AI Research Paper Feed

TLDR: DFlash is a speculative decoding framework that uses block diffusion for parallel drafting, achieving over 6x acceleration in language model inference compared to traditional methods.

TLDR: KV-CoRE is a method to evaluate the data-dependent compressibility of kv-caches in large language models, revealing patterns linked to model architecture and training data across multiple languages.

TLDR: This paper establishes a tight law of the iterated logarithm for empirical KL-infinity statistics, applicable to very general data conditions.

TLDR: Copyright Detective is an interactive forensic system designed to detect and analyze potential copyright risks in the outputs of large language models (LLMs).

TLDR: The paper introduces new Radon--Wasserstein gradient flows for efficient high-dimensional sampling using interacting particles with linear scaling costs.

TLDR: LongR is a framework that improves long-context reasoning in reinforcement learning by using a dynamic 'Think-and-Read' mechanism and dense utility rewards, achieving significant gains on benchmarks like LongBench v2.

TLDR: ADANA, an optimizer with time-varying schedules for hyperparameters, improves large-scale language model training efficiency by up to 40% compared to AdamW.

TLDR: AIMMD is a new algorithm that uses machine learning to improve the efficiency of transition path sampling for studying molecular processes.

TLDR: The paper presents a method to achieve differentially private prediction with reduced dependence on the number of queries, improving efficiency in streaming settings.

TLDR: The paper introduces a Multi-Field Tool Retrieval framework to improve how Large Language Models select external tools by addressing challenges in tool documentation and user query alignment.

TLDR: The paper introduces UOT-ERRANT, a new metric for evaluating grammatical error correction systems by optimally transporting edit vectors, showing improved performance and interpretability.

TLDR: The Clifford Kolmogorov-Arnold Network (ClKAN) is a new architecture for approximating functions in Clifford algebra spaces, utilizing Randomized Quasi Monte Carlo methods and novel batch normalization strategies for improved scalability and efficiency.

TLDR: Diamond Maps are a new model for generative tasks that efficiently align with user preferences by enabling quick adaptation to rewards during inference.

TLDR: The paper introduces Concept DAS (CDAS), a novel intervention-based model steering method that uses distribution matching to achieve more faithful and stable control compared to traditional preference-optimization methods.

TLDR: The Hinge Regression Tree (HRT) is a new method for creating oblique decision trees using a Newton method that improves split quality and convergence speed, outperforming traditional tree models.

TLDR: The paper addresses fairness in machine learning under group-conditional prior probability shift and introduces a method to maintain fairness when label prevalences change across demographic groups between training and deployment.

TLDR: PACE introduces a more efficient method for mathematical reasoning in language models by using minimal exploration, outperforming traditional methods with less computational cost.

TLDR: This paper presents a unified convergence analysis for the SAG, SAGA, and IAG algorithms, providing a simpler and more comprehensive understanding of their performance.

TLDR: This paper presents a new construction of a transformer that can solve the PARITY problem using a single layer with practical features, and establishes a lower bound proving that a single-layer, single-head transformer cannot solve PARITY.

TLDR: FedRandom is a novel technique that improves the stability and accuracy of participant contribution assessments in federated learning by generating more consistent samples, significantly reducing estimation errors.