless human-AI interaction, improving task success rates and reducing latency and bandwidth use in AI systems.

TLDR: The paper proposes a deterministic functional-topological framework for understanding real-world perception, enabling systems to generalize from limited data by discovering the compact manifold structure of physical phenomena through self-supervised methods.

TLDR: Semantic Soft Bootstrapping improves long context reasoning in language models by using a self-distillation technique without requiring reinforcement learning, resulting in significant accuracy improvements on math benchmarks.

TLDR: ReflexFlow introduces a new learning objective to reduce exposure bias in Flow Matching, improving generation quality significantly across several datasets.